A bone layer confidence-driven robotic bone cutting force fusion safety control system
The robot bone cutting force fusion safety control system driven by bone layer confidence solves the problems of insufficient real-time performance, stability and safety of the control system in the bone layer change scenario in the existing technology, and realizes stable fusion and safe control in the bone layer state.
Patent Information
- Application Number
- CN202611233893.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing bone cutting robot control systems struggle to balance real-time performance, stability, and safety in scenarios with frequent changes in bone layers. Single force sensor feedback is susceptible to noise, and single prediction models are easily affected by sample differences and model errors. Furthermore, the weights lack a basis in the bone layer state, resulting in control oscillations and insufficient safety.
A robot bone cutting force fusion safety control system driven by bone layer confidence is established. Through six-dimensional force signal compensation and preprocessing, feature principal component mapping, bone layer confidence discrimination, hierarchical prediction and intra-layer fixed weight fusion, combined with fuzzy PID control, an interpretable bone layer weight mechanism is established. The weights are adjusted in real time and a residual safety envelope is constructed to ensure control stability and safety.
It achieves stable fusion of real-time feedback and model prediction in bone layer change scenarios, reduces the risk of force mutation at bone layer interfaces, improves control accuracy and robustness, and enhances the safety of surgical robots.
Smart Images

Figure CN123101255A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of robot force control technology, specifically relating to a robot bone cutting force fusion safety control system driven by bone layer confidence. Background Technology
[0002] Existing bone-cutting robot control systems typically rely on six-dimensional force sensor feedback, empirical models, or machine learning prediction models. While single force sensor feedback can reflect the actual cutting state, it is susceptible to zero-point drift, tool posture changes, sensor noise, and heterogeneous disturbances in bone tissue. Single prediction models offer some smoothness and trend prediction capabilities, but are also affected by sample differences, bone density variations, and model errors. Neither approach, when used alone, can effectively balance real-time performance, stability, and safety in scenarios with frequent changes in bone layer.
[0003] In existing technologies, there is a certain foundation for establishing bone layer cutting force prediction models using statistical features, PCA, SVM, or GA-SVM, and for using fuzzy PID to regulate feed rate. However, if only the real-time force principal components and the predicted force principal components are globally weighted with fixed weights, there are still three shortcomings: First, the weights lack the basis of bone layer state and model error, and are easily regarded as empirical weighting; second, the noise levels, prediction errors, and control response requirements of cortical bone, cancellous bone, and transition layers are different, and a single weight cannot cover all bone layers; third, if the prediction model or control rules are directly switched when switching bone layers, a sudden increase in residuals and feed rate oscillations may occur.
[0004] Therefore, there is an urgent need for a bone layer confidence-driven robot bone cutting force fusion safety control system that establishes an interpretable bone layer weight mechanism between real-time feedback and model prediction, maintains fixed weights within the bone layer to suppress weight oscillations, switches weights and control rules through confidence and hysteresis rules when switching bone layers, and actively protects against model mismatch, bone layer mutations, and depth overruns through residual safety envelopes. Summary of the Invention
[0005] To address the aforementioned problems in existing technologies, this invention proposes a bone layer confidence-driven robotic bone cutting force fusion safety control system. This system is rationally designed, overcomes the shortcomings of existing technologies, and exhibits excellent performance.
[0006] A bone layer confidence-driven robotic bone cutting force fusion safety control system includes: The data acquisition module is used to acquire six-dimensional force sensing signals, robot end-effector pose, and cutting parameters during the robot's layered cutting of bone tissue. The six-dimensional force sensing signals include force component signals along three mutually orthogonal directions and torque component signals around three mutually orthogonal axes. The cutting parameters include spindle speed, feed rate, and depth of cut. The signal compensation and preprocessing module is used to perform zero-point drift correction, gravity and its torque component compensation, filtering and noise reduction and outlier suppression on the six-dimensional force sensor according to the robot end pose and the installation pose relationship between the cutting tool and the six-dimensional force sensor, so as to obtain the compensated six-dimensional bone tissue cutting force sensor signal. The feature principal component mapping module is used to extract windowed statistical features from the six-dimensional bone tissue cutting force sensory signal, and to perform fixed mapping using standardized parameters and principal component analysis load matrix determined and locked during the training phase to obtain the real-time first force principal component. The bone layer confidence discrimination module is used to construct a bone layer scoring function based on the real-time first principal component, the real-time first principal component change rate, the cutting depth, the unit cutting energy, and the real-time prediction residual. It obtains the confidence of cortical bone, cancellous bone, and transition layer through a normalized exponential function, and determines the current bone layer status based on the continuous window criterion and the hysteresis switching rule. The stratified prediction module is used to select a stratified prediction sub-model for bone tissue cutting force based on the bone layer state determined in the previous data window. When the previous data window is a stable cortical bone state or a stable cancellous bone state, the prediction sub-model corresponding to the stable bone layer is invoked, and a predicted first principal component aligned with the current data window is generated based on the information from the previous data window. When the previous data window is a transitional layer state, the prediction sub-model corresponding to the most recently confirmed stable bone layer is invoked to generate a predicted first principal component aligned with the current data window, and the predicted first principal component is used only for real-time prediction residual monitoring. When the bone layer state in the current data window is determined to be a transitional layer state, or when the bone layer state in the current data window is inconsistent with the stable bone layer state on which the predicted first principal component was generated, the predicted first principal component does not participate in closed-loop fusion control. The intra-layer fixed weight fusion module is used to determine the intra-layer fixed fusion weight of the real-time first principal component based on the current stable bone layer state, prediction model error variance, real-time feedback fluctuation variance, bone layer confidence and normalized residual; the intra-layer fixed fusion weight remains unchanged within the same stable bone layer control cycle and is updated when the stable bone layer state changes; in the transition layer state, the fusion weight of the real-time first principal component is set to 1; The fusion feedback generation module is used to fuse the real-time first principal component and the predicted first principal component according to the fixed fusion weight within the layer to obtain the fusion feedback first principal component. The residual safety monitoring module is used to construct a comprehensive risk index based on normalized residual, peak force, depth of cut, feed rate and feed acceleration, and to switch between normal control mode, speed limit and load reduction mode and stop protection mode according to the comprehensive risk index. The hierarchical fuzzy PID control module is used to call the fuzzy PID control rule corresponding to the current bone layer state based on the deviation between the fused feedback first principal component and the target first principal component and the comprehensive risk index, and output the feed speed command of the robot end effector.
[0007] Furthermore, in the principal component mapping module, the windowed statistical features include one or more of the following: mean, peak-to-peak value, root mean square value, variance, skewness, kurtosis, energy ratio in the principal direction, and unit cutting energy. The windowed statistical feature vector is represented as follows: ; in, The mean of the window. The root mean square value, Peak-to-peak value For window variance, For skewness, For ravine, The energy ratio in the dominant direction, The cutting energy is expressed as a unit.
[0008] The unit cutting energy Calculated based on the cutting force, feed rate, and equivalent cutting volume within the current window: ; ; in, For the first A data window, Force in the cutting direction, For feed rate, The equivalent cutting volume for the current window. For the equivalent cutting cross-sectional area, This represents the increment of the cutting depth within the window. To prevent positive numbers with a denominator of zero, This represents the sampling time interval between adjacent sampling moments.
