Anti-jitter control algorithm based on double-threshold judgment

Abstract

The invention belongs to the technical field of computers, and particularly relates to an anti-jitter control algorithm based on double-threshold judgment. According to the algorithm, by introducing two judgment mechanisms of a man-machine interaction force threshold value and a time threshold value, the intention of a doctor can be intelligently recognized, a follow-up mode and a holding mode can be automatically switched, mechanical arm movement shaking caused by invalid operation or hand vibration is effectively reduced, and the precision, stability and safety of an operation are improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and more particularly, relates to an anti-shake control algorithm based on dual-threshold judgment. Background Art

[0002] With the rapid development of minimally invasive surgical techniques and surgical robotic systems, doctors are increasingly demanding precision, response speed, and system stability during surgeries. This is especially true in high-precision procedures such as neurosurgery, ophthalmology, and orthopedics, where doctors must manipulate robotic arms with precise movements to perform complex operations. Therefore, effectively controlling the stable motion of robotic arms to avoid malfunctions caused by hand tremors, ineffective contact, or fatigue has become a key technical challenge in the field of human-machine interaction control.

[0003] Most existing anti-shake control technologies use a single threshold judgment mechanism, which determines the doctor's operating intention by setting a fixed interaction force threshold or grip time threshold, thereby controlling the start, stop, and state switching of the robotic arm. This type of method is simple to implement, low-cost, and has a fast system response speed, and therefore has been widely used in early surgical assistance robotic arm equipment. However, in actual surgical operations, this single threshold judgment method has significant limitations: its anti-interference ability is weak, and it is easy to misjudge the doctor's unintentional contact or slight hand tremor as a valid instruction, causing the false triggering of the tracking mode; conversely, when the doctor's operating force is weak or lacks continuity due to fatigue, it may be ignored by the system, resulting in the robotic arm failing to respond to the operating intention in a timely manner, seriously affecting the stability and continuity of the operation.

[0004] In addition, although some systems have introduced auxiliary means such as image recognition and posture perception, and use cameras and artificial intelligence algorithms to identify changes in doctors' gestures to determine operational intentions, thereby improving the interaction dimension, due to the complexity of image processing itself and its high sensitivity to the external environment, recognition errors often occur in actual applications due to factors such as occlusion and lighting changes. The response delay problem is also prominent and cannot meet real-time requirements.

[0005] In terms of control mode design, traditional devices often use "passive following" or "active locking" approaches. The former relies entirely on the surgeon's hand to drive the robotic arm's movement, making active anti-shake control impossible. The latter uses motor braking to maintain the robotic arm's posture, but this can lead to response delays when the surgeon needs to quickly resume operation, making it unsuitable for high-frequency dynamic control. Furthermore, existing systems generally lack ergonomic optimization in hardware interaction design. The rigid connection of the operating handle easily amplifies micro-movements of the hand, causing force feedback distortion. Long-term use can also easily increase the surgeon's physical burden, affecting the continuity of surgery.

[0006] More critically, most current systems are still stuck in the fixed parameter setting stage, lacking the ability to adaptively identify and dynamically learn individual physician characteristics. For example, under varying operating styles, usage strength, or fatigue states of different physicians, traditional systems are unable to automatically adjust control thresholds based on real-time interactive data, which can easily lead to misjudgments or untimely mode switching. Furthermore, in terms of multi-degree-of-freedom collaborative control, existing control logic is mostly designed for single-axis or simple action scenarios, and has yet to develop a jitter coupling analysis and intelligent control mechanism for robotic arms with more than five degrees of freedom, making it difficult to support the stable execution of highly complex, continuous, and multi-directional concurrent operations within space.

