Method and system for intelligent control of force of flexible arm of physiotherapy robot

Abstract

This application discloses a method and system for intelligent force control of a flexible arm in a physiotherapy robot, comprising: collecting tactile and bioelectrical impedance data of the contact area through multimodal sensors to construct a digital anatomical feature atlas including muscle density and bone location; analyzing the atlas using a deep learning model to determine the sensitivity level of the contact area; combining this level with real-time physiological feedback from the patient to dynamically calculate and optimize a hierarchical progressive target force sequence; driving the flexible arm to apply force according to the sequence, and continuously updating the atlas to achieve closed-loop control. This application constructs an intelligent closed loop from perception to execution, and by integrating anatomical features and physiological feedback, achieves dual adaptive adjustment of force, significantly improving the adaptability, safety, and effectiveness of physiotherapy.

Description

Technical Field

[0001] This application relates to the fields of robotics and medical rehabilitation, specifically to a method and system for intelligent force control of a flexible arm used in a physiotherapy robot. Background Technology

[0002] Physiotherapy robots, as an important component of modern medical rehabilitation systems, play a role in assisting patient rehabilitation and reducing the workload of physical therapists by executing precise and repeatable physical therapy movements. In physiotherapy robots, the flexible arm, as the core actuator that directly contacts the patient, directly affects the treatment effect and patient safety through the precision and safety of its applied force. With the increasing demand for rehabilitation medicine and the deepening of personalized treatment concepts, higher requirements are being placed on the intelligence and adaptability of physiotherapy robot control systems.

[0003] Current physiotherapy robots largely rely on pre-set programs or fixed parameters for force control of their flexible arms. This approach exhibits significant limitations when dealing with the complex and diverse physiological structures of the human body and individual differences among patients. For example, different parts of the body, such as the muscle-dense area of ​​the shoulder and the vulnerable bone area near the clavicle, have vastly different tolerances and requirements for treatment force. However, existing systems typically lack the ability to sense these anatomical features in real time and dynamically, making it difficult to achieve targeted force adjustment. This "one-size-fits-all" control method may result in insufficient force at sites requiring strong pressure, thus affecting treatment effectiveness; while excessive force at sensitive or vulnerable sites may cause patient discomfort or even secondary injury.

[0004] Furthermore, existing force control systems generally lack a real-time linkage mechanism with the patient's physiological state. The treatment process is a dynamic and interactive process, and the patient's physiological responses, such as heart rate fluctuations and changes in muscle tension, are key indicators for measuring treatment comfort and safety. Existing technical solutions often ignore these real-time physiological feedback signals, failing to form an effective closed-loop regulation system. The technical challenge lies in how to accurately identify and quantify key anatomical features such as muscle density and bone distribution at the contact site using sensors at the end of the flexible arm, and based on these features, achieve safe, smooth, and effective dynamic force optimization. A single tactile sensor cannot provide sufficiently rich and in-depth information to accurately characterize subcutaneous tissue properties, resulting in the system being unable to provide sufficient basis for complex force decisions. Therefore, how to enable physiotherapy robots to perceive the anatomical characteristics of the treatment site in real time and combine this with the patient's physiological feedback to achieve a layered, progressive, and dynamically adaptive force control is a technical problem that urgently needs to be solved in the field of physiotherapy robot technology. Summary of the Invention

[0005] One aspect of this application is to provide a method for intelligent force control of a flexible arm for a physiotherapy robot, which solves the technical problem in the prior art where the force control of physiotherapy robots relies on fixed parameters and lacks the ability to adaptively adjust to human anatomical features and real-time physiological states, thus making it difficult to balance treatment safety and effectiveness.

[0006] To achieve the above objectives, this application provides a method for intelligent force control of a flexible arm in a physiotherapy robot, comprising the following steps:

[0007] Step one involves using a multimodal sensor configured on the end effector of the flexible arm of the physiotherapy robot to collect raw sensor data in real time at the contact points between the flexible arm and the patient's body. This raw sensor data includes at least tactile pressure distribution data and bioelectrical impedance signals. First, the raw sensor data is preprocessed, for example, by applying Gaussian filtering to the pressure data and performing baseline correction on the bioelectrical impedance signals, resulting in a standardized first data set. Then, multi-dimensional features are extracted from the first data set. For example, the pressure centroid and peak pressure of the contact area are calculated based on the pressure distribution data to form a first feature set. Simultaneously, time-frequency analysis is performed on the bioelectrical impedance signals to extract the impedance amplitude and phase information at different frequencies, forming a second feature set. In the feature fusion stage, dimensionality reduction fusion techniques such as principal component analysis (PCA) are used to merge the high-dimensional feature sets, generating a compact and information-rich contact feature vector.