[0009] Furthermore, in the feature principal component mapping module, the principal component consistent mapping includes: using the mean, standard deviation, and PCA loading matrix determined during the training phase to perform a fixed mapping on the online windowed statistical features, specifically: ; ; ; ; in, For the first Statistical feature vectors extracted from each data window. and These are the feature mean vector and feature standard deviation vector determined during the training phase, respectively. This indicates element-wise division. It is a vector of all 1s. To standardize the feature vector, The number of training samples. For the first The standardized feature vector corresponding to each training sample To train the sample covariance matrix, Principal component loading matrix, It is an eigenvalue diagonal matrix. This is the first principal component load vector. It is the primary principal component in real time.
[0010] Furthermore, in the bone layer confidence discrimination module, the real-time prediction residual... and normalized residuals They are respectively: ; ; in, Indicates according to the first Information from the first data window, and based on the information from the first data window. The stable bone layer state determined by the first data window selects the corresponding prediction sub-model, generating a model that corresponds to the first... The first principal component of the prediction, time-aligned with the data window, when the... When a data window is in a transitional layer state, the predicted first principal component is generated using the prediction sub-model corresponding to the most recently confirmed stable bone layer. Indicates bone layer The standard deviation of the residuals is below; Calculate the first principal component in real time The amplitude is obtained by normalization. The normalized real-time first principal component amplitude, real-time first principal component change rate, cutting depth, unit cutting energy, and normalized residual are used to construct the normalized eigenvector: ; in, This is the normalized value of the real-time principal component amplitude. This is the normalized value of the rate of change of the principal components in real time. This is a normalized value for the depth of cut. This is the normalized value of the unit cutting energy.
[0011] Furthermore, in the bone layer confidence discrimination module, the bone layer scoring function includes a cortical bone scoring function, a cancellous bone scoring function, and a transition layer scoring function, which are as follows: ; ; ; in, , and These represent the scores for cortical bone, cancellous bone, and transition layer, respectively. to , to , to The corresponding scoring parameters are determined by constraint optimization based on training samples with bone layer category labels, with the goal of minimizing the sum of bone layer classification cross-entropy loss and regularization term, and remain fixed during online bone layer confidence calculation. Entropy is the confidence level. For the bone layer dominant septum, the expression is: ; when At that time, it was stipulated ; ; in, Indicates the first Bone layer in each data window The normalized confidence level, , Indicates cortical bone, Indicates cancellous bone, Indicates a transition layer; and These represent the confidence scores for the largest and second-largest bone layers in the previous data window, respectively. The confidence scores for cortical bone, cancellous bone, and transition layer were obtained using a normalized exponential function. ; in, , and They represent the first The confidence scores for cortical bone, cancellous bone, and transition layer are given in each data window. ; The hysteresis switching rules include stable bone layer entry rules and transition layer entry rules; for stable bone layers... The stable bone layer entry rule is as follows: ; The transition layer entry rule is as follows: ; and: ; in, For the first The current bone layer status corresponding to each data window , Indicates a transition layer. To determine the number of consecutive decision windows, and They are respectively stable bone layers The entry confidence threshold and exit confidence threshold, The threshold for entering the transition layer. To enter the stable bone layer The upper limit of the normalized residual allowed at that time. To exit the stable bone layer The normalized residual threshold, The hysteresis width.
[0012] Furthermore, the hierarchical prediction module includes a cortical bone prediction sub-model and a cancellous bone prediction sub-model; each prediction sub-model is based on the first... The data window contains spindle speed, feed rate, depth of cut, real-time first principal force component, real-time rate of change of the first principal force component, and unit cutting energy, predicting the first... The first principal component of each data window, with the input vector being: ; in, , and For the first Each data window displays the spindle speed, feed rate, and depth of cut. For the first The real-time principal components of each data window. This represents the rate of change of the principal components in real time. The unit cutting energy; To stabilize the bone layer The corresponding input vector of the prediction sub-model; The cortical bone prediction sub-model and the cancellous bone prediction sub-model are support vector machine regression models with parameters optimized by a genetic algorithm, and the prediction function is: ; in, To stabilize the bone layer The corresponding prediction sub-model outputs the predicted first principal component. and The first one obtained from training the support vector regression model Group of Lagrange multipliers, To stabilize the bone layer The corresponding number of support vectors, For kernel function, To stabilize the bone layer The corresponding prediction sub-model Support vectors, This is the bias term; the genetic algorithm is used to optimize the penalty parameters, kernel function parameters, and insensitive loss parameters of the support vector regression model. When the current bone layer state is a transitional layer state, the prediction sub-model corresponding to the most recently confirmed stable bone layer only outputs the predicted first principal component used to calculate the normalized residual. The predicted first principal component does not participate in the generation of the fusion feedback.
[0013] Furthermore, in the intralayer fixed weight fusion module, the stable bone layer... The corresponding prediction model error variance and real-time feedback fluctuation variance They are respectively: ; ; in, To stabilize the bone layer The corresponding number of validation samples, and The first The real-time first principal components and the time-aligned predicted first principal components of each validation sample. To stabilize the bone layer The following is the number of calibration samples used to estimate the variance of real-time feedback fluctuations. To stabilize the bone layer The mean of the first principal component in real time; Stable bone layer Initial real-time feedback fusion weights for: ; At the moment of transition of stable bone layer state The corrected intra-layer fixed fusion weights are: ; in, , This is the normalized residual adjustment coefficient. and They are respectively stable bone layers Allowed real-time feedback fusion weight lower and upper limits; fixed fusion weights within a layer. It remains unchanged within the same stable bone layer control cycle until the next stable bone layer state switch.
[0014] Furthermore, in the fusion feedback generation module, the real-time first principal component and the predicted first principal component are fused under the same stable bone layer state; If the current bone layer state is a stable bone layer and is consistent with the bone layer state when the predicted value was generated, the first principal component fusion form is: ; in, No. The real-time first principal component corresponding to each data window According to the first The information in the first data window is generated and compared with the first... The first principal component of the prediction is aligned with each data window. The fusion weights are for the first principal components in real time, and ; If the current bone layer state is a transitional state, or if the current bone layer state is inconsistent with the bone layer state at the time of generating the predicted value, the first principal component of the prediction will not participate in the fusion of the current control cycle. .
[0015] Furthermore, the comprehensive risk indicators in the residual safety monitoring module Characterized as: ; Among them, normalized residuals As a residual risk item, For peak force risk items, This is a deep boundary crossing risk item. For the risk item of exceeding the speed limit, To mitigate the risk of exceeding the limit; , , , and These are the weighting coefficients for the residual risk term, peak force risk term, depth out-of-bounds risk term, velocity out-of-bounds risk term, and acceleration out-of-bounds risk term, respectively, and the sum of these weighting coefficients is 1; the peak force risk term, depth out-of-bounds risk term, velocity out-of-bounds risk term, and acceleration out-of-bounds risk term are respectively: ; ; ; ; ; in, For the first The cutting force vector of each data window. Current bone layer status The preset upper limit threshold of platform cutting force, and These are the current bone layer states. The preset maximum and minimum process cutting depths, Current bone layer status The preset maximum feed rate, Current bone layer status The preset maximum feed acceleration, , For feed acceleration; when At this time, the system is in normal control mode; when When the system enters speed-limiting and load-reducing mode; when At that time, the system enters a safe deceleration to zero control phase; among which, The load reduction trigger threshold is a comprehensive risk indicator. This is the threshold for stopping the triggering of the comprehensive risk indicator, and .