[0007] In summary, current anti-shake control technology has shortcomings in terms of anti-interference ability, mode discrimination accuracy, real-time response efficiency, human-machine interaction experience, and multi-degree-of-freedom collaborative control. It cannot meet the comprehensive requirements of stability, safety, and intelligence in modern high-precision surgical environments. Therefore, there is an urgent need for a human-machine interaction control method that can combine a dual judgment mechanism of force threshold and time threshold, and possess high-frequency data acquisition capabilities, adaptive mode switching logic, and multi-modal perception feedback. This method can achieve more precise, stable, and smooth human-machine collaboration, significantly improving the practicality and safety performance of surgical robotic arm systems. Summary of the Invention

[0008] To this end, a dual-threshold anti-shake control algorithm is needed. This algorithm is suitable for surgical robotic systems with human-machine interaction capabilities, particularly in precision medical scenarios where the human hand pulls the robotic arm to achieve follow-up operations. By incorporating two judgment mechanisms—a "human-machine interaction force threshold" and a "time threshold"—this algorithm intelligently identifies the surgeon's intent and automatically switches between "follow-up mode" and "hold mode." This effectively reduces robotic arm jitter caused by ineffective manipulation or hand tremors, improving surgical precision, stability, and safety.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] A dual-threshold-based anti-shake control algorithm is used to control a five-degree-of-freedom follower robotic arm. A force sensor is placed between the arm and the robotic arm, and a small handle is designed at the front end of the support platform. The human hand pulls the small handle to pull the robotic arm, causing it to follow the human hand's movement. The anti-shake control algorithm includes the following steps:

[0011] Step 1: Initialize the human-computer interaction force threshold, the time threshold for holding the small handle, and the human-computer interaction force duration threshold, load the CNN-SVM fusion electromyography fatigue recognition model, and initialize it to the follow-up mode;

[0012] Step 2: Collect the operator's upper arm sEMG signal and input it into the CNN-SVM fusion electromyography fatigue recognition model for feature extraction, and use the SVM classifier to identify the fatigue state;

[0013] Step 3: Determine whether the fatigue state has reached the set threshold. If so, increase the human-computer interaction force threshold, delay the switching time of the handle holding time threshold or the interaction force duration threshold, and remind the user to take a break. Otherwise, go to step 4.

[0014] Step 4: Determine whether the human-computer interaction force is greater than a threshold value based on the human-computer interaction force detected by the force sensor. If yes, go to step 5; otherwise, go to step 6.

[0015] Step 5: Determine whether the time the hand holds the small handle is less than a threshold. If so, determine whether the duration of the human-computer interaction force is greater than a threshold. If so, switch to the hold mode; if not, return to the previous state and re-determine.

[0016] Step 6: Determine whether the time the hand holds the small handle exceeds the threshold. If so, determine whether the duration of the human-computer interaction force is greater than the threshold. If so, switch to follow-up mode; if not, return to re-determine, and continuously update the status and mode in real time to ensure timely response to doctor's needs.

[0017] This technical solution is further optimized, in the initialization phase of step 1:

[0018] (1) Setting the human-computer interaction force threshold Fth;

[0019] (2) Setting the grip time threshold Tgrip;

[0020] (3) Setting the interaction force duration threshold Tdur;

[0021] (4) The system initial mode is set to "follow-up mode".

[0022] This technical solution is further optimized, in the signal acquisition and preprocessing stage in step 2:

[0023] (1) Attach surface electromyography electrodes to the operator's upper arm to collect electromyographic signals in real time;

[0024] (2) The raw sEMG signal is transmitted to the processing unit through the acquisition module for preprocessing operations such as filtering, denoising, and normalization;

[0025] (3) Synchronously collect human-computer interaction force signals and the state of holding the small handle.

[0026] This technical solution is further optimized, in the electromyographic fatigue identification stage in step 2:

[0027] (1) The processed electromyographic signal is input into CNN to extract features and then sent to SVM classifier for judgment;

[0028] (2) If the recognition result is "fatigue", the system enters fatigue buffer mode and executes the following sub-process:

[0029] a) Increase the Fth threshold and extend the Tgrip and Tdur time determination range;

[0030] b) Pause the movement of the robot arm and maintain the current position;

[0031] c) Activate the sound and light prompt function to remind the doctor of fatigue status;

[0032] (3) If the recognition result is "non-fatigue", the normal dual-threshold judgment control logic is entered.