[0008] Step two involves processing the contact feature vector. A pre-defined anatomical inverse analysis model is used to extract muscle density distribution and bone coordinates from the tissue beneath the contact area, thereby constructing or updating a digital anatomical feature map. This step utilizes an anatomical inverse analysis model based on estimation algorithms such as Extended Kalman Filter (EKF). The physical model underlying this algorithm is established based on statistical regression analysis of clinical trial data from a large population with diverse physical characteristics (age, sex, body mass index), combined with biophysical finite element simulation results of human tissues. To accommodate individual differences, the model can rapidly collect response data during the initial contact phase using a set of pre-defined probing actions (e.g., slight, multi-angle pressure loading). This data is then used to calibrate key parameters in the model online (such as equivalent fat layer thickness and baseline skin conductivity), providing a personalized foundation for subsequent accurate analysis.

[0009] In real-time operation, the EKF algorithm executes a prediction-update loop. In the prediction step, the algorithm uses the tissue state estimated at the previous time step and the state transition equation to predict the muscle density and bone location at the current time step. In the update step, the algorithm uses the latest contact feature vector generated in step S100 as the actual observation value at the current time step, combines it with the observation equation, calculates the residual between the predicted value and the observed value, and corrects the predicted state accordingly. In this way, the EKF algorithm can effectively and stably and accurately inversely solve for the physical properties of deep tissues from noisy sensor data. Finally, the estimated muscle density distribution information and the identified skeletal keypoint coordinate information are registered into a local two-dimensional coordinate system with the center of the contact area as the origin, forming a digitized, real-time updated anatomical feature map 304, such as... Figure 3 As shown.

[0010] Step three involves feeding the anatomical feature atlas as input to a pre-trained deep learning model for region attribute classification. This model (e.g., a convolutional neural network (CNN)) analyzes spatial features such as muscle density and bone distance in the atlas to determine the tissue characteristics of different regions and assigns a site sensitivity level to each region accordingly.

[0011] Step four involves acquiring the sensitivity level of the described body part and dynamically calculating a hierarchical progressive intensity sequence by combining this with real-time physiological feedback signals collected from external physiological monitoring devices. This step first selects an initial progressive intensity template from a pre-built intensity strategy library based on the body part sensitivity level. Then, the system collects physiological signals such as the patient's heart rate and skin conductance, and calculates a quantified physiological stress index. Finally, this physiological stress index is used as a dynamic adjustment coefficient to modulate the initial template in real time, generating a personalized progressive intensity sequence that responds to changes in the patient's physiological state in real time.

[0012] Furthermore, steps three and four—namely, determining the site sensitivity level and calculating the force progression sequence—can also be implemented using a reinforcement learning-based personalized treatment strategy self-optimization method. In this method, the entire force control process is modeled as a Markov decision process. The system's state is defined by a composite vector containing information such as the current anatomical feature map, real-time physiological feedback signals, and the force applied by the flexible arm. The system's action space is defined as the options for adjusting the force value at the next moment, such as a set of discrete or continuous actions like "increase by 0.5 Newtons," "maintain unchanged," or "decrease by 0.5 Newtons." The reward function is designed to guide the system in learning the optimal strategy; for example, a positive reward is given when a change in bioelectrical impedance representing muscle relaxation is detected, a negative reward (penalty) is given when the physiological stress index increases, and a larger negative reward is given when the force exceeds the safety boundary. By employing reinforcement learning algorithms such as Deep Q-Networks (DQN) or policy gradients, the system (acting as an agent) can autonomously learn an optimal control strategy through continuous interaction with the patient, by constantly trying (selecting actions) and observing the results (state transitions and rewards). This strategy can map any state to the optimal action, thereby replacing the calculation method based on preset rules. This enables a shift in force control from reactive adjustment to learning-based optimization, forming a personalized and continuously evolving treatment model for each patient.