[0016] Furthermore, in the hierarchical fuzzy PID control module, the stable bone layer... The corresponding target first principal component is determined by the mean of the real-time first principal component of the independent stable calibration window during the model development phase. The independent stable calibration window is a window with a clearly defined bone layer category, located within the corresponding stable bone layer cutting interval, and not in the bone layer transition region. The expression for the mean of the real-time first principal component is: ; in, To stabilize the bone layer The corresponding set of independent stable calibration windows, The number of valid windows in the set of independent and stable calibration windows. For the first The real-time first principal component corresponding to each independent stable calibration window; the target first principal component of the transition layer. This is a conservative target value preset based on the safety calibration data of the transition layer; No. The first principal component of the objective for each control cycle is: ; ; Stable bone layer The control error and error change rate are as follows: ; in, To control the cycle; Current stable bone layer The PID control gain is as follows: ; ; ; in, , and They are respectively stable bone layers The corresponding proportional, integral, and differential reference gains; , and These are the proportional, integral, and differential gain corrections obtained through fuzzy inference based on control error, error rate of change, current stable bone layer confidence level, and comprehensive risk index; when At that time, the proportional, integral, and differential reference gains respectively adopt the preset transition layer reference gains. , and The proportional, integral, and differential gain corrections are each limited to a preset transition layer correction range. , and Inside; Current bone layer status Control output below for: ; Robot end effector feed rate command for: ; in, Current bone layer status The corresponding speed adjustment coefficient, The comprehensive risk mitigation coefficient, and These are the current bone layer states. The preset minimum and maximum feed rates are: ; The feed rate command also satisfies the following acceleration constraints: ; in, Current bone layer status The preset maximum allowable feed acceleration.
[0017] The beneficial technical effects of this invention are as follows: 1. This invention places the real-time principal component, the predicted principal component, and the target principal component in the same fixed PCA principal component space, avoiding closed-loop control deviations caused by inconsistencies in the scales of the predicted quantity, the feedback quantity, and the control target.
[0018] 2. This invention introduces a bone layer scoring function, normalized index confidence, and hysteresis switching mechanism, which can not only distinguish between cortical bone and cancellous bone, but also identify the uncertain state at the bone layer interface as a safe state of the transition layer, thereby reducing the risk of force mutation at the bone layer interface.
[0019] 3. The fusion weights of this invention are not global empirical constants, but rather fixed weights within each layer based on the error variance of the prediction model for each bone layer, the variance of real-time feedback noise, the confidence level of the bone layer, and the normalized residuals. The weights remain fixed within each layer and are updated according to hysteresis conditions between layers, which improves the stability and interpretability of the fusion process.
[0020] 4. This invention utilizes real-time predictive residuals, peak force, depth of cut, feed rate, and feed acceleration to construct a residual safety envelope, which can proactively enter a speed-limiting and load-reducing or stop protection mode when there is model mismatch, sensor abnormality, bone layer mutation, or depth exceeding the limit, thereby enhancing the cutting safety of the surgical robot.
[0021] 5. This invention unifies bone layer confidence discrimination, layer prediction, intra-layer fixed weight fusion, residual safety supervision, and layered fuzzy PID control into a closed-loop control system, which can improve the control accuracy, stability, and robustness of bone tissue robots. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the robot bone tissue layer cutting and fusion safety control system of the present invention.
[0023] Figure 2 This is a flowchart of the safety control process for dual-source principal component fusion driven by bone layer confidence in this invention.
[0024] Figure 3 This is a schematic diagram of the pose coupling gravity compensation and coordinate mapping of the six-dimensional force sensor of the present invention.
[0025] Figure 4 The flowchart shows the preprocessing of cutting force signals, feature extraction, and PCA fixed load matrix mapping.
[0026] Figure 5 This is a flowchart illustrating the construction process of the bone tissue stratification prediction model and weight table of the present invention.
[0027] Figure 6 This is a schematic diagram of the bone layer confidence discrimination, hysteresis switching and intralayer fixed weight mechanism of the present invention.
[0028] Figure 7This is a block diagram of the dual-source principal component fusion control and residual safety envelope of the present invention.
[0029] Figure 8 PC1-MAE represents the relative trajectory level for different control strategies.
[0030] Figure 9 The confidence level of the region representing the continuous cutting trajectory. Detailed Implementation
[0031] The specific embodiments of the present invention will be further described below with reference to specific examples: A bone layer confidence-driven robotic bone cutting force fusion safety control system, such as... Figure 1 and 2 As shown, it includes: An initialization module is used to establish initialization information for a bone tissue layer cutting task. The initialization information includes the principal component of the target cutting force, the bone layer safety threshold, the cutting depth boundary, the feed rate boundary, the feed acceleration boundary, and the robot cutting trajectory. The principal component of the target cutting force can be obtained from ex vivo bone cutting calibration data, the safety force threshold, or preoperative bone density prior mapping.
[0032] The data acquisition module is used to acquire six-dimensional force sensing signals, robot end-effector pose, and cutting parameters during the robot's layered cutting of bone tissue. The six-dimensional force sensing signals include force components along three mutually orthogonal directions and torque components around three mutually orthogonal axes. The cutting parameters include spindle speed, feed rate, and depth of cut. The signal compensation and preprocessing module is used to perform zero-point drift correction, gravity and its induced torque component compensation, filtering and noise reduction, and outlier suppression on the force and torque components in the six-dimensional force sensing signals based on the robot end-effector pose and the installation pose relationship between the cutting tool and the six-dimensional force sensor, to obtain a compensated six-dimensional bone tissue cutting force sensing signal. The feature principal component mapping module is used to perform windowed statistical feature extraction on the six-dimensional bone tissue cutting force sensing signal and perform fixed mapping using standardized parameters and principal component analysis load matrices determined and locked during the training phase to obtain the real-time first principal force component. The bone layer confidence discrimination module is used to construct a bone layer scoring function based on the real-time first principal component, the real-time first principal component change rate, the cutting depth, the unit cutting energy, and the real-time prediction residual. It obtains the confidence scores for cortical bone, cancellous bone, and transition layer using a normalized exponential function, and determines the current bone layer state based on continuous window criteria and hysteresis switching rules. When the confidence score for cortical bone or cancellous bone meets the entry threshold within multiple consecutive windows and the normalized residual does not exceed the limit, it determines entry into the corresponding bone layer. When the confidence score for the transition layer increases, the normalized residual exceeds the limit, or the confidence score for the stable bone layer falls below the exit threshold, it determines entry into the safe transition layer state. The layer prediction module is used to select the bone tissue cutting force based on the bone layer state determined in the previous data window. A hierarchical prediction sub-model is used. When the current data window is in a stable cortical bone state or a stable cancellous bone state, the prediction sub-model corresponding to the stable bone layer is invoked, and a first principal component of prediction aligned with the current data window is generated based on the information from the previous data window. When the current data window is in a transitional state, the prediction sub-model corresponding to the most recently confirmed stable bone layer is invoked to generate a first principal component of prediction aligned with the current data window, and the first principal component of prediction is used only for real-time prediction residual monitoring. When the bone layer state of the current data window is determined to be a transitional state, or when the bone layer state of the current data window is inconsistent with the stable bone layer state on which the first principal component of prediction is based, the first principal component of prediction does not participate in closed-loop fusion control. An intra-layer fixed-weight fusion module is used to determine the intra-layer fixed fusion weight of the real-time first principal component based on the current stable bone layer state, prediction model error variance, real-time feedback fluctuation variance, bone layer confidence, and normalized residual. The intra-layer fixed fusion weight remains unchanged within the same stable bone layer control cycle and is updated when the stable bone layer state changes. In the transition layer state, the fusion weight of the real-time first principal component is set to 1. A fusion feedback generation module is used to fuse the real-time first principal component and the predicted first principal component according to the intra-layer fixed fusion weight to obtain a fused feedback first principal component. Compared to single sensor feedback, the fused feedback principal component has stronger noise suppression capability. Compared to pure model prediction, the fused feedback principal component can maintain its responsiveness to the actual cutting state. The residual safety monitoring module is used to construct a comprehensive risk index based on normalized residuals, peak force, depth of cut, feed rate, and feed acceleration, and to switch between normal control mode, speed limiting and load reduction mode, and stop protection mode according to the comprehensive risk index; that is, when the comprehensive safety index is in the normal range, closed-loop control is maintained; when the comprehensive safety index enters the medium risk range, the feed rate is reduced; when the comprehensive safety index exceeds the stop threshold, an alarm or stop command is triggered. The hierarchical fuzzy PID control module is used to call the fuzzy PID control rule corresponding to the current bone layer state based on the deviation between the fused feedback first principal component and the target first principal component and the comprehensive risk index, and output the feed speed command of the robot end effector.