[0033] This technical solution is further optimized. In step 2, a force sensor is installed between the robotic arm and the supporting structure to collect the interactive force applied by the doctor by holding the small handle in real time; the system continuously monitors the changes in the interactive force through high-frequency sampling and determines whether the current force value is higher or lower than the set threshold Fth.

[0034] This technical solution is further optimized, the switching logic in the follow-up mode:

[0035] (1) If it is detected that the current human-computer interaction force is lower than Fth;

[0036] (2) At the same time, the time of holding the small handle exceeds Tgrip;

[0037] (3) The interaction force remains below the threshold for a period exceeding Tdur;

[0038] (4) The system automatically switches to "hold mode" and the robotic arm maintains its current position to avoid disturbances caused by invalid contact.

[0039] This technical solution is further optimized, the switching logic in the hold mode:

[0040] (1) If the human-machine interaction force is detected to be higher than Fth;

[0041] (2) The time of holding the small handle is less than Tgrip;

[0042] (3) At the same time, the interaction force remains above the threshold for a period exceeding Tdur;

[0043] (4) The system switches to "follow-up mode" so that the robotic arm moves in real time following the movements of the human hand.

[0044] This technical solution is further optimized, and the structure of the CNN-SVM fusion electromyography fatigue recognition model is as follows:

[0045] Step 1: CNN feature extraction network training phase

[0046] (1) Sample construction: Collect EMG training samples and corresponding labels. The samples include: original sEMG data + fatigue / non-fatigue labels;

[0047] (2) Network structure:

[0048] (a) Construct a CNN-Softmax network for the training phase;

[0049] (b) The CNN structure includes: convolutional layer, pooling layer, and fully connected layer; the output uses the Softmax activation function for preliminary classification; the Adam optimizer and the cross entropy loss function are used for model optimization training; the Softmax output is trained using the cross entropy loss function: Where y_i is the true label, is the Softmax output.

[0050] (c) Model saving: After training is completed, the CNN structure is removed from the Softmax layer and saved as a "feature extractor" model;

[0051] Step 2: Statistical learning classifier fusion training phase

[0052] (a) Feature extraction: Use the trained CNN model to extract features from the training and test samples to obtain high-dimensional deep semantic feature vectors. CNN performs feature extraction through convolution operations, and the calculation formula is as follows:

[0053] Convolutional layer: Z^(l)=f(W^(l)*X^(l-1)+b^(l)), where Z^(l) is the output feature map of the lth layer, W^(l) is the convolution kernel weight, * represents the convolution operation, b^(l) is the bias term, and f(x) is the activation function (such as ReLU: f(x)=max(0,x)).

[0054] Pooling layer: P^(l) = Pooling(Z^(l))

[0055] The fully connected layer outputs a feature vector: F = φ(Z^(L)), where φ represents the feature vector formed by flattening and feeding it into the fully connected network.

[0056] (b) Statistical learning classifier training: The extracted features and labels are used to train a traditional machine learning classifier. SVM is used as the main classifier to build a CNN-SVM fusion model. The feature vector F extracted by CNN is input into the SVM classifier, and its discriminant function is:

[0057] y=sign(w^T*F+b)

[0058] Among them, w is the weight vector, b is the bias, and sign(·) is the sign function.

[0059] If the kernel function form is used (taking the radial basis kernel as an example):

[0060] y=sign(Σα_i y_i K(F,F_i)+b)

[0061] Where K(F,F_i)=exp(-γ||F-F_i||^2)

[0062] Step 3, final model structure:

[0063] (a) Input layer: original EMG signal;

[0064] (b) Feature extraction layer: CNN network with the Softmax layer removed;

[0065] (c) Classifier layer: SVM;

[0066] (d) Output layer: fatigue / non-fatigue binary classification results.