[0013] Step five involves safety verification and iterative optimization of the force progression sequence calculated in step four. This step compares the force value and its rate of change at each level of the sequence with preset global safety constraints (such as maximum permissible contact pressure, maximum force increase rate, etc.). If any value in the sequence exceeds the safety constraints, the sequence is adjusted using an iterative optimization algorithm until it fully complies with the safety constraints, thus obtaining the optimized force control sequence.

[0014] Step six involves generating a low-level drive command sequence for the flexible arm based on the optimized force control sequence and pre-stored individual patient data. This step retrieves individual patient data such as age and medical history, and generates a set of personalized calibration coefficients. These calibration coefficients are then applied to the optimized force control sequence for fine-tuning, resulting in the final personalized treatment force plan. Subsequently, based on the kinematic and dynamic model of the flexible arm, inverse dynamics calculations are used to resolve this plan into low-level drive commands such as motor torque and target angle required by each joint of the robot, forming the final execution parameters.

[0015] Step seven involves the robot controller executing the underlying drive command sequence to drive the flexible arm to apply progressively layered forces to the patient's treatment area. Simultaneously, throughout the force application process, the multimodal sensor continuously monitors real-time data of the contact area and uses this data to cyclically update the anatomical feature atlas, thus forming a closed-loop feedback and control replanning mechanism. This mechanism, by continuously monitoring and cyclically updating the anatomical atlas during execution, constructs a complete closed-loop control system, achieving a highly adaptive response to changes in the treatment process. If abnormal fluctuations occur in the real-time updated atlas or physiological feedback signals, the system immediately uses this new information as input to dynamically replan the subsequent force control sequence, ensuring that the system can respond quickly and appropriately to any changes during treatment.

[0016] The technical solution provided in this application constructs a complete intelligent control closed loop from perception and decision-making to execution through the aforementioned steps. By employing multimodal sensor fusion technology, it achieves precise perception of the anatomical features of the treatment site, solving the problems of traditional methods that rely on limited information acquisition and cannot deeply understand tissue characteristics. Based on this, combined with real-time physiological feedback signals from the patient, it dynamically generates personalized force control strategies, effectively avoiding safety risks or poor therapeutic effects caused by improper force, thus improving the personalization level of treatment. Finally, through a closed-loop control mechanism of continuous monitoring and cyclical updates, the physiotherapy robot is endowed with the ability to respond to changes in the patient's state in real time and dynamically adjust the treatment plan, significantly enhancing the adaptability, safety, and intelligence of the entire treatment process. Furthermore, the introduction of a reinforcement learning-based self-optimization method enables the system to learn from experience and continuously optimize treatment strategies, providing a technical path for achieving personalized and intelligent rehabilitation treatment. Attached Figure Description

[0017] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating a method for intelligent force control of a flexible arm in a physiotherapy robot, as provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the structure of the end effector of the flexible arm of the physiotherapy robot in contact with the patient's body in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the data processing flow for generating anatomical feature maps from raw sensor data in an embodiment of this application.

[0021] Figure 4 This is a schematic diagram of the dynamic generation and optimization process of the intensity progression sequence in the embodiments of this application.

[0022] Figure 5 This is a schematic diagram of an anatomical feature atlas and site sensitivity level classification generated in an embodiment of this application.

[0023] Figure 6 This is a performance comparison chart between the policy optimization method based on reinforcement learning in this application embodiment and the traditional rule-based method.

[0024] Figure 7 This is a schematic diagram of the structure of the multimodal sensor module described in the embodiments of this application and the typical signals it collects, wherein... Figure 7 (a) is a pressure sensor array. Figure 7 (b) is a four-electrode bioelectrical impedance configuration. Figure 7 (c) shows the real-time pressure signal curve. Figure 7 (d) is the bioelectrical impedance spectrum.

[0025] Figure 8 This is a schematic diagram of the personalized force control strategy applied to patients of different age groups in the embodiments of this application. Detailed Implementation

[0026] The present application will now be described in further detail with reference to the accompanying drawings.