[0033] Specifically, in the data acquisition module, the six-dimensional force / torque signal is acquired by a six-dimensional force sensor located between the robot end effector and the cutting tool. The robot end pose is provided by the robot controller or an external positioning system. The six-dimensional force / torque signal includes at least longitudinal cutting force, transverse cutting force, axial cutting force, and stress torque.
[0034] Specifically, in the signal compensation and preprocessing module, such as Figure 3 As shown, gravity component compensation includes: establishing a rotation matrix between the sensor coordinate system and the cutting tool coordinate system; calculating the force and torque components of the tool's gravity in the sensor coordinate system based on the cutting tool's mass, center of mass position, and robot end-effector posture; and subtracting the force, torque, and zero-point drift components from the original six-dimensional force / torque signal.
[0035] Assuming a fixed installation pose relationship exists between the sensor coordinate system and the tool coordinate system, the controller calculates the force and torque components of the tool's gravity in the sensor coordinate system based on the robot's end effector posture, the cutting tool's mass, and the position of its center of mass. It then combines this with zero-point drift correction to obtain the compensated bone cutting force signal.
[0036] Filtering and noise reduction include one or more of infinite impulse response low-pass filtering, median filtering, or sliding window smoothing, while outlier suppression includes peak removal based on amplitude thresholds, principal component residual thresholds, window statistics thresholds, or Huber loss criteria.
[0037] Specifically, in the principal component mapping module, the windowed statistical features include one or more of the following: mean, peak-to-peak value, root mean square value, variance, skewness, kurtosis, energy ratio in the principal force direction, and unit cutting energy. The windowed statistical feature vector is represented as follows: (1) in, The mean of the window. The root mean square value, Peak-to-peak value For window variance, For skewness, For ravine, The energy ratio in the dominant direction, The cutting energy is expressed as a unit.
[0038] Unit cutting energy Calculated based on the cutting force, feed rate, and equivalent cutting volume within the current window: (2) (3) in, For the first A data window, Force in the cutting direction, For feed rate, The equivalent cutting volume for the current window. For the equivalent cutting cross-sectional area, This represents the increment of the cutting depth within the window. To prevent positive numbers with a denominator of zero, This represents the sampling time interval between adjacent sampling moments.
[0039] like Figure 4 As shown, the principal component consistent mapping includes: using the mean, standard deviation, and PCA loading matrix determined during the training phase to perform a fixed mapping on the online windowed statistical features, specifically: (4) (5) (6) (7) in, For the first Statistical feature vectors extracted from each data window. and These are the feature mean vector and feature standard deviation vector determined during the training phase, respectively. This indicates element-wise division. It is a vector of all 1s. To standardize the feature vector, The number of training samples. For the first The standardized feature vector corresponding to each training sample To train the sample covariance matrix, Principal component loading matrix, It is an eigenvalue diagonal matrix. This is the first principal component load vector. As the first principal component in real time, the real-time principal component, the predicted principal component, and the target principal component are all constrained to the same fixed principal component space, thereby avoiding closed-loop control bias caused by online re-PCA or feature scale drift.
[0040] Based on this, the bone layer confidence level is calculated according to the real-time principal components, principal component change rate, cutting depth, unit cutting energy, and real-time-prediction residual. The bone layer status includes cortical bone, cancellous bone, and transition layer. The controller only confirms bone layer switching when the confidence level of a certain bone layer exceeds the corresponding threshold for multiple consecutive windows; when the confidence level is insufficient or the real-time-prediction residual exceeds the limit, the controller enters the transition layer safety control cycle.
[0041] Specifically, in the bone layer confidence discrimination module, the real-time prediction residual... and normalized residuals They are respectively: (8) (9) in, Indicates according to the first Information from the first data window, and based on the information from the first data window. The stable bone layer state determined by the first data window selects the corresponding prediction sub-model, generating a model that corresponds to the first... The first principal component of the prediction, time-aligned with the data window, when the... When a data window is in a transitional layer state, the predicted first principal component is generated using the prediction sub-model corresponding to the most recently confirmed stable bone layer. Indicates bone layer The standard deviation of the residuals is below; Calculate the first principal component in real time The amplitude is obtained by normalization. The normalized real-time force principal component amplitude, real-time force principal component change rate, cutting depth, unit cutting energy, and normalized residual are used to construct the normalized eigenvector. : (10) in, This is the normalized value of the real-time principal component amplitude. This is the normalized value of the rate of change of the principal components in real time. This is a normalized value for the depth of cut. This is the normalized value of the unit cutting energy.
[0042] Specifically, in the bone layer confidence discrimination module, such as Figure 6 As shown, the bone layer scoring functions include cortical bone scoring functions, cancellous bone scoring functions, and transition layer scoring functions, which are as follows: (11) (12) (13) in, , and These represent the scores for cortical bone, cancellous bone, and transition layer, respectively. to , to , to The corresponding scoring parameters are determined by constraint optimization based on training samples with bone layer category labels, with the goal of minimizing the sum of bone layer classification cross-entropy loss and regularization term, and remain fixed during online bone layer confidence calculation. Entropy is the confidence level. For the bone layer dominant septum, the expression is: (14) when At that time, it was stipulated (15) (16) in, Indicates the first Bone layer in each data window The normalized confidence level, ; and These represent the confidence scores for the largest and second-largest bone layers in the previous data window, respectively; the confidence scores for cortical bone, cancellous bone, and transition layer are obtained using a normalized exponential function. (17) in, , and They represent the first The confidence scores for cortical bone, cancellous bone, and transition layer are given in each data window. ; Equations (13) to (17) allow the confidence level of the transition layer to be simultaneously affected by the residual surge, the principal component change rate, the confidence entropy, and the bone layer dominance interval, enabling proactive identification of uncertain states at the cortical-cancellous bone junction.
[0043] Hysteresis switching rules include entry rules for stable bone layers and entry rules for transitional layers. For stable bone layers... The rules for entering the stable bone layer are as follows: (18) The transition layer entry rules are as follows: (19) and: (20) in, For the first The current bone layer status corresponding to each data window , Indicates a transition layer. To determine the number of consecutive decision windows, and They are respectively stable bone layers The entry confidence threshold and exit confidence threshold, The threshold for entering the transition layer. To enter the stable bone layer The upper limit of the normalized residual allowed at that time. To exit the stable bone layer The normalized residual threshold, The hysteresis width is defined by equations (18) to (20). Frequent shaking of the bone layer state in the junctional region can be avoided by using equations (18) to (20).