[0067] Different from the existing technology, the above technical solution has the following beneficial effects:

[0068] 1) Dual threshold judgment enhances the ability to identify false touches and avoids frequent false switching;

[0069] 2) Adaptive following logic improves operational sensitivity and stability;

[0070] 3) The real-time feedback mechanism meets the response speed requirements of high-precision surgical operations;

[0071] 4) Reduce the burden on doctors, reduce hand fatigue, and improve the ability to continue surgery;

[0072] 5) Adaptable to a variety of application scenarios, especially suitable for minimally invasive surgical systems with high precision requirements in neurosurgery, orthopedics, otolaryngology, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is the structural diagram of the five-axis robotic arm;

[0074] Figure 2 This is a flow chart of the anti-shake control algorithm based on dual threshold judgment;

[0075] Figure 3 Schematic diagram of the CNN-SVM fusion electromyography fatigue recognition model. DETAILED DESCRIPTION

[0076] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0077] See Figure 1 As shown in the figure, the present invention designs a five-degree-of-freedom follower robotic arm. A force sensor is installed between the arm frame and the robotic arm. The base has four movable wheels, allowing it to stand independently. It has two operating modes: hold and follow. In the follow mode, the electric actuators activate, and the robotic arm moves upward via counterweights at each joint, causing the arm holder to push the surgeon's arm from below. Friction between the surgeon's arm and the arm frame causes the arm frame to move and follow the surgeon's arm. In the hold mode, the arm frame maintains its position by locking the electric actuators and supports the weight of the surgeon's arm. This locking of the electric actuators effectively addresses fatigue and hand tremors caused by long hours of work. Each state is automatically switched by analyzing signals from the force sensors and encoders in each joint. In all patients, dynamic mode is primarily selected for tumor resection, while static mode is selected for gripping or joint operations. Supporting the arm reduces shoulder and arm tension when holding lightweight surgical instruments such as scalpels, thereby reducing physical fatigue for the surgeon. Apart from this, this machine also offers wider surgical angles for surgical tools as the surgeon’s arms can be fixed in place.

[0078] A force sensor is set between the arm and the robotic arm, and a small handle is designed at the front end of the support platform. The small handle is pulled by the human hand to pull the robotic arm, so that the robotic arm moves with the human hand.

[0079] See also Figure 2 The present invention preferably adopts an anti-shake control algorithm based on dual threshold judgment, and the specific steps are as follows:

[0080] Step 1: System initialization phase:

[0081] (1) Setting the human-computer interaction force threshold Fth;

[0082] (2) Setting the grip time threshold Tgrip;

[0083] (3) Setting the interaction force duration threshold Tdur;

[0084] (4) Set the fatigue threshold fatigue_th;

[0085] (5) Loading the CNN-SVM fusion electromyography fatigue recognition model;

[0086] (6) The system initial mode is set to "follow-up mode".

[0087] See Figure 3 Figure 2 shows a schematic diagram of a CNN-SVM fusion electromyography fatigue recognition model, which is used to intelligently identify the degree of muscle fatigue during doctor operations. This model uses a deep convolutional neural network to extract features from surface electromyography (sEMG) signals and a support vector machine (SVM) to efficiently classify and judge the extracted feature vectors, thereby improving the accuracy and robustness of fatigue detection.

[0088] The model structure includes the following two stages:

[0089] Step 1: CNN feature extraction network training phase

[0090] (1) Sample construction: Collect EMG training samples and corresponding labels. The samples include: original sEMG data + fatigue / non-fatigue labels;

[0091] (2) Network structure:

[0092] (a) Construct a CNN-Softmax network for the training phase;

[0093] (b) The CNN structure includes: convolutional layer, pooling layer, and fully connected layer; the output uses the Softmax activation function for preliminary classification; the Adam optimizer and the cross entropy loss function are used for model optimization training; the Softmax output is trained using the cross entropy loss function: Where y_i is the true label, is the Softmax output.

[0094] (c) Model saving: After training is completed, the CNN structure is removed from the Softmax layer and saved as a "feature extractor" model;

[0095] Step 2: Statistical learning classifier fusion training phase

[0096] (a) Feature extraction: Use the trained CNN model to extract features from the training and test samples to obtain high-dimensional deep semantic feature vectors. CNN performs feature extraction through convolution operations, and the calculation formula is as follows:

[0097] Convolutional layer: Z^(l)=f(W^(l)*X^(l-1)+b^(l)), where Z^(l) is the output feature map of the lth layer, W^(l) is the convolution kernel weight, * represents the convolution operation, b^(l) is the bias term, and f(x) is the activation function (such as ReLU: f(x)=max(0,x)).