[0027] Example 1

[0028] Please see Figure 1 This application provides a method for intelligent control of the force of a flexible arm in a physiotherapy robot. This method achieves intelligent, adaptive closed-loop control of the treatment force through multimodal perception and real-time feedback.

[0029] Step S100: Using a multimodal sensor configured on the end effector of the flexible arm of the physiotherapy robot, raw sensing data of the contact area between the flexible arm and the patient's body is collected in real time. The raw sensing data includes at least tactile pressure distribution data and bioelectrical impedance signals.

[0030] Specifically, such as Figure 2 As shown, the flexible arm 102 of the physiotherapy robot 100 is equipped with a multimodal sensor module 101 at its end, which directly contacts the contact area 201 of the patient's body 200. In a preferred embodiment, the structure and signal acquisition schematic diagram of the multimodal sensor module 101 are shown below. Figure 7 As shown. Figure 7 (a) shows an 8×8 pressure sensor array of 40mm×40mm, where different gray levels represent different pressure values. Figure 7 (b) shows a four-electrode bioelectrical impedance configuration, labeled with four electrodes: E1, E2, E3, and E4. Figure 7(c) is a real-time pressure signal curve, showing the change of pressure value over time, and indicating the contact phase, stabilization phase and release phase. Figure 7 (d) shows the bioelectrical impedance spectrum, illustrating the curve of impedance amplitude as a function of frequency.

[0031] When fusing raw sensor data, data preprocessing is performed first. For example, a two-dimensional Gaussian filtering algorithm is applied to pressure distribution data to smooth the data and effectively remove glitches introduced by sensor noise or minor vibrations. For bioelectrical impedance signals, digital bandpass filtering is performed to filter out power frequency interference, and a baseline correction algorithm is used to eliminate signal drift caused by changes in electrode contact, thereby obtaining a standardized first set of data.

[0032] Subsequently, multi-dimensional features are extracted from the first data set. For example... Figure 3 As shown, multiple statistical and morphological features can be calculated from the pressure distribution data 301, such as the total pressure in the contact area, the location of the peak pressure, the coordinates of the pressure centroid, the average pressure value, and the spatial gradient of the pressure distribution. These features constitute the first feature set. Simultaneously, a Fast Fourier Transform (FFT) analysis can be performed on the bioelectrical impedance signal 302 to extract the impedance amplitude and phase angle information at different excitation frequencies. This information is highly sensitive to the tissue's hydration state and cell structure, thus constituting the second feature set.

[0033] During the feature fusion stage, since the dimensions of the first and second feature sets may be inconsistent, dimension alignment can be performed first, for example, by copying features from the lower-dimensional feature set or padding with zeros. Then, to reduce data redundancy and extract the most critical information, Principal Component Analysis (PCA) can be used to project the aligned high-dimensional feature set into a lower-dimensional feature space, generating a contact feature vector 303 with higher information density and a more compact structure. Furthermore, based on this contact feature vector, a pre-trained Support Vector Machine (SVM) classifier can be applied to identify the macroscopic state of the current contact (e.g., "initial light touch detection," "stable pressure therapy," etc.) in real time for status monitoring or log recording.

[0034] Step S200: Process the contact feature vector and extract the muscle density distribution information and bone position coordinate information of the tissue below the contact area through a preset anatomical reverse analysis model, thereby constructing or updating a digital anatomical feature map.

[0035] The key to this step lies in retrieving the internal structural information of subcutaneous tissue from external sensor signals. Specifically, this anatomical inverse analysis model can be an estimation algorithm based on the Extended Kalman Filter (EKF). The physical model upon which this model relies is established based on statistical regression analysis of clinical trial data from a large number of individuals with different physical characteristics (age, sex, body mass index), combined with the results of biophysical finite element simulations of human tissues. To accommodate individual differences, the model can rapidly collect response data during the initial contact phase through a set of preset probing actions (e.g., mild, multi-angle pressure loading), and use this data to calibrate key parameters in the model online (such as equivalent fat layer thickness and baseline skin conductivity), thereby providing a personalized basis for subsequent accurate analysis.