[0044] During bone layer switching, the controller first enters the transition layer safety control cycle, calling the transition layer safety prediction sub-model, the transition layer fixed fusion weight, and the transition layer fuzzy PID control rule base. After the cortical bone confidence or cancellous bone confidence satisfies the hysteresis switching condition for multiple consecutive windows, it then switches to the corresponding cortical bone or cancellous bone prediction sub-model, intra-layer fixed fusion weight, and fuzzy PID control rule base.
[0045] Specifically, in the stratified prediction module, the bone tissue cutting force stratified prediction sub-model includes a cortical bone prediction sub-model, a cancellous bone prediction sub-model, and a transition layer safety prediction sub-model. The cortical and cancellous bone prediction sub-models adopt the GA-SVM model, while the transition layer prediction sub-model adopts a conservative prediction model or a rate-limited safety model. Each bone layer corresponds to the prediction error variance, real-time feedback noise variance, and residual threshold. Based on these statistics, a bone layer weight table is generated to obtain the cortical bone weight, cancellous bone weight, and transition layer weight. The weights remain fixed within the same bone layer control cycle and are updated after the bone layer state meets the hysteresis switching condition, thus combining the stability of fixed weights with the adaptability of stratified control.
[0046] During the training phase, bone tissue cutting samples were collected from different individuals with different bone densities and cutting parameters, and labeled according to cortical bone, cancellous bone, and transitional layers. Each sample underwent compensation, filtering, feature extraction, and PCA mapping to obtain the dataset used for training the hierarchical prediction model.
[0047] GA-SVM prediction sub-models were established for cortical bone and cancellous bone respectively. A genetic algorithm was used to optimize the SVM penalty factor C and kernel function parameters. and insensitive loss parameters The fitness function is composed of prediction error, model complexity, and bone layer sample balance to improve the model's generalization ability. The transition layer can establish a conservative prediction sub-model or limit the prediction output within a safe range to prevent abrupt changes in feed rate at the bone layer interface.
[0048] The prediction model error variance for each bone layer is calculated on the validation set; the real-time feedback noise variance for each bone layer is calculated in the no-cut or stable cutting segments; and the upper and lower limits of the fusion weights for each bone layer are determined by combining the response speed constraint. This forms a bone layer weight table, which includes at least the cortical bone fixation and fusion weights, cancellous bone fixation and fusion weights, and transition layer fixation and fusion weights. This weight table is written to the controller before surgery or experimentation and is read according to the current bone layer status or updated during bone layer switching.
[0049] like Figure 5 As shown, each prediction sub-model is based on the first... The data window contains spindle speed, feed rate, depth of cut, real-time principal force components, real-time rate of change of principal force components, and unit cutting energy, used to predict the first... The principal components of each data window, with the input vector as follows: ;(twenty one) in, , and For the first Each data window displays the spindle speed, feed rate, and depth of cut. For the first The real-time principal components of each data window. This represents the rate of change of the principal components in real time. The unit cutting energy; Bone layer The corresponding input vector of the prediction sub-model; The cortical bone prediction sub-model and the cancellous bone prediction sub-model are support vector machine regression models with parameters optimized by a genetic algorithm. The prediction function is: ;(twenty two) in, To stabilize the bone layer The corresponding prediction sub-model outputs the predicted first principal component. and The first one obtained from training the support vector regression model Group of Lagrange multipliers, To stabilize the bone layer The corresponding number of support vectors, For kernel function, To stabilize the bone layer The corresponding prediction sub-model Support vectors, This is the bias term; the genetic algorithm is used to optimize the penalty parameters, kernel function parameters, and insensitive loss parameters of the support vector regression model. The overall fitness function of the genetic algorithm is: ;(twenty three) ;(twenty four) in, This is a penalty term for model complexity. , and These are the penalty factor, kernel function parameters, and insensitive loss parameters for the support vector machine regression model, respectively.
[0050] When the current bone layer state is a transitional layer state, the prediction sub-model corresponding to the most recently confirmed stable bone layer only outputs the predicted first principal component used to calculate the normalized residual. The predicted first principal component does not participate in the generation of the fusion feedback.
[0051] Specifically, in the intralayer fixed weight fusion module, the stable bone layer... The corresponding prediction model error variance and real-time feedback fluctuation variance They are respectively: (25) (26) in, To stabilize the bone layer The corresponding number of validation samples, and The first The real-time first principal components and the time-aligned predicted first principal components of each validation sample. To stabilize the bone layer The following is the number of calibration samples used to estimate the variance of real-time feedback fluctuations. To stabilize the bone layer The mean of the first principal component in real time; Stable bone layer Initial real-time feedback fusion weights for: (27) At the moment of transition of stable bone layer state The corrected intra-layer fixed fusion weights are: (28) in, , This is the normalized residual adjustment coefficient. and They are respectively stable bone layers Allowed real-time feedback fusion weight lower and upper limits; fixed fusion weights within a layer. The control remains unchanged within the same stable bone layer control cycle until the next stable bone layer state switch; Equation (27) makes the prediction model more biased towards real-time feedback when the error is larger, and reduces the proportion of real-time feedback when the noise of real-time feedback is larger; Equation (28) further incorporates bone layer confidence, transition layer confidence and normalized residual into the weight constraints. After entering the same bone layer control cycle, the controller remains unchanged. It remains unchanged until the conditions for the next bone layer hysteresis switching are met.
[0052] After obtaining the fixed fusion weights within the layer, the controller fuses the real-time force principal components with the predicted force principal components to obtain the fusion feedback force principal components. Specifically, in the fusion feedback generation module, the real-time first force principal components and the predicted first force principal components are fused under the same stable bone layer state; when the current bone layer state is inconsistent with the bone layer state corresponding to the generation of the predicted first force principal components, or when the current bone layer state is a transition layer state, the predicted first force principal components do not participate in the fusion of the current control cycle, and the fusion weight of the real-time first force principal components is set to 1; like Figure 7 As shown, if the current bone layer state is a stable bone layer and is consistent with the bone layer state when the predicted value was generated, then the stable bone layer... The first principal component fusion form is as follows: (29) in, No. The real-time first principal component corresponding to each data window According to the first The information in the first data window is generated and compared with the first... The first principal component of the prediction is aligned with each data window. The fusion weights are for the first principal components in real time, and ; If the current bone layer state is a transitional layer state, or if the current bone layer state is inconsistent with the bone layer state when the predicted value was generated, then: (30)
[0053] After obtaining the principal components of the fused feedback force, the controller further constructs the residual safety envelope. Specifically, in the residual safety monitoring module, the residual safety envelope is composed of comprehensive safety indices. Characterization: (31) Among them, normalized residuals As a residual risk item, For peak force risk items, This is a deep boundary crossing risk item. For the risk item of exceeding the speed limit, To mitigate the risk of exceeding the limit; , , , and These are the weighting coefficients for the residual risk term, peak force risk term, depth out-of-bounds risk term, velocity out-of-bounds risk term, and acceleration out-of-bounds risk term, respectively, and the sum of these weighting coefficients is 1. The peak force risk item, depth out-of-bounds risk item, velocity out-of-bounds risk item, and acceleration out-of-bounds risk item are respectively: (32) (33) (34) (35) (36) in, For the first The cutting force vector of each data window. Current bone layer status The preset upper limit threshold of platform cutting force, and These are the current bone layer states. The preset maximum and minimum process cutting depths, Current bone layer status The preset maximum feed rate, Current bone layer status The preset maximum feed acceleration, , For feed acceleration; when At this time, the system is in normal control mode; when When the system enters speed-limiting and load-reducing mode; when At that time, the system enters a safe deceleration to zero control phase; among which, The load reduction trigger threshold is a comprehensive risk indicator. This is the threshold for stopping the triggering of the comprehensive risk indicator, and , and This refers to the platform engineering control threshold.