[0098] Pooling layer: P^(l) = Pooling(Z^(l))

[0099] The fully connected layer outputs a feature vector: F = φ(Z^(L)), where φ represents the feature vector formed by flattening and feeding it into the fully connected network.

[0100] (b) Statistical learning classifier training: The extracted features and labels are used to train a traditional machine learning classifier. SVM is used as the main classifier to build a CNN-SVM fusion model. The feature vector F extracted by CNN is input into the SVM classifier, and its discriminant function is:

[0101] y=sign(w^T*F+b)

[0102] Among them, w is the weight vector, b is the bias, and sign(·) is the sign function.

[0103] If the kernel function form is used (taking the radial basis kernel as an example):

[0104] y=sign(Σα_i y_i K(F,F_i)+b)

[0105] Where K(F,F_i)=exp(-γ||F-F_i||^2)

[0106] Step 3, final model structure:

[0107] (a) Input layer: original EMG signal;

[0108] (b) Feature extraction layer: CNN network with the Softmax layer removed;

[0109] (c) Classifier layer: SVM;

[0110] (d) Output layer: fatigue / non-fatigue binary classification results.

[0111] Step 2: Signal acquisition and preprocessing stage:

[0112] (1) Attach surface electromyography electrodes to the doctor's upper arm to collect electromyographic signals in real time;

[0113] (2) The raw sEMG signal is transmitted to the processing unit through the acquisition module for preprocessing operations such as filtering, denoising, and normalization;

[0114] (3) Synchronously collect human-computer interaction force signals and the state of holding the small handle.

[0115] Step 3: EMG fatigue identification stage:

[0116] (1) The processed electromyographic signal is input into CNN to extract features and then sent to SVM classifier for judgment;

[0117] (2) If the recognition result is "fatigue", the system enters fatigue buffer mode and executes the following sub-process:

[0118] a) Increase the Fth threshold and extend the Tgrip and Tdur time determination range;

[0119] b) Pause the movement of the robot arm and maintain the current position;

[0120] c) Activate the sound and light prompt function to remind the doctor of fatigue status;

[0121] (3) If the recognition result is "non-fatigue", the normal dual-threshold judgment control logic is entered.

[0122] Step 4: Judgment logic in follow-up mode:

[0123] (1) If it is detected that the current human-computer interaction force is lower than Fth;

[0124] (2) The time of holding the small handle exceeds Tgrip;

[0125] (3) At the same time, the interaction force remains below the threshold for a period exceeding Tdur;

[0126] (4) The system automatically switches to "hold mode" and locks the current position of the robotic arm.

[0127] Step 5: Judgment logic in hold mode:

[0128] (1) If the human-machine interaction force is detected to be higher than Fth;

[0129] (2) The time of holding the small handle is less than Tgrip;

[0130] (3) At the same time, the interaction force remains above the threshold for a period exceeding Tdur;

[0131] (4) The system automatically switches to "follow-up mode" and the robotic arm begins to follow the movement of the human hand.

[0132] Step 6: Status update and control output phase:

[0133] (1) The system executes myoelectric fatigue recognition and control logic judgment cyclically in each control cycle;

[0134] (2) Output mode status and control signals to the execution module in real time to achieve dynamic mode switching and motion control;

[0135] (3) At the same time, the historical status is recorded in the system log for subsequent analysis and optimization.

[0136] First, the present invention improves the quality of surgery by providing doctors with anti-shake support during surgery, and can promptly avoid the adverse effects on doctors and patients caused by the high fatigue of doctors who focus on surgery for a long time, thereby safeguarding medical health. Secondly, the high-precision robotic arms and advanced control systems equipped with medical exoskeletons can flexibly interact with doctors and provide the motion assistance they need. This is especially critical for surgeries that require extremely high precision, such as neurosurgery, cardiac surgery, and ophthalmic surgery. By improving surgical accuracy, medical exoskeletons can reduce errors and complications during surgery, thereby significantly reducing surgical risks. In addition, the fatigue detection algorithm model built into the medical exoskeleton system can monitor the doctor's fatigue in real time, issue timely warnings for problems that may be caused by excessive fatigue, and provide decision-making support for doctors. This not only improves the safety of surgery, but also provides doctors with valuable time to make adjustments during complex operations.