[0036] In real-time operation, the EKF algorithm executes a prediction-update loop. In the prediction step, the algorithm uses the tissue state estimated at the previous time step and the state transition equation to predict the muscle density and bone location at the current time step. In the update step, the algorithm uses the latest contact feature vector generated in step S100 as the actual observation value at the current time step, combines it with the observation equation, calculates the residual between the predicted value and the observed value, and corrects the predicted state accordingly. In this way, the EKF algorithm can effectively and stably and accurately inversely solve for the physical properties of deep tissues from noisy sensor data. Finally, the estimated muscle density distribution information and the identified skeletal keypoint coordinate information are registered into a local two-dimensional coordinate system with the center of the contact area as the origin, forming a digitized, real-time updated anatomical feature map 304, such as... Figure 3 As shown.

[0037] Step S300: The anatomical feature map is used as input to a pre-trained deep learning model for region attribute classification to determine the tissue characteristics of different regions in the map, and a site sensitivity level is assigned to each region accordingly.

[0038] Specifically, the deep learning model can be a lightweight convolutional neural network (CNN). The model is trained through supervised learning, and the training dataset contains a large number of anatomical feature maps collected from known body parts, each with manually labeled anatomical tags (such as "biceps brachii belly", "supraclavicular region", "scapular spine", etc.).

[0039] In real-time applications, the anatomical feature map generated in step S200 is formatted as a multi-channel image and input into the CNN model. For example, this image may contain three channels, representing muscle density values, estimated distance to bones, and the phase angle of bioelectrical impedance, respectively. The CNN model, through its cascaded convolutional and pooling layers, automatically learns and extracts spatial feature patterns from the map, such as high-density regions far from bones or low-density regions adjacent to bones. The model's fully connected layers then classify the attributes of each pixel or grid region in the map based on the extracted features.

[0040] The classification results might be labeled as "high-density muscle area," "low-density fat area," or "area adjacent to bone." The system then converts these physical attribute labels into a quantified site sensitivity level based on a pre-defined mapping rule table. For example, the rule could be defined as follows: the sensitivity level for areas adjacent to bone is the highest, level 5; areas with low muscle density are level 4; medium-density muscle areas are level 3; high-density muscle areas (usually with higher tolerance) are level 2; and pure fat areas are level 1. In this way, the entire contact area is assigned a refined sensitivity level distribution map, such as... Figure 5 As shown. Figure 5 It visually presents a feature map generated based on muscle density (represented by a heatmap background) and bone location (represented by X-shaped markers), and divides the map into regions with different sensitivity levels.

[0041] Step S400: Obtain the sensitivity level of the location and, in conjunction with the patient's physiological feedback signals collected in real time from external physiological monitoring equipment, dynamically calculate a layered progressive intensity sequence.

[0042] like Figure 4 As shown, the system first selects an initial intensity progression template 401 from a pre-built intensity strategy library based on the highest or average site sensitivity level for the current contact area output in step S300. This strategy library is built based on the experience of a large number of clinical physiotherapists and biomechanical safety data, and stores standardized intensity loading curves for sites with different sensitivity levels. For example, for a region with a sensitivity level of 5, the initial intensity of the template is very low, and the intensity escalation rate is extremely slow; while for a region with a sensitivity level of 2, the initial intensity of the template is higher, and the escalation rate is faster.

[0043] Simultaneously, the system connects to external physiological monitoring devices (such as heart rate bracelets and skin conductance sensors) worn by the patient via Bluetooth or other wireless communication methods, collecting real-time physiological feedback signals 402, including heart rate, heart rate variability (HRV), and skin conductance (GSR). These signals effectively reflect the patient's autonomic nervous system activity, thereby indirectly measuring their level of tension, pain, or discomfort. The system normalizes these multi-dimensional physiological signals and calculates a comprehensive, quantitative physiological stress index using a weighted fusion algorithm.

[0044] Finally, the system uses this physiological stress index as a dynamic adjustment coefficient to modulate the selected initial intensity progression template in real time. For example, if the physiological stress index is higher than the normal baseline, indicating that the patient may be feeling tense or uncomfortable, the system will proportionally reduce the target intensity values ​​of all subsequent levels in the template and slow down the rate of intensity change. Conversely, if the physiological stress index is stable or decreasing, the planned intensity progression can be maintained or moderately accelerated. Through this dual adaptive regulation, a personalized intensity progression sequence 403 that considers both static anatomical features and responds to dynamic physiological states is ultimately generated.