[0054] Specifically, in the hierarchical fuzzy PID control module, the stable backbone layer... The corresponding target first principal component is determined by the mean of the real-time first principal component of the independent stable calibration window during the model development phase. The independent stable calibration window is a window with a clearly defined bone layer category, located within the corresponding stable bone layer cutting interval, and not in the bone layer transition region. The expression for the mean of the real-time first principal component is: (37) in, To stabilize the bone layer The corresponding set of independent stable calibration windows, The number of valid windows in the set of independent and stable calibration windows. For the first The real-time first principal component corresponding to each independent stable calibration window; the target first principal component of the transition layer. This is a conservative target value preset based on the safety calibration data of the transition layer; No. The first principal component of the objective for each control cycle is: (38) (39) Stable bone layer The control error and error change rate are as follows: (40) in, To control the cycle; Current stable bone layer The PID control gain is as follows: (41) (42) (43) in, , and They are respectively stable bone layers The corresponding proportional, integral, and differential reference gains; , and These are the proportional, integral, and differential gain corrections obtained through fuzzy inference based on control error, error rate of change, current stable bone layer confidence level, and comprehensive risk index; when At that time, the proportional, integral, and differential reference gains respectively adopt the preset transition layer reference gains. , and The proportional, integral, and differential gain corrections are each limited to a preset transition layer correction range. , and internal number; Current bone layer status Control output below for: (44) Robot end effector feed rate command for: (45) in, Current bone layer status The corresponding speed adjustment coefficient, The comprehensive risk mitigation coefficient, and These are the current bone layer states. The preset minimum and maximum feed rates are: ; The feed rate command also satisfies the following acceleration constraints: (46) in, Current bone layer status The maximum allowable feed acceleration is preset; when the current bone layer state is the transition layer state, the pure real-time first principal component feedback in the fusion feedback generation module and the conservative feed speed and feed acceleration limit corresponding to the transition layer are adopted.
[0055] In a preferred embodiment, to further improve the continuity of the bone layer status output, the controller can also perform recursive smoothing of the bone layer status: (47) (48) in, Smoothing factor (0 < < 1), This is the smoothed confidence vector. Equations (47)-(48) can be used as an alternative or supplementary scheme to the continuous window criterion, making the bone layer state output smoother and more continuous.
[0056] To suppress the impact of abnormal peaks on principal component mapping, window features can also employ Huber-type suppression functions: (49) Based on this function, anomaly suppression weights can be constructed: (50) (51) in, This is the threshold parameter for the Huber function. The normalized residual. Equations (49)-(51) can be used as alternative implementations of the outlier suppression step in step S3 or S4.
[0057] Furthermore, the controller can introduce a weighted soft-start function during bone layer switching: (52) in, This is the moment of bone layer switching. is the weight transition time constant of the current bone layer. Equation (52) can make the fusion weight smoothly transition from the old value to the new value when switching bone layers, avoiding control output oscillation caused by sudden weight changes.
[0058] Equations (47) to (52) above are optional implementations and do not change the basic closed-loop structure of the present invention, which includes bone layer confidence discrimination, layer prediction, intra-layer fixed weight fusion, residual safety envelope and layered fuzzy PID control.
[0059] To verify the effectiveness of the proposed bone layer confidence-driven robotic bone tissue layer cutting dual-source force principal component fusion safety control method, cutting experiments were conducted on an isolated porcine vertebral lamina platform.
[0060] I. Experimental Platform and Parameters The experimental platform consisted of a six-DOF collaborative robotic arm (Elite EC66, repeatability ±0.02 mm), a high-speed electric spindle (AM-BL2453AE, maximum power 257.5 W, maximum no-load speed 28,715 r / min), a single-groove medical carbide ball end mill (ball diameter 4 mm), and a six-dimensional force / torque sensor (Kunwei KWR75B, accuracy 0.1% FS, Fx / Fy / Fz range ±200 N, Mx / My / Mz range ±8 N·m). A total of 21 fresh, frozen, isolated porcine vertebral lamina samples were collected from 6 adult pigs (80–110 kg), covering the T12–L6 segments.
[0061] The spindle speed range is 4,000–7,000 r / min, the feed rate is 1–3 mm / s, and the depth of cut is 0.15–0.50 mm. The six-dimensional force / torque signal is sampled at 996 Hz and processed using a fourth-order causal Butterworth IIR filter with a 40 Hz cutoff frequency. The window length is 200 points (0.2008 s), the step size is 50 points (0.0502 s), and the window-level feature update frequency is 19.92 Hz. A fixed PCA space retains three principal components, explaining a cumulative variance of 92.6%, with PC1 serving as the primary closed-loop feedback variable.
[0062] II. Spatial Bone Region Recognition Effect On the test set, the accuracy of spatial bone region identification was 93.2%, and the macro-F1 score was 91.8%. The precision, recall, and F1 scores for the three types of bone regions are shown in Table 1. These results demonstrate that the bone layer confidence discrimination mechanism proposed in this invention can effectively distinguish between cortical bone regions, operational transition zones, and cancellous bone regions.
[0063] Table 1. Recognition performance of the internal clearance of the spatial bone region III. Tracking effect of stable bone area Four control strategies were set up: (1) single measurement feedback; (2) GA-SVR prediction feedback; (3) fixed weight fusion (the prediction branch weight is fixed at 0.2); (4) the complete method of the present invention. The tracking performance under the stable cutting state of cortical bone is shown in Table 2. The PC1-MAE of the complete method of the present invention is 3.61±0.13, which is 40.7% lower than that of single measurement feedback (6.09±0.18) and 14.9% lower than that of fixed weight fusion (4.24±0.16); the overshoot is reduced from 39.8% to 21.4%.
[0064] Table 2. Dimensionless mechanical state tracking performance of different control strategies during cortical bone stabilization cutting. The tracking performance under stable cutting conditions of cancellous bone is shown in Table 3. The PC1-MAE of the complete method is 0.60±0.03, which is 43.4% lower than that of single measurement feedback (1.06±0.05) and 15.5% lower than that of fixed weight fusion (0.71±0.03).
[0065] Table 3. Dimensionless mechanical state tracking performance of different control strategies during stabilizing cancellous bone cutting. With a single measurement feedback trajectory-level PC1-MAE of 100%, the complete method drops to approximately 59.3%, fixed-weight fusion to approximately 69.5%, and GA-SVR prediction feedback to approximately 87.7%. Figure 8 As shown.
[0066] IV. Control State Switching Effect To evaluate the effectiveness of the hysteresis switching mechanism in suppressing state jitter, three state update strategies were compared: (1) direct switching with a single threshold; (2) regional confidence discrimination (without hysteresis); and (3) regional confidence + hysteresis + continuous confirmation (in this invention). The results are shown in Table 4.
[0067] Table 4. Switching performance of different control state update strategies The method of this invention reduces unnecessary switching from 13 times to 4 times, a reduction of 69.2%, indicating that the hysteresis and continuous confirmation mechanism can effectively suppress state jitter at the bone layer boundary. The average confirmation delay increased from 0.036 s to 0.112 s, demonstrating a controllable trade-off between state stability and response speed.
[0068] V. Confidence Evolution During Continuous Bone Region Transition The confidence evolution of representative continuous cutting trajectories is as follows: Figure 9 As shown. Within the stable cortical bone region. Dominant; after entering the operational transition zone, Decline and Increase; after entering the cancellous bone area, Gradually gaining dominance. This result demonstrates that the confidence level can smoothly transition and the control state can switch in an orderly manner during continuous bone region transitions, verifying the practical effectiveness of the bone layer confidence-driven state management mechanism.