[0137] Medical exoskeleton systems provide support and strength assistance, helping medical staff maintain correct posture during surgery and reducing muscle fatigue and discomfort caused by prolonged standing and repetitive movements. They also provide assurance during surgery—the exoskeleton system's design reduces hand tremors during operation, thereby improving surgical stability and precision. This not only reduces the physical and psychological burden on doctors, but also reduces the burden on them during surgery. In these ways, medical exoskeleton systems not only reduce the workload of medical staff, but also improve their work efficiency and quality of life, ultimately promoting the sustainable development of the entire medical industry.

[0138] It should be noted that, in this document, relational terms such as first and second, etc., are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "include," "comprise," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. Without further limitation, elements defined by the phrase "include..." or "comprising..." do not exclude the presence of additional elements in the process, method, article, or terminal device comprising the elements. Furthermore, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the number itself; "above," "below," "within," etc., are understood to include the number itself.

[0139] Although the above embodiments have been described, those skilled in the art may make additional changes and modifications to these embodiments once they know the basic creative concepts. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present invention.

Claims

1. An anti-shake control algorithm based on dual threshold judgment, characterized in that: The anti-shake control algorithm is used to control a five-degree-of-freedom follower robotic arm. A force sensor is set between the arm and the robotic arm, and a small handle is designed at the front end of the support platform. The human hand pulls the small handle to pull the robotic arm, making the robotic arm move with the human hand. The anti-shake control algorithm includes the following steps: Step 1: Initialize the human-computer interaction force threshold, the time threshold for holding the small handle, and the human-computer interaction force duration threshold, load the CNN-SVM fusion electromyography fatigue recognition model, and initialize it to the follow-up mode; Step 2: Collect the operator's upper arm sEMG signal and input it into the CNN-SVM fusion electromyography fatigue recognition model for feature extraction, and use the SVM classifier to identify the fatigue state; Step 3: Determine whether the fatigue state has reached the set threshold. If so, increase the human-computer interaction force threshold, delay the switching time of the handle holding time threshold or the interaction force duration threshold, and remind the user to take a break. Otherwise, go to step 4. Step 4: Determine whether the human-computer interaction force is greater than a threshold value based on the human-computer interaction force detected by the force sensor. If yes, go to step 5; otherwise, go to step 6. Step 5: Determine whether the time the hand holds the small handle is less than a threshold. If so, determine whether the duration of the human-computer interaction force is greater than a threshold. If so, switch to the hold mode; if not, return to the previous state and re-determine. Step 6: Determine whether the time the hand holds the small handle exceeds the threshold. If so, determine whether the duration of the human-computer interaction force is greater than the threshold. If so, switch to follow-up mode; if not, return to re-determine, and continuously update the status and mode in real time to ensure timely response to doctor's needs.

2. The anti-shake control algorithm based on dual threshold judgment according to claim 1, characterized in that: In the initialization phase of step 1: (1) Setting the human-computer interaction force threshold Fth; (2) Setting the grip time threshold Tgrip; (3) Setting the interaction force duration threshold Tdur; (4) The system's initial mode is set to "follow-up mode".

3. The anti-shake control algorithm based on dual threshold judgment according to claim 2, characterized in that: The signal acquisition and preprocessing stage in step 2: (1) Attach surface electromyography electrodes to the operator's upper arm to collect electromyographic signals in real time; (2) The raw sEMG signal is transmitted to the processing unit through the acquisition module for preprocessing operations such as filtering, denoising, and normalization; (3) Synchronously collect human-computer interaction force signals and the state of holding the small handle.