[0045] Furthermore, to achieve a higher level of personalization and self-optimization, steps S300 and S400 can also be implemented using a reinforcement learning-based self-optimization method for personalized treatment strategies. In this method, the entire force control process is modeled as a Markov decision process (MDP). The system's state is a composite vector containing the main features of the current anatomical atlas, real-time physiological feedback signals, and the force value applied by the flexible arm. The system's action space is defined as the options for force adjustment within the next control cycle, such as a set of discrete actions {increase by 0.5 Newtons, remain unchanged, decrease by 0.5 Newtons}. The reward function is designed to guide the system to learn the strategy most beneficial to the patient; for example, a positive reward is given when bioelectrical impedance signals indicate muscle relaxation, a negative reward is given when the physiological stress index significantly increases, and a larger penalty is given when the applied force touches a safety boundary. By employing algorithms such as Deep Q-Networks (DQN), the system (agent) learns an optimal control strategy π (State) -> Action through trial and error in continuous interaction with the patient. This strategy can directly output the optimal force adjustment action based on the currently perceived complete state, thus replacing the calculation method based on preset rules and templates. This allows force control to form a continuously evolving and improving personalized treatment model for each patient. Figure 6As shown, compared with traditional rule-based methods that have stable performance but no significant improvement, reinforcement learning-based methods can continuously learn and adapt in ongoing treatment, and their comprehensive treatment score can achieve significant and continuous growth, ultimately forming an optimized personalized strategy.

[0046] Step S500: Perform security verification and iterative optimization on the force progression sequence calculated in step S400.

[0047] This step is crucial to ensuring the treatment process meets safety requirements. The system has a built-in global set of safety constraints, set by clinical experts according to biomechanical safety standards. This set includes hard boundaries such as maximum permissible contact pressure (e.g., not exceeding 50 kPa), maximum force increase rate (e.g., not exceeding 10 N / s), and minimum effective treatment force (e.g., not less than 2 N). The verification process involves checking each target force value and its rate of change in the force progression sequence point by point.

[0048] If the value or rate of change at any point in the sequence exceeds the safety constraints, the system will initiate an iterative optimization procedure. This procedure prioritizes satisfying all safety constraints, with maintaining the smoothness and continuity of the intensity as a secondary optimization objective. For example, if the intensity increment at a certain step is too large, the optimization algorithm will reduce it to within the safety threshold and smoothly distribute the remaining increments across the preceding time steps, thus avoiding sudden jumps in intensity. After this step, an optimized intensity control sequence 404 is finally obtained that meets both personalized treatment needs and strictly adheres to safety regulations.

[0049] Step S600: Based on the optimized force control sequence and by integrating pre-stored individual patient difference data, generate the underlying drive command sequence for the flexible arm.

[0050] Before execution, the system retrieves pre-stored individualized data from the patient's file. This data includes the patient's age, gender, body mass index (BMI), medical history (such as osteoporosis), history of skin allergies, and specific treatment contraindications set by the therapist. The system then transforms this static data into a set of personalized calibration coefficients using a personalized adjustment rule base. For example, for an older patient diagnosed with osteoporosis, the system automatically applies a calibration coefficient that reduces the overall strength level by 20%. Figure 8 This is a schematic diagram of a specific application example of this application, which shows how the system applies different basic treatment strategies according to the patient's age group (youth, middle-aged, elderly), such as adjusting the target intensity, climbing rate and dynamic safety boundary, so as to achieve age-adaptive control for different physiological conditions.

[0051] This set of calibration coefficients is applied to the optimized force control sequence, and each target force value in the sequence is fine-tuned to form the final personalized treatment force plan. Subsequently, based on the precise kinematic and dynamic model of the physiotherapy robot's flexible arm, the system performs inverse dynamics calculations. This calculation process accurately decomposes the target contact force required to be achieved on the end effector at each moment in the treatment plan into the underlying drive commands such as motor torque, target rotation angle, or angular velocity required to be output by each joint of the robot's flexible arm. These commands constitute the execution parameters that are finally sent to the robot's underlying controller.

[0052] In step S700, the robot controller executes the underlying drive command sequence to drive the flexible arm to apply layered and progressive force to the patient's treatment site. At the same time, during the entire process of force application, the multimodal sensor continuously monitors the real-time data of the contact area and uses the real-time data to cyclically update the anatomical feature atlas, thereby forming a closed-loop feedback and control replanning.

[0053] After receiving the final execution parameters, the robot controller precisely drives the motors of each joint of the flexible arm, causing its end effector to press or massage the patient's treatment area according to the planned force sequence. Crucially, during execution, the data acquisition and decision-making process of steps S100 to S400 is not interrupted, but rather runs continuously in the background at a high frequency (e.g., 10 times per second).

[0054] This means that the anatomical feature map and physiological stress index are updated dynamically in real time. This constitutes a tight closed-loop feedback control circuit. For example, if the patient's muscles tense momentarily due to discomfort during compression, the multimodal sensors will immediately capture the changes in pressure distribution and bioelectrical impedance, resulting in an updated anatomical feature map showing increased muscle density. Simultaneously, physiological monitoring equipment may also detect an increase in heart rate. This new information is immediately input into the decision-making process, and the system will instantly replan the subsequent force control sequence, potentially reducing the force or pausing the increase in force to allow the patient to relax. This real-time feedback and replanning mechanism enables the physical therapy robot to respond quickly, intelligently, and appropriately to subtle changes during treatment, thus achieving a high degree of adaptive control.

[0055] Example 2

[0056] This application also provides a system for intelligent force control of a flexible arm in a physiotherapy robot. The internal structure and working principle of this system correspond to the method described in Embodiment 1. This system can be integrated into the physiotherapy robot controller or used as a standalone control module. The system includes:

[0057] A data acquisition and fusion module is configured to perform the function described in step S100 of embodiment 1, namely, to acquire raw sensing data through a multimodal sensor and perform preprocessing, feature extraction and fusion to generate a contact feature vector.

[0058] An anatomical atlas construction module is configured to perform the function described in step S200 of Embodiment 1, namely, to receive contact feature vectors and extract tissue information through an anatomical inverse analysis model to construct or update a digitized anatomical feature atlas.

[0059] A region attribute classification module is configured to perform the function described in step S300 of embodiment 1, that is, to input anatomical feature maps into a deep learning model to classify attributes of different regions and assign site sensitivity levels.

[0060] A force sequence generation module is configured to perform the functions described in steps S400 and S500 of Embodiment 1, namely, to dynamically calculate, verify and optimize a hierarchical force control sequence by combining the site sensitivity level and the patient's physiological feedback signal.

[0061] A drive instruction generation module is configured to perform the function described in step S600 of embodiment 1, namely, to generate the underlying drive instruction sequence of the flexible arm based on the optimized force control sequence and by integrating individual patient differential data.

[0062] An execution and feedback control module is configured to perform the function described in step S700 of embodiment 1, namely, to control the robot to execute drive commands and, during execution, continuously call other modules to cyclically update the anatomical feature atlas and reprogram the control sequence, thereby achieving closed-loop feedback control.

[0063] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A system for intelligent force control of a flexible arm in a physiotherapy robot, characterized in that, include: The data acquisition and fusion module is used to acquire and process tactile pressure distribution data and bioelectrical impedance signals from multimodal sensors to generate a contact feature vector. The anatomical atlas construction module is used to construct a digital anatomical feature atlas representing muscle density distribution and bone location based on the contact feature vector using an anatomical inverse analysis model. The region attribute classification module is used to classify the digitized anatomical feature maps to determine the site sensitivity level; The force sequence generation module is used to calculate and optimize the force control sequence by combining the sensitivity level of the site and the real-time collected physiological feedback signals of the patient. The drive command generation module is used to generate the underlying drive command sequence of the flexible arm based on the force control sequence and by integrating individual patient difference data. The execution and feedback control module is used to control the flexible arm to execute the underlying drive instruction sequence, and during the execution process, continuously triggers the data acquisition and fusion module, the anatomical map construction module, the regional attribute classification module, and the force sequence generation module to cyclically update the digitized anatomical feature map and dynamically reprogram the force control sequence.

Images