Upper limb rehabilitation dynamic control method based on three-dimensional visual intention recognition

Abstract

The invention provides an upper limb rehabilitation dynamic control method based on three-dimensional visual intention recognition, which comprises the following steps: synchronously acquiring three-dimensional visual observation data of upper limbs of a patient and pose data of an execution mechanism of a rehabilitation mechanical arm, constructing an upper limb pose vector containing an upper limb kinematics state, and associating the upper limb pose vector with an external traction displacement increment; based on the upper limb attitude vector and the external drafting displacement increment, generating a passive attitude reference vector representing the theoretical upper limb attitude under the pure passive drafting condition; calculating a residual error between the actually observed upper limb attitude vector and the passive attitude reference vector, and separating an active movement contribution amount of the patient from the residual error; generating a patient motion intention vector containing direction and intensity information based on the active motion contribution; according to the motion intention vector of the patient, an instruction used for controlling the auxiliary torque of the rehabilitation mechanical arm is generated, the real active intention of the patient can be accurately separated and extracted, and stable and accurate dynamic auxiliary control is achieved accordingly.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of rehabilitation training and intelligent auxiliary control, and particularly to a dynamic control method for upper limb rehabilitation based on three-dimensional visual intention recognition. BACKGROUND

[0002] An upper limb rehabilitation robot system drives the patient's limbs to move regularly through a mechanical arm, which helps to restore joint mobility and muscle function. Among them, the semi-active training mode encourages patients to try to actively force to participate while the mechanical arm applies external stretching, which can better promote neuromuscular function reconstruction. In order to make the training process more adaptive, the existing technology usually uses a three-dimensional vision system to capture the patient's upper limb movement in real time, and infers the patient's intention based on the movement trajectory, and then dynamically adjusts the size of the auxiliary force to form a closed-loop interactive control.

[0003] However, in semi-active training, the limb movement trajectory observed by the three-dimensional vision system is essentially a superimposed result of the passive stretching displacement of the mechanical arm and the active displacement generated by the patient's autonomous muscle force. Since the vision system can only obtain the spatial position change, it cannot separate the two different movement components from the physical level, resulting in a serious coupling of passive and active components in the observed signal. This coupling makes the intention direction and intensity identified directly from the trajectory change lose its true physiological significance and cannot accurately represent the patient's own force intention. In actual application, the existing method is easy to misjudge the displacement caused by passive traction as the patient's active movement, which leads to the mechanical arm withdrawing force too early, so that the patient cannot maintain the movement due to lack of force and the training is interrupted; or the patient's weak active force is submerged in the strong passive signal and cannot be detected, which leads to insufficient assistance and affects the training effect. Such identification errors will cause the mismatch between man and machine, not only reducing the training efficiency and safety, but also dampening the enthusiasm of the patient for long-term participation, hindering the rehabilitation process.

[0004] Therefore, there is an urgent need for an upper limb rehabilitation training method that can accurately separate and extract the patient's real active intention from the coupled passive traction and active force observation, and realize stable and precise dynamic auxiliary control accordingly. SUMMARY

[0005] Therefore, in order to solve the problems brought by the prior art, the present application provides a dynamic control method for upper limb rehabilitation based on three-dimensional visual intention recognition.

[0006] In a first aspect, the present disclosure provides a dynamic control method for upper limb rehabilitation based on three-dimensional visual intention recognition, which comprises: S1, synchronously collecting three-dimensional visual observation data of the patient's upper limb and actuator pose data of the rehabilitation mechanical arm, constructing an upper limb posture vector containing the kinematic state of the upper limb, and associating it with the external stretching displacement increment; S2, generating a passive posture reference vector representing a theoretical posture of the upper limb under pure passive stretching based on the upper limb posture vector and an external stretching displacement increment; S3, calculating a residual error between an actually observed upper limb posture vector and the passive posture reference vector, and separating a voluntary motion contribution of the patient from the residual error; S4, generating a patient motion intention vector containing direction and intensity information based on the voluntary motion contribution; S5, generating an instruction for controlling an auxiliary torque of a rehabilitation robot arm according to the patient motion intention vector.

[0007] Optionally, the S1 comprises: converting the three-dimensional visual observation data and the actuator pose data to the same coordinate system; extracting key points representing positions of joints of the upper limb from the three-dimensional visual observation data; calculating joint angles and end displacement according to the key points to form the upper limb posture vector; calculating a displacement increment along a preset stretching direction according to the actuator pose data as the external stretching displacement increment; correlating the upper limb posture vector and the external stretching displacement increment at the same time to form a synchronous data pair.

[0008] Optionally, the S2 comprises: calculating a length change proportion of the upper limb caused by unit stretching according to the upper limb posture vector and the external stretching displacement increment; estimating passive angle changes of each joint based on the proportion and the external stretching displacement increment; generating the passive posture reference vector according to the passive angle changes.

[0009] Optionally, the estimating passive angle changes of each joint comprises: based on the proportion and the external stretching displacement increment, according to calculating passive angle changes of the joint wherein, is the length change proportion of the upper limb caused by unit stretching, C(t) is the external stretching displacement increment, is an equivalent arm length of the joint pre-calibrated.

[0010] Optionally, the S3 comprises: calculating a residual error between an actually observed upper limb posture vector and the passive posture reference vector in each degree of freedom; setting a decision threshold for each component of the residual, and regarding the residual component with an amplitude exceeding the decision threshold as an effective active contribution; performing temporal smoothing processing on the residual component regarded as an effective active contribution, to form a stable active motion contribution amount.

[0011] Optionally, the setting of the decision threshold for each component of the residual includes: setting different decision thresholds for the joint angle residual and the end displacement residual respectively; wherein, for the end displacement residual, the presence of an effective active contribution is determined by calculating the average amplitude of the residual displacement in each direction in three-dimensional space, and comparing the average amplitude with a preset end amplitude threshold.

[0012] Optionally, the S4 includes: fusing the contributions of multiple joints in the active motion contribution amount into a scalar displacement along the stretching direction; determining a motion intention direction according to the sign and size of the scalar displacement; calculating a normalized intention strength according to the amplitude of the scalar displacement and a preset reference value; performing anti-jitter processing on the intention direction and intention strength to form the patient motion intention vector.

[0013] Optionally, the S5 includes: mapping the intensity information in the patient motion intention vector into an auxiliary gain coefficient; combining the auxiliary gain coefficient, intention direction information and a reference maximum auxiliary torque to calculate a target auxiliary torque; performing smoothing filtering and safety limiting on the target auxiliary torque to generate the final control instruction.

[0014] In a second aspect, the present disclosure provides an electronic device, which includes a memory and at least one processor, the memory stores a computer program, and the processor is configured to execute the computer program to implement the method of the first aspect.

[0015] In a third aspect, the present disclosure provides a computer storage medium, which stores a computer program, and the computer program is executed to implement the method of the first aspect.

[0016] The present disclosure has the following advantages compared with the prior art: 1) By constructing a unified spatiotemporal data representation and passive response model, the motion component generated by the passive traction of the robot arm can be quantitatively separated from the composite motion trajectory observed by three-dimensional vision, thereby filtering out the interference of external driving on the intention recognition signal. This makes the active contribution extracted subsequently purely from the patient's own muscle contraction, significantly improving the physiological authenticity and accuracy of motion intention recognition, and fundamentally solving the problem of intention misjudgment caused by the coupling of active and passive motion.

[0017] 2) By designing a threshold-based residual screening mechanism and temporal smoothing process, the active motion segments with physiological significance can be robustly identified from the separated active signal, and fused and quantified into intention vectors containing explicit direction and continuous intensity. This process effectively suppresses the interference of measurement noise and small physiological tremors, avoids frequent jumps and jitter of the intention output, and makes the recognition result not only sensitive to changes in patient's wishes, but also has good stability, providing a reliable basis for subsequent control.

[0018] 3) Based on the accurately identified intention vector, the assistive torque of the rehabilitation robot arm is generated and dynamically adjusted in real time through a configurable gain mapping and smoothing output mechanism. This control strategy enables the assistive force to form a continuous and smooth adaptive relationship with the patient's real-time intention direction and effort level, achieving on-demand assistance. This not only avoids interruptions in training due to premature force removal, but also prevents the impact on training effectiveness and safety due to insufficient or excessive assistance, forming a stable, coordinated, and personalized human-robot interaction closed loop.

[0019] 4) Anti-interference and safety protection mechanisms such as smoothing and amplitude limiting are designed throughout the entire process from data alignment, model construction, component separation to intention generation and control. This not only improves the robustness of the algorithm in dealing with real scene noise and individual differences, but also ensures that the entire system output is always within the pre-set safe and reasonable range. This improves the training effectiveness and accuracy while ensuring the safety of the rehabilitation process and the comfort of the patient, enhancing the practicality and reliability of the technology in clinical complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure, together with the description.

[0021] Figure 1 A flowchart of the upper limb rehabilitation dynamic control method based on three-dimensional vision intention recognition provided by the embodiments of the present disclosure is shown. Figure 2 A patient motion intention generation and assistive control instruction mapping schematic diagram provided by the embodiments of the present disclosure is shown.

[0022] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present disclosure more clearly, and should not be used to limit the scope of protection of the present disclosure.

[0024] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0026] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0027] Figure 1 A flowchart of the dynamic control method for upper limb rehabilitation based on three-dimensional visual intent recognition provided in this disclosure embodiment is shown below. Figure 1 As shown, it may include the following steps: S1: Simultaneously collect three-dimensional visual observation data of the patient's upper limb and the positional data of the actuator of the rehabilitation robotic arm, construct an upper limb posture vector containing the kinematic state of the upper limb, and associate it with the external stretching displacement increment.

[0028] The three-dimensional visual observation data of the patient's upper limbs and the pose data of the actuator of the rehabilitation robotic arm are spatiotemporally aligned and standardized to provide a unified and comparable basic data representation for subsequent separation of passive and active motion components. This is achieved through the following sub-steps.

[0029] S1.1: Acquire and calibrate the 3D point cloud and the original pose data of the actuator, and complete the coordinate system alignment and data standardization.