4. The anti-shake control algorithm based on dual threshold judgment according to claim 3, characterized in that: In the step 2, the electromyographic fatigue identification stage: (1) The processed electromyographic signal is input into CNN to extract features and then sent to SVM classifier for judgment; (2) If the recognition result is "fatigue", the system enters fatigue buffer mode and executes the following sub-process: a) Increase the Fth threshold and extend the Tgrip and Tdur time determination range; b) Pause the movement of the robot arm and maintain the current position; c) Activate the sound and light prompt function to remind the doctor of fatigue status; (3) If the recognition result is "non-fatigue", the normal dual-threshold judgment control logic is entered.

5. The anti-shake control algorithm based on dual threshold judgment as claimed in claim 2, characterized in that: In step 2, a force sensor is installed between the robotic arm and the support structure to collect the interactive force applied by the doctor by holding the small handle in real time; the system continuously monitors the changes in the interactive force through high-frequency sampling and determines whether the current force value is higher or lower than the set threshold Fth.

6. The anti-shake control algorithm based on dual threshold judgment according to claim 2, characterized in that: The switching logic in the following mode: (1) If it is detected that the current human-computer interaction force is lower than Fth; (2) At the same time, the time of holding the small handle exceeds Tgrip; (3) The interaction force remains below the threshold for a period exceeding Tdur; (4) The system automatically switches to "hold mode" and the robotic arm maintains its current position to avoid disturbances caused by invalid contact.

7. The anti-shake control algorithm based on dual threshold judgment according to claim 2, characterized in that: The switching logic in the hold mode is: (1) If the human-computer interaction force is detected to be higher than Fth; (2) The time of holding the small handle is less than Tgrip; (3) At the same time, the interaction force remains above the threshold for a period exceeding Tdur; (4) The system switches to "follow-up mode" so that the robotic arm moves in real time following the movements of the human hand.

8. The anti-shake control algorithm based on dual threshold judgment according to claim 1, characterized in that: The structure of the CNN-SVM fusion electromyography fatigue recognition model is as follows: Step 1: CNN feature extraction network training phase (1) Sample construction: Collect EMG training samples and corresponding labels. The samples include: original sEMG data + fatigue / non-fatigue labels; (2) Network structure: (a) Construct a CNN-Softmax network for the training phase; (b) The CNN structure includes: convolutional layer, pooling layer, and fully connected layer; the output uses the Softmax activation function for preliminary classification; the Adam optimizer and the cross entropy loss function are used for model optimization training; the Softmax output is trained using the cross entropy loss function: Where y_i is the true label, is the Softmax output; (c) Model saving: After training is completed, remove the Softmax layer from the CNN structure and save it as a "feature extractor" model; Step 2: Statistical learning classifier fusion training phase (a) Feature extraction: Use the trained CNN model to extract features from the training and test samples to obtain high-dimensional deep semantic feature vectors. CNN performs feature extraction through convolution operations, and the calculation formula is as follows: Convolutional layer: Z^(l)=f(W^(l)*X^(l-1)+b^(l)), where Z^(l) is the output feature map of the lth layer, W^(l) is the convolution kernel weight, * represents the convolution operation, b^(l) is the bias term, and f(x) is the activation function; Pooling layer: P^(l) = Pooling(Z^(l)) The fully connected layer outputs a feature vector: F = φ(Z^(L)), where φ represents the feature vector formed by flattening and feeding it into the fully connected network. (b) Statistical learning classifier training: The extracted features and labels are used to train a traditional machine learning classifier. SVM is used as the main classifier to build a CNN-SVM fusion model. The feature vector F extracted by CNN is input into the SVM classifier, and its discriminant function is: y=sign(w^T*F+b) Where w is the weight vector, b is the bias, sign(·) is the sign function, If the kernel function form is used, take the radial basis kernel as an example: y=sign(Σα_iy_iK(F,F_i)+b) Where K(F,F_i)=exp(-γ||F-F_i||^2) Step 3, final model structure: (a) Input layer: original EMG signal; (b) Feature extraction layer: CNN network with the Softmax layer removed; (c) Classifier layer: SVM; (d) Output layer: fatigue / non-fatigue binary classification results.

Images