[0069] VI. Risk Response and Safety Protection Effectiveness During the continuous bone region conversion phase, the complete method of this invention was compared with the fixed-weight fusion strategy, and the risk response results are shown in Table 5. The number of platform force threshold exceeding, software risk supervision mode entry, Unload entry, and Stop entry events decreased from 10, 24, 20, and 4 to 4, 18, 16, and 2, respectively, and the cumulative duration of Unload decreased from 9.60 s to 6.72 s. The maximum physical cutting force decreased from 41.7 ± 3.8 N to 34.5 ± 3.1 N, and the maximum feed acceleration decreased from 9.8 ± 1.1 mm / s² to 6.4 ± 0.8 mm / s².
[0070] Table 5. Lamina-pairing events and overall descriptive event burden during continuous bone region transitions. In summary, this invention achieves accurate identification of bone layer status, significantly improves stable tracking accuracy, effectively suppresses state switching jitter, and strongly constrains engineering risks on an isolated porcine vertebral lamina platform, verifying the effectiveness and engineering feasibility of the bone layer confidence-driven robotic bone tissue layer cutting dual-source force principal component fusion safety control method.
[0071] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A safety control system for robot bone cutting force fusion driven by bone layer confidence, characterized in that, include: The data acquisition module is used to acquire six-dimensional force sensing signals, robot end-effector pose, and cutting parameters during the robot's layered cutting of bone tissue. The six-dimensional force sensing signals include force component signals along three mutually orthogonal directions and torque component signals around three mutually orthogonal axes. The cutting parameters include spindle speed, feed rate, and depth of cut. The signal compensation and preprocessing module is used to perform zero-point drift correction, gravity and its torque component compensation, filtering and noise reduction and outlier suppression on the six-dimensional force sensor according to the robot end pose and the installation pose relationship between the cutting tool and the six-dimensional force sensor, so as to obtain the compensated six-dimensional bone tissue cutting force sensor signal. The feature principal component mapping module is used to extract windowed statistical features from the six-dimensional bone tissue cutting force sensory signal, and to perform fixed mapping using standardized parameters and principal component analysis load matrix determined and locked during the training phase to obtain the real-time first force principal component. The bone layer confidence discrimination module is used to construct a bone layer scoring function based on the real-time first principal component, the real-time first principal component change rate, the cutting depth, the unit cutting energy, and the real-time prediction residual. It obtains the confidence of cortical bone, cancellous bone, and transition layer through a normalized exponential function, and determines the current bone layer status based on the continuous window criterion and the hysteresis switching rule. The stratified prediction module is used to select a stratified prediction sub-model for bone tissue cutting force based on the bone layer state determined in the previous data window. When the previous data window is a stable cortical bone state or a stable cancellous bone state, the prediction sub-model corresponding to the stable bone layer is invoked, and a predicted first principal component aligned with the current data window is generated based on the information from the previous data window. When the previous data window is a transitional layer state, the prediction sub-model corresponding to the most recently confirmed stable bone layer is invoked to generate a predicted first principal component aligned with the current data window, and the predicted first principal component is used only for real-time prediction residual monitoring. When the bone layer state in the current data window is determined to be a transitional layer state, or when the bone layer state in the current data window is inconsistent with the stable bone layer state on which the predicted first principal component was generated, the predicted first principal component does not participate in closed-loop fusion control. The intra-layer fixed weight fusion module is used to determine the intra-layer fixed fusion weight of the real-time first principal component based on the current stable bone layer state, prediction model error variance, real-time feedback fluctuation variance, bone layer confidence and normalized residual; the intra-layer fixed fusion weight remains unchanged within the same stable bone layer control cycle and is updated when the stable bone layer state changes; in the transition layer state, the fusion weight of the real-time first principal component is set to 1; The fusion feedback generation module is used to fuse the real-time first principal component and the predicted first principal component according to the fixed fusion weight within the layer to obtain the fusion feedback first principal component. The residual safety monitoring module is used to construct a comprehensive risk index based on normalized residual, peak force, depth of cut, feed rate and feed acceleration, and to switch between normal control mode, speed limit and load reduction mode and stop protection mode according to the comprehensive risk index. The hierarchical fuzzy PID control module is used to call the fuzzy PID control rule corresponding to the current bone layer state based on the deviation between the fused feedback first principal component and the target first principal component and the comprehensive risk index, and output the feed speed command of the robot end effector.
2. The safety control system for robot bone cutting force fusion driven by bone layer confidence according to claim 1, characterized in that, In the principal component mapping module, the windowed statistical features include one or more of the following: mean, peak-to-peak value, root mean square value, variance, skewness, kurtosis, energy ratio in the principal direction, and unit cutting energy. The windowed statistical feature vector is represented as follows: ; in, The mean of the window. The root mean square value, Peak-to-peak value For window variance, For skewness, For ravine, The energy ratio in the dominant direction, The unit cutting energy; The unit cutting energy Calculated based on the cutting force, feed rate, and equivalent cutting volume within the current window: ; ; in, For the first A data window, Force in the cutting direction, For feed rate, The equivalent cutting volume for the current window. For the equivalent cutting cross-sectional area, This represents the increment of the cutting depth within the window. To prevent positive numbers with a denominator of zero, This represents the sampling time interval between adjacent sampling moments.
3. The bone layer confidence-driven robot bone cutting force fusion safety control system according to claim 2, characterized in that, In the feature principal component mapping module, the principal component consistent mapping includes: using the mean, standard deviation, and PCA loading matrix determined during the training phase to perform a fixed mapping on the online windowed statistical features, specifically: ; ; ; ; in, For the first Statistical feature vectors extracted from each data window. and These are the feature mean vector and feature standard deviation vector determined during the training phase, respectively. This indicates element-wise division. It is a vector of all 1s. To standardize the feature vector, The number of training samples. For the first The standardized feature vector corresponding to each training sample To train the sample covariance matrix, Principal component loading matrix, It is an eigenvalue diagonal matrix. This is the first principal component load vector. It is the primary principal component in real time.
4. The bone layer confidence-driven robot bone cutting force fusion safety control system according to claim 3, characterized in that, In the bone layer confidence discrimination module, the real-time prediction residual and normalized residuals They are respectively: ; ; in, Indicates according to the first Information from the first data window, and based on the information from the first data window. The stable bone layer state determined by the first data window selects the corresponding prediction sub-model, generating a model that corresponds to the first... The first principal component of the prediction, time-aligned with the data window, when the... When a data window is in a transitional layer state, the predicted first principal component is generated using the prediction sub-model corresponding to the most recently confirmed stable bone layer. Indicates bone layer The standard deviation of the residuals is below; Calculate the first principal component in real time The amplitude is obtained by normalization. The normalized real-time first principal component amplitude, real-time first principal component change rate, cutting depth, unit cutting energy, and normalized residual are used to construct the normalized eigenvector: ; in, This is the normalized value of the real-time principal component amplitude. This is the normalized value of the rate of change of the principal components in real time. This is a normalized value for the depth of cut. This is the normalized value of the unit cutting energy.
5. The safety control system for robot bone cutting force fusion driven by bone layer confidence according to claim 4, characterized in that, The bone layer confidence discrimination module includes three bone layer scoring functions: cortical bone scoring function, cancellous bone scoring function, and transition layer scoring function, which are as follows: ; ; ; in, , and These represent the scores for cortical bone, cancellous bone, and transition layer, respectively. to , to , to The corresponding scoring parameters are determined by constraint optimization based on training samples with bone layer category labels, with the goal of minimizing the sum of bone layer classification cross-entropy loss and regularization term, and remain fixed during online bone layer confidence calculation. Entropy is the confidence level. For the bone layer dominant septum, the expression is: ; when At that time, it was stipulated ; ; in, Indicates the first Bone layer in each data window The normalized confidence level, , Indicates cortical bone, Indicates cancellous bone, Indicates a transition layer; and These represent the confidence scores for the largest and second-largest bone layers in the previous data window, respectively. The confidence scores for cortical bone, cancellous bone, and transition layer were obtained using a normalized exponential function. ; in, , and They represent the first The confidence scores for cortical bone, cancellous bone, and transition layer are given in each data window. ; The hysteresis switching rules include stable bone layer entry rules and transition layer entry rules; for stable bone layers... The stable bone layer entry rule is as follows: ; The transition layer entry rule is as follows: ; and: ; in, For the first The current bone layer status corresponding to each data window , Indicates a transition layer. To determine the number of consecutive decision windows, and They are respectively stable bone layers The entry confidence threshold and exit confidence threshold, The threshold for entering the transition layer. To enter the stable bone layer The upper limit of the normalized residual allowed at that time. To exit the stable bone layer The normalized residual threshold, The hysteresis width.
6. The bone layer confidence-driven robot bone cutting force fusion safety control system according to claim 5, characterized in that, The hierarchical prediction module includes a cortical bone prediction sub-model and a cancellous bone prediction sub-model; each prediction sub-model is based on the first... The data window contains spindle speed, feed rate, depth of cut, real-time first principal force component, real-time rate of change of the first principal force component, and unit cutting energy, predicting the first... The first principal component of each data window, with the input vector being: ; in, , and For the first Each data window displays the spindle speed, feed rate, and depth of cut. For the first The real-time principal components of each data window. This represents the rate of change of the principal components in real time. The unit cutting energy; To stabilize the bone layer The corresponding input vector of the prediction sub-model; The cortical bone prediction sub-model and the cancellous bone prediction sub-model are support vector machine regression models with parameters optimized by a genetic algorithm, and the prediction function is: ; in, To stabilize the bone layer The corresponding prediction sub-model outputs the predicted first principal component. and The first one obtained from training the support vector regression model Group of Lagrange multipliers, To stabilize the bone layer The corresponding number of support vectors, For kernel function, To stabilize the bone layer The corresponding prediction sub-model Support vectors, This is the bias term; the genetic algorithm is used to optimize the penalty parameters, kernel function parameters, and insensitive loss parameters of the support vector regression model. When the current bone layer state is a transitional layer state, the prediction sub-model corresponding to the most recently confirmed stable bone layer only outputs the predicted first principal component used to calculate the normalized residual. The predicted first principal component does not participate in the generation of the fusion feedback.
7. The bone layer confidence-driven robot bone cutting force fusion safety control system according to claim 6, characterized in that, In the intralayer fixed weight fusion module, the stable bone layer The corresponding prediction model error variance and real-time feedback fluctuation variance They are respectively: ; ; in, To stabilize the bone layer The corresponding number of validation samples, and The first The real-time first principal components and the time-aligned predicted first principal components of each validation sample. To stabilize the bone layer The following is the number of calibration samples used to estimate the variance of real-time feedback fluctuations. To stabilize the bone layer The mean of the first principal component in real time; Stable bone layer Initial real-time feedback fusion weights for: ; At the moment of transition of stable bone layer state The corrected intra-layer fixed fusion weights are: ; in, , This is the normalized residual adjustment coefficient. and They are respectively stable bone layers Allowed real-time feedback fusion weight lower and upper limits; fixed fusion weights within a layer. It remains unchanged within the same stable bone layer control cycle until the next stable bone layer state switch.
8. The bone layer confidence-driven robot bone cutting force fusion safety control system according to claim 7, characterized in that, In the fusion feedback generation module, the real-time first principal component and the predicted first principal component are fused under the same stable bone layer state; If the current bone layer state is a stable bone layer and is consistent with the bone layer state when the predicted value was generated, the first principal component fusion form is: ; in, No. The real-time first principal component corresponding to each data window According to the first The information in the first data window is generated and compared with the first... The first principal component of the prediction is aligned with each data window. The fusion weights are for the first principal components in real time, and ; If the current bone layer state is a transitional state, or if the current bone layer state is inconsistent with the bone layer state at the time of generating the predicted value, the first principal component of the prediction will not participate in the fusion of the current control cycle. 。 9. A safety control system for robot bone cutting force fusion driven by bone layer confidence according to claim 8, characterized in that, The comprehensive risk indicators in the residual safety monitoring module Characterized as: ; Among them, normalized residuals As a residual risk item, For peak force risk items, This is a deep boundary crossing risk item. For the risk item of exceeding the speed limit, To mitigate the risk of exceeding the limit; , , , and These are the weighting coefficients for the residual risk term, peak force risk term, depth out-of-bounds risk term, velocity out-of-bounds risk term, and acceleration out-of-bounds risk term, respectively, and the sum of these weighting coefficients is 1. The peak force risk item, depth out-of-bounds risk item, velocity out-of-bounds risk item, and acceleration out-of-bounds risk item are respectively: ; ; ; ; ; in, For the first The cutting force vector of each data window. Current bone layer status The preset upper limit threshold of platform cutting force, and These are the current bone layer states. The preset maximum and minimum process cutting depths, Current bone layer status The preset maximum feed rate, Current bone layer status The preset maximum feed acceleration, , For feed acceleration; when At this time, the system is in normal control mode; when When the system enters speed-limiting and load-reducing mode; when At that time, the system enters a safe deceleration to zero control phase; among which, The load reduction trigger threshold is a comprehensive risk indicator. This is the threshold for stopping the triggering of the comprehensive risk indicator, and .
10. A safety control system for robot bone cutting force fusion driven by bone layer confidence according to claim 9, characterized in that, In the hierarchical fuzzy PID control module, the stable bone layer The corresponding target first principal component is determined by the mean of the real-time first principal component of the independent stable calibration window during the model development phase. The independent stable calibration window is a window with a clearly defined bone layer category, located within the corresponding stable bone layer cutting interval, and not in the bone layer transition region. The expression for the mean of the real-time first principal component is: ; in, To stabilize the bone layer The corresponding set of independent stable calibration windows, The number of valid windows in the set of independent and stable calibration windows. For the first The real-time first principal component corresponding to each independent stable calibration window; the target first principal component of the transition layer. This is a conservative target value preset based on the safety calibration data of the transition layer; No. The first principal component of the objective for each control cycle is: ; ; Stable bone layer The control error and error change rate are as follows: ; in, To control the cycle; Current stable bone layer The PID control gain is as follows: ; ; ; in, , and They are respectively stable bone layers The corresponding proportional, integral, and differential reference gains; , and These are the proportional, integral, and differential gain corrections obtained through fuzzy inference based on control error, error rate of change, current stable bone layer confidence level, and comprehensive risk index; when At that time, the proportional, integral, and differential reference gains respectively adopt the preset transition layer reference gains. , and The proportional, integral, and differential gain corrections are each limited to a preset transition layer correction range. , and Inside; Current bone layer status Control output below for: ; Robot end effector feed rate command for: ; in, Current bone layer status The corresponding speed adjustment coefficient, The comprehensive risk mitigation coefficient, and These are the current bone layer states. The preset minimum and maximum feed rates are: ; The feed rate command also satisfies the following acceleration constraints: ; in, Current bone layer status The preset maximum allowable feed acceleration.