[0030] First, a fixed 3D vision acquisition device, such as a point cloud acquisition system based on a depth camera, is deployed in the rehabilitation training area. Simultaneously, it is ensured that the end effector of the rehabilitation robotic arm has trackable markers or that its control system can directly provide the end effector pose. Before training begins, spatial calibration between the 3D vision coordinate system and the actuator's base coordinate system must be completed. Specifically, a calibration board with known geometric dimensions, such as a planar calibration board with a specific black and white checkerboard pattern, is used. Multiple sets of images with different poses are acquired and solved, and a mature camera calibration algorithm (such as Zhang Zhengyou's calibration method) is employed to obtain the calibration matrix that transforms the camera coordinate system to the unified coordinate system for rehabilitation training. After calibration, the calibration accuracy can be verified by calculating the reprojection error of feature points, ensuring that it meets the millimeter-level spatial positioning requirements of rehabilitation training.

[0031] A unified coordinate system for rehabilitation training is typically defined based on a fixed reference point on the base of the rehabilitation robotic arm or the training platform. For example, with the center of the robotic arm base as the origin, two mutually perpendicular directions in the horizontal plane are designated as the X and Y axes, and the vertical direction as the Z axis, a right-handed coordinate system fixed in the training environment is established. This coordinate system facilitates a unified description of the spatial pose relationships between the patient's upper limb, the rehabilitation robotic arm, and the training target.

[0032] For each sampling time t, the raw point cloud data of the patient's upper limb is obtained from the 3D vision system, denoted as Simultaneously, the raw pose data of the actuator's end effector is read from the actuator control system and recorded as follows: This data includes end-effector position and attitude information.

[0033] Subsequently, the aforementioned calibration matrix will be used to... and The data is uniformly transformed to the same rehabilitation training coordinate system. Specifically, the transformation relationships obtained through calibration, which typically include rotation, translation, and possible scale factors, are used to transform the point cloud data and actuator end-effector pose data from the original camera coordinate system to the rehabilitation training coordinate system.

[0034] At the end position of the actuator Using the same coordinate system for unified processing, we finally obtained the calibration point cloud data D(t) and the end effector pose data Q(t) in the unified coordinate system.

[0035] This step achieves precise geometric alignment between the patient point cloud and the robotic arm end effector trajectory in the same physical reference system, laying a foundation for comparability and consistency for subsequent description of upper limb posture changes and external stretching displacements from the same spatial reference.

[0036] S1.2: Extract key anatomical points of the upper limb from the calibration point cloud and construct a set of key points.

[0037] Based on the obtained calibrated point cloud data D(t), key points representing the main anatomical structures of the upper limb, such as the shoulder joint, elbow joint, wrist joint, and hand extremities, need to be extracted. In practice, the calibrated point cloud D(t) is first initially screened according to a pre-defined upper limb activity space range. This space range can be determined based on the patient's sitting posture and the position of the training table, for example, within a volume area 0.5 to 1.5 meters from the camera and 0.6 to 1.6 meters high. Subsequently, based on a pre-loaded human upper limb skeletal structure model, the 3D positions of key anatomical points such as the shoulder, elbow, and wrist are identified and located from the screened point cloud through point cloud segmentation, cluster analysis, or a trained deep learning model. The deep learning model that can be used is a human pose estimation model based on 3D point clouds or 2D projected images, such as those built on architectures like PointNet++, Voxel-Pose, or HRNet and their variants. The model is trained using point cloud data containing a large amount of labeled 3D coordinates of upper limb key points or synchronous multi-view 2D image data to learn the mapping relationship from input data to the spatial location of key points. Specific algorithms for key point detection include, but are not limited to: direct regression of key point coordinates, prediction of key point heatmaps followed by finding extreme values, or instance segmentation followed by calculation of joints.

[0038] With elbow joint key points For example, its spatial location can be approximated by the geometric center of its nearest set of points. Using a similar principle and method, key points of the shoulder joint can be extracted sequentially. Key points of the wrist joint and the extremities of the hand This process ultimately forms the set of key points for the upper limb at the current moment: .

[0039] This step transforms the dense and redundant original point cloud D(t) into a set K(t) consisting of a small number of key points with clear anatomical significance, thus providing a highly simplified and stable geometric skeleton representation for subsequent calculations of upper limb joint angles, limb segment spatial vectors, and end-effector trajectories, effectively reducing the complexity of subsequent calculations.

[0040] S1.3: Calculate the joint angle and end displacement increment based on the key point set to construct the three-dimensional posture vector of the upper limb.

[0041] Based on the obtained key point set K(t), by calculating the spatial geometric relationship between the key points, the real-time angles of each major joint and the spatial displacement changes of the end-effector hand can be derived, thereby constructing a digital posture vector that can comprehensively describe the instantaneous spatial state and kinematic changes of the upper limb.

[0042] To calculate the shoulder joint flexion and extension angle For example, the line connecting the shoulder joint to the elbow joint can be defined as the upper arm vector, and the line connecting the elbow joint to the wrist joint can be defined as the forearm vector. The angle can be obtained by calculating the angle between the projections of these two vectors onto a specific reference plane (such as the sagittal plane). For changes in the position of the distal end of the hand, it is necessary to calculate its displacement increment over consecutive time steps. The displacement increment of the distal end of the hand along each coordinate axis between adjacent sampling time steps is calculated to obtain... , and In practical applications, the direction of displacement increment calculation should be consistent with the preset stretching direction. The calculated angles of major joints, such as the shoulder, elbow, and wrist, should be used as the basis for this calculation. , , Combined with the displacement increments of the end-effector hand in the three coordinate axes, this forms the current three-dimensional posture vector of the upper limb: .

[0043] This step transforms the discrete set of keypoint spatial coordinates K(t) into a posture vector S(t) with clear kinematic and physical meaning. This vector fully characterizes the configuration of each joint of the upper limb at the current moment and the overall motion trend of the end effector, providing a precise and quantitative geometric and kinematic description basis for distinguishing between passive motion components generated by external mechanical traction and active motion components generated by the patient's own muscle contraction in subsequent steps.

[0044] S1.4: Calculate the external stretching displacement increment of the robotic arm in a unified coordinate system and form a synchronized data pair with the attitude vector.

[0045] Based on the obtained actuator end-effector pose data Q(t) transformed to a unified coordinate system, the spatial position information of the actuator end-effector in the rehabilitation training coordinate system is extracted, and its scalar position along the preset stretching direction (usually consistent with the longitudinal axis of the upper limb) is denoted as... Between two adjacent sampling times, the increment of external traction displacement applied to the patient's limb by the actuator end effector can be calculated as follows: , Where C(t) is the external stretching displacement increment corresponding to time t, in millimeters; This is the position scalar of the end point in the stretching direction at that moment; This corresponds to the position at the previous sampling moment. The system's sampling period needs to be set based on the control cycle of the actuator and the frame rate of the vision system, typically ranging from 10 to 50 milliseconds, to ensure the timeliness of the data.

[0046] Subsequently, the calculated posture vector S(t) at the same time t is associated and bound with the stretch C(t), thus forming a data sequence pair {S(t), C(t)} that is strictly synchronized in time and has a unified spatial reference. This operation ensures that in each frame of subsequent processing, the posture information of the patient's upper limb can accurately correspond to the stretch output by the actuator at that moment.

[0047] This step precisely matches and correlates the upper limb posture vector S(t), which describes the patient's own movement state, with the actuator stretch displacement C(t), which represents the external driving force, in both time and space dimensions. The resulting synchronized data pair provides indispensable and structurally complete underlying input data for subsequent steps to build a passive response model and separate the patient's active contribution from the observed synthetic motion.

[0048] In the technical solution of this disclosure, by accurately aligning and standardizing the three-dimensional visual observation data and the positional data of the actuator in time and space, and extracting key points of the upper limb with clear anatomical significance to construct a three-dimensional posture sequence and synchronous external stretching representation under a unified coordinate system, a geometrically consistent, time-synchronized and physically clear underlying data foundation is provided for the entire intention recognition and control process. This enables reliable comparison and separation of motion components originating from the patient and external mechanical drives within a unified reference framework.

[0049] S2: Based on the upper limb posture vector and the external stretching displacement increment, generate a passive posture reference vector that represents the theoretical posture of the upper limb under pure passive stretching conditions.

[0050] Based on the synchronous data obtained in step S1, the equivalent length change in the stretching direction is calculated, the passive transfer ratio is derived, and a smooth vector representing the theoretical upper limb posture under pure passive stretching conditions is finally generated. This calculation of the equivalent length change in the stretching direction, the derivation of the passive transfer ratio, and the generation of a smooth vector representing the theoretical upper limb posture under pure passive stretching conditions provide a reliable reference for accurately separating the patient's active movement contribution in subsequent steps. This is specifically achieved through the following sub-steps.

[0051] S2.1: Calculate the equivalent length increment of the upper limb in the stretching direction based on the synchronous data.

[0052] Based on the synchronized data pair output in step S1, the physical quantities reflecting the overall extension and contraction of the upper limb along the stretching direction need to be extracted from the geometric information implicit in the posture vector S(t). Specifically, using the coordinates of the shoulder joint and the end point of the hand in the key point set, the projected distance of the upper limb in the preset stretching direction at the current moment is calculated, and this distance is defined as the equivalent length of the upper limb. The stretching direction is usually set according to the training mode, such as along the longitudinal axis of the upper limb from the shoulder to the hand. The equivalent length change of the upper limb due to stretching between adjacent sampling times, i.e., the length increment, is denoted as... When calculating the equivalent length, the coordinate vectors of the shoulder joint and the end of the hand need to be projected onto the stretching direction vector and their difference in magnitude needs to be calculated. The sampling period of the system remains consistent with step S1, typically between 10 and 50 milliseconds.

[0053] This step condenses the complex attitude vector S(t), which contains multiple joint angles and end-effector displacements, into a one-dimensional scalar variation sequence along the main stretching direction. This sequence visually characterizes the observed overall extension and contraction of the upper limb and is strictly synchronized with the external stretching displacement C(t) at the same moment. This transformation simplifies the dimensions of subsequent modeling, enabling the analysis of the relationship between external driving force and passive deformation based on scalar relationships, and laying a clear and easily processed data foundation for constructing a passive transfer model.

[0054] S2.2: Establish the passive transfer ratio coefficient between external stretching displacement and upper limb length increment.

[0055] In obtaining the equivalent length increment After determining the corresponding external stretching displacement C(t), the proportional relationship between the two needs to be quantified to describe the degree of upper limb length change caused by a unit external stretching displacement under ideal passive conditions. Define the passive transfer proportionality coefficient. , for The ratio to C(t). Considering the potential for minor physiological resistance, soft tissue viscoelasticity, or sensor noise during actual rehabilitation, the original calculated... Outliers may occur that exceed reasonable limits. Therefore, it is necessary to limit the value to a reasonable range determined in advance based on biomechanical characteristics and equipment parameters. Subsequent processes will use the limited value as the effective passive transmission ratio; for simplicity, it will still be denoted as [value missing]. However, it should be understood that this is the value after the limit is applied.

[0056] This step establishes the relationship between the external mechanical displacement C(t) and the observed change in upper limb length. Dynamic correlation coefficient between This coefficient quantitatively characterizes the efficiency of converting exogenous stretching drive into a passive geometric response of the limb, and is a key link connecting known external inputs with the inference of internal passive joint changes.

[0057] S2.3: Calculate the passive response increment of the joint based on the passive transmission ratio, and construct the original passive attitude reference vector by combining it with the reference attitude.

[0058] Based on the obtained passive transfer ratio The external stretching amount C(t) can be further extrapolated to the level of joint and distal displacement to estimate the angular changes of the major joints of the upper limb and the displacement changes of the distal end under pure passive stretching. Specifically, the passive displacement increment of the distal end along the preset stretching direction... Can be directly from The calculations are obtained and can be decomposed into each coordinate axis. To balance model complexity and practicality, a joint that dominates the change in length in the stretching direction is usually chosen as a representative, such as the elbow joint. It is assumed that the contribution of this joint's motion to the equivalent length change can be represented by a fixed reference lever arm. To represent this equivalently, the passive angle increment of the joint... The following formula can be used for estimation: , in, The passive angle increment of the dominant joint at time t can be expressed in degrees or radians. C(t) represents the passive transmission ratio; C(t) represents the extension displacement increment of the actuator, in millimeters. The equivalent lever arm length of the joint, expressed in millimeters, can be obtained through individualized anthropometric measurements or calibration based on a standard model, typically ranging from 200 to 350 millimeters. The calibration based on a standard model can be implemented through the following exemplary process: After the patient wears or comes into contact with the rehabilitation robotic arm, the system drives it to execute a preset, slow passive motion trajectory, while simultaneously recording the end-effector displacement and joint angle changes observed in three-dimensional vision. Based on this data, the average lever arm length of the equivalent joint (such as the elbow joint) that plays a dominant role in the current movement mode is deduced through geometric relationship fitting or parameter identification algorithms. This process can be assisted by optical motion capture systems or robot-embedded encoders as measurement devices. For dynamic changes in the lever arm that may be caused by changes in joint angles, this average can be used in the simplified model. The value is estimated.

[0059] In rehabilitation training, this model quantitatively describes the theoretical joint angle changes that a patient's upper limb should produce due to passive properties such as tissue viscoelasticity under a given external stretch C(t). Its core significance lies in providing a purely passive response benchmark for subsequent steps. This allows for the precise isolation of the contribution of robotic arm traction from observed compound movements, ensuring that the subsequently extracted active contribution truly reflects the patient's neuromuscular control ability and providing an objective physiological basis for assessing the patient's active participation.

[0060] After obtaining the passive angle increments of the dominant joints, they need to be integrated into a complete passive attitude description. To this end, a predefined passive reference attitude is introduced. This posture characterizes the initial joint angles and distal positions of the patient's upper limb when it is completely relaxed and unaffected by any active force exertion. It is typically measured and recorded before training through a slow, smooth, purely passive stretching process. Therefore, the original passive posture vector at the current moment... This can be constructed by superimposing passive angle increments onto the baseline attitude: , in, Let be the original passive attitude vector at time t. The passive reference attitude vector is of constant form. It is represented in vector form as the passive adjustment of the degrees of freedom of each joint caused by the current stretch.

[0061] This step achieves the mapping from one-dimensional stretching displacement and proportionality coefficient to the theoretical passive posture of a multi-degree-of-freedom upper limb. The resulting... A reference trajectory that is directly related to the external drive and reflects the idealized passive response was constructed. This provides a crucial calculation template and comparison benchmark for comparing the actual observed attitude with this theoretical reference in subsequent steps, thereby extracting the part that exceeds the passive response, i.e., the active motion contribution.

[0062] S2.4: Perform temporal smoothing on the original passive attitude reference vector and output the final passive attitude reference response.

[0063] Due to potential minor backlashes in mechanical transmissions, inherent noise in measurement systems, or the nonlinearity exhibited by human soft tissue under transient stretching, the directly calculated original passive attitude vector... It may contain unwanted high-frequency fluctuations. To avoid these transient disturbances being misidentified as active signals in subsequent active component separation steps, it is necessary to... Time-domain smoothing is performed to extract a more stable and smoother passive response trend. This is achieved through... After smoothing, the final passive attitude reference vector is obtained. The smoothing coefficient can be adjusted according to the smoothness of the patient's movement and the system response requirements. This smoothing process effectively suppresses high-frequency noise and transient impact components in the original passive posture reference vector, resulting in a more accurate final passive response output. It presents a smooth, continuous curve, which is more in line with the passive limb movement pattern caused by smooth stretching in actual situations.

[0064] This step produces a smoothed passive attitude vector. This will be used as the final output of step S2, namely the passive attitude reference vector, and passed to subsequent processing steps. Ensure... The smoothness and stability of the equation are crucial because subsequent steps will calculate the actual observed attitude S(t) and... The difference, a robust and stable theoretical reference, can ensure that the difference mainly captures the patient's real active force information, rather than the fluctuations or noise from the model itself, thereby improving the accuracy and reliability of active intent recognition.

[0065] In the technical solution of this disclosure, a quantitative response model for passive deformation of multiple joints of the upper limb, based on synchronized posture and stretching data, is constructed. By calculating the equivalent length increment in the stretching direction and its dynamic proportional relationship with the external stretching, the theoretical posture trajectory that the upper limb should present under pure passive traction conditions is deduced, and a stable passive response reference quantity is generated through smoothing processing. This provides a crucial calculation benchmark and reference template for accurately separating the passive component brought about by mechanical traction from the actual observed composite motion.

[0066] S3: Calculate the residual between the actual observed upper limb posture vector and the passive posture reference vector, and separate the patient's active movement contribution from the residual.

[0067] By comparing the observed upper limb posture with the theoretical passive posture reference constructed in step S2, the difference residuals between the two are calculated. A reasonable screening and smoothing mechanism is designed to separate and extract the active motor contribution from the residuals, thus achieving the separation of active and passive components at the kinematic level. This is specifically achieved through the following sub-steps.

[0068] S3.1: Calculate the joint angle residual and end displacement residual between the actual observed attitude vector and the passive attitude vector.

[0069] The smoothed passive attitude reference vector, based on the output of step S2, serves as the passive attitude reference vector. The attitude vector S(t), which is synchronized with the actual observed attitude vector, first requires calculating the difference between the two at each degree of freedom of motion. These differences constitute the original residual basis for separating the active contributions. The attitude vector S(t) typically contains angular information from multiple joints, such as the shoulder joint angle. elbow joint angle wrist joint angle And the displacement increments of the end-effector in the three coordinate axes. , , The corresponding passive attitude vector This includes the corresponding passive prediction values: , , , , , .

[0070] For each joint angle degree of freedom, calculate the difference between the measured value and the passively predicted value to obtain the residual of the joint angle: , , .

[0071] For the spatial displacement of the end-effector, the formula for calculating the residual displacement along each coordinate axis is as follows: , , , in, , , This represents the residual displacement at the end in the corresponding direction.

[0072] Finally, the residual angles of all joints are combined into a set of joint residuals. Combine the end residual displacements in all directions into an end residual vector. .

[0073] This step quantifies the deviation between the observed complex motions and the theoretical passive posture reference through direct algebraic differencing. These residuals theoretically encompass all motion components beyond the passive response, which may include the patient's actual active force exertion, as well as minor fluctuations caused by measurement noise, model errors, and soft tissue deformation, providing the most original data material for subsequent screening and purification.

[0074] S3.2: Construct a mask for determining the active and non-active regions of each degree of motion freedom based on a preset residual threshold.

[0075] During passive stretching, even when the patient is completely relaxed, minute residual fluctuations may occur due to mechanical gaps, sensor noise, or the viscoelastic response of biological tissues. These fluctuations do not originate from active neural drive, and indiscriminately treating all residuals as active contributions would introduce significant errors. Therefore, a reasonable threshold mechanism is needed to identify physiologically significant active components from the residuals.

[0076] A residual threshold is preset for each joint degree of freedom to determine whether the current residual may originate from active force. These thresholds are as follows: for the shoulder joint... elbow joint wrist joint The upper limit of fluctuation in the residual angle of each joint can be determined during the initial passive stretching test before training begins by statistically analyzing the residual angle. Taking the elbow joint as an example, its residual angle threshold is set as follows: The threshold typically ranges from 2 to 8 degrees, and its specific value can be determined by the upper limit of residual fluctuation obtained from the initial purely passive testing phase. Based on this threshold, an active contribution determination mask for the elbow joint is generated. : when hour, ; when hour, , in, A value of 1 indicates that the elbow joint is currently identified as an active region, while a value of 0 indicates that it is identified as a non-active region. This represents the absolute value of the elbow joint residual angle. Similarly, threshold values ​​are set for the shoulder and wrist joints respectively. and The typical value range is 2 to 10 degrees, and a corresponding mask is generated. and The masks of the three joints are combined to obtain the active joint mask set. .

[0077] For the end displacement residual, in order to comprehensively evaluate its overall deviation in three-dimensional space, its average amplitude is first calculated. : , in, This represents the average amplitude of the end residual displacement, in millimeters. A threshold value is set for the end residual amplitude. The typical value range is 3 to 10 millimeters. Based on this threshold, an end-effector active contribution determination mask is generated. : when hour, , when hour, , in, A value of 1 indicates that the end is determined to have made an active contribution at the current moment. This ultimately forms the active end mask. .

[0078] This step employs thresholding to create a binary filter for each degree of freedom of motion. It automatically categorizes small residuals with amplitudes less than the threshold, potentially originating from noise or passive fluctuations, as non-active components, while only marking residuals with amplitudes significantly exceeding the range of passive fluctuations as potential active contributors. This process significantly suppresses interference from spurious active signals, allowing the subsequent extraction of active contributions to focus more precisely on the degrees of freedom of motion that the patient is actually exerting force on and that have clear physiological significance.

[0079] S3.3: Use the decision mask to extract the original active contribution of each degree of freedom from the original residual.

[0080] Based on the joint residual set calculated in step S3.1 Terminal residual vector and the active mask generated in step S3.2 and This allows for the screening of the original residuals and the extraction of the active contribution amount that has passed the threshold test.

[0081] In practice, the residual of each degree of freedom is multiplied by its corresponding mask value. If the mask value is 0, the residual of that degree of freedom is set to zero; if the mask value is 1, the original residual value is retained. Taking the elbow joint as an example, its active contribution angle... The calculation formula is: , When mask When it is 0, regardless of the original residual What is the value, and what is the product result? All values ​​are 0, indicating that the residual of the elbow joint is considered passive fluctuation or noise, and its active contribution is determined to be zero. When the mask... When it is 1, equal to the original residual This indicates that the residual is fully preserved as an active contribution. Active contribution angles of the shoulder and wrist joints. and Similarly, the following calculations can be performed: , , For end-point displacement, a uniform end-point mask is used. Controlling the residuals in three directions. Taking the X direction as an example, it actively contributes displacement. The calculation formula is: , when When it is 0, =0; when When it is 1, equal Active contribution displacement in the Y and Z directions and It can be obtained using the same method.

[0082] Finally, all the active contribution angles and end-effector active contribution displacements that have been filtered by the mask are combined to form the original active contribution vector: .

[0083] This step utilizes a pre-constructed mask to filter the original residuals, eliminating small residuals that are deemed likely to originate from external disturbances or measurement noise, and retaining only the residuals with significant amplitudes that pass the threshold test as the active contributions for each degree of freedom. The resulting original active contribution vector... Structurally clear and explicit, each non-zero component directly corresponds to a degree of freedom of motion identified as having active patient participation, which is a key step in extracting the patient's true active intention from the kinematic coupling signal.

[0084] S3.4: Perform time smoothing and amplitude constraints on the original active contribution amount to generate the final active contribution amount vector.

[0085] Original active contribution vector Undesirable jumps may occur due to instantaneous fluctuations in single-frame data, boundary effects in threshold judgment, or residual noise. To obtain a more stable, smoother output of the active contribution that conforms to physiological and safety constraints, it is necessary to perform time-domain smoothing and impose reasonable amplitude limits. Perform time-series smoothing to obtain the smoothed active contribution vector. .

[0086] Furthermore, to prevent the calculated active contribution from being abnormally large at certain times, it is necessary to... Each component is subjected to amplitude constraints to ensure it remains within a preset safe and reasonable range. All components, after smoothing and amplitude limiting, are ultimately combined to form the final active contribution vector with controlled overall amplitude. .

[0087] This step, through time smoothing, effectively eliminates spikes and jitters in the original data caused by transient interference, making... It can reflect the patient's continuous and stable active exertion trend over a longer time scale. By constraining the amplitude, it ensures that the output active contribution is always within a safe and controllable physical range, avoiding unstable or dangerous commands from subsequent control modules due to individual outliers.

[0088] In the technical solution of this embodiment, by calculating the residual between the actual observed posture and the theoretical passive posture obtained in step S2, and applying a masking filtering mechanism based on a physiologically reasonable threshold, the active contribution generated by the patient's autonomic neuromuscular control is separated from the kinematic signal. After temporal smoothing and safe amplitude constraint processing, a stable active contribution vector that conforms to the physiological range is finally output, solving the coupling problem between passive traction and active force exertion in the observed signal, and laying a direct data foundation for subsequent extraction of the patient's true movement intention.

[0089] S4: Based on the active motion contribution, generate a patient motion intention vector containing direction and intensity information.

[0090] Figure 2 A schematic diagram illustrating the mapping between patient movement intention generation and assistive control commands provided in an embodiment of this disclosure is shown. For example... Figure 2 As shown, the extracted final active contribution vector, distributed across multiple degrees of freedom, is... Through reasonable fusion, classification, and normalization, it is transformed into a digital expression that directly and clearly reflects the patient's rehabilitation training intentions—the actual intention vector. This vector contains two core elements: intention direction and intention intensity, providing a clear and quantifiable decision-making basis for subsequent dynamic auxiliary control. This is achieved through the following sub-steps.

[0091] S4.1: Weighted fusion of the active contributions of multiple joints, compressed into an equivalent active displacement scalar in the stretching direction.

[0092] Final active contribution vector It includes the active angular changes of major joints in the upper limb, such as the shoulder, elbow, and wrist joints. , , And the active displacement changes of the end-effector in three spatial directions. , , In order to use a unified scalar at the control level to describe the overall active force exertion trend of the patient along the predetermined training stretch direction, it is necessary to map and compress the contributions of these multi-degree-of-freedom quantities onto this principal direction to form an equivalent active displacement quantity.

[0093] Assuming the training stretching direction is predefined, for example, the positive Z-axis in a rehabilitation coordinate system represents the stretched limb. By assigning a weighting coefficient to each active contribution component, its equivalent contribution in the stretching direction is calculated and linearly superimposed to obtain the equivalent active displacement scalar. Its calculation formula can be expressed as: , in, Let be the equivalent active displacement scalar at time t, in millimeters. In the field of rehabilitation, this weighted fusion process has clear clinical significance: it comprehensively maps the discrete active angular changes and end-effector displacements of multiple joints such as the shoulder, elbow, and wrist onto a preset rehabilitation training direction (such as flexion / extension), forming a single scalar. This allows the system to ignore differences in motor details and focus directly on the patient's overall effort trend and level of cooperation in the core training direction. It provides key data conversion for generating concise and stable high-level motor intentions, which is in line with the clinical thinking of focusing on function-oriented movements in rehabilitation training. A positive value indicates that the patient's active force tends to cooperate along the direction of stretching, while a negative value indicates that the tendency is opposite to the direction of stretching, thus resisting it. , , These are the equivalent length conversion coefficients corresponding to the active angles of the shoulder, elbow, and wrist joints, respectively. Their physical meaning is the equivalent length change caused by a unit change in joint angle in the stretching direction, with the unit being millimeters per degree, and a typical value range between 1 and 5. The weighting coefficients are for the active displacement of the distal end in the stretching direction, i.e., the Z-direction, and range from 0.5 to 1.5. These weighting coefficients need to be comprehensively set during the individual patient calibration phase, based on the anatomical dimensions of the upper limb, joint kinematic relationships, and specific training task objectives.

[0094] An exemplary integrated calibration procedure may include: guiding the patient to attempt several active movements along the target training direction without external stretching or assistance; and simultaneously recording the changes in the active angles of each joint. , , and the displacement of the end in the stretching direction By using parameter identification methods such as multiple linear regression, a scalar fit is obtained that enables fusion. The combination of weighted coefficients that best represents the overall trend of movement. Evaluation indicators can be used for the fusion results. Consistency with the actual observed end-effector displacement direction and amplitude, such as minimizing the mean square error. This calibration process quantifies the individual patient's kinematic characteristics and training direction into specific weight parameters.

[0095] This linear weighted fusion step aggregates the active contribution information scattered across multiple motion dimensions into a scalar measured along a single principal direction. This scalar provides a unified and physically meaningful benchmark for the subsequent concise and efficient determination of the patient's overall intention direction and intensity.

[0096] S4.2: Determine the direction of motion intention based on the equivalent active displacement scalar and the static dead zone threshold.

[0097] Obtain the equivalent active displacement scalar Then, it needs to be converted into a discrete and clear intention direction classification, which is usually divided into three states: active cooperation along the stretching direction, active resistance in the opposite direction of the stretching direction, and no obvious directional intention, i.e., static.

[0098] To achieve this classification, a static dead zone threshold needs to be set. The unit is millimeters, and its typical value ranges from 1 to 5 millimeters. This threshold is used to filter out displacements with extremely small amplitudes that may be caused by residual noise or physiological micro-tremors, which are insufficient to constitute a clear intention to move. The settings need to take into account the measurement noise levels of the 3D vision system and the robotic arm, as well as the minimum effective active displacement to be detected. They are usually determined through system calibration and pre-experimentation.

[0099] The determination of the intended direction follows these rules: when the equivalent active displacement Greater than the dead zone threshold At that time, it was determined that the patient had an intention to actively cooperate in the direction of stretching; when Less than the negative dead zone threshold At that time, it was determined that the patient had an active resistance intention opposite to the direction of stretching; when The absolute value is less than or equal to the threshold. If no clear directional intention is found, then the intention is considered to be static. This determination process can generate an initial intention direction code. : like ,but , like ,but , like ,but , In this code, 1 represents active cooperation along the stretching direction, such as flexion or elevation; -1 represents active resistance opposite to the stretching direction; and 0 represents no obvious directional intent. Furthermore, to calculate the intensity of intent, the amplitude of the equivalent active displacement, i.e., its absolute value, needs to be obtained. , in, The amplitude of the equivalent active displacement is expressed in millimeters.

[0100] This step, by introducing a dead-zone threshold, achieves a robust mapping from continuous displacement scalars to discrete intention states. It not only outputs a clear intention direction classification, but also... This also provides basic amplitude information for the next step of intensity calculation. This completes the initial conversion from geometric motion quantities to high-level intentional semantics.

[0101] S4.3: Generate a normalized intention intensity based on the equivalent active displacement amplitude and the reference maximum active displacement.

[0102] Intent strength needs to be quantified to determine the patient's level of active participation or effort in the identified direction. This requires calculating the equivalent active displacement amplitude at the current moment. The intensity value is normalized to a value between 0 and 1 by comparing it with a reference value that characterizes the patient’s maximum active ability in that direction.

[0103] Set a reference maximum active displacement The unit is millimeters. This reference value is usually obtained through a specialized test at the beginning of treatment. It reflects the maximum effective displacement amplitude that the patient can generate autonomously in the direction of stretching during a specific training task, and its typical range is between 10 and 50 millimeters. The specialized test refers to a standardized assessment conducted under the guidance of a therapist: the patient is asked to perform several active movements in a predetermined rehabilitation training direction without external assistance, making maximum effort. To ensure the standardization and repeatability of the movements, the test is conducted under a fixed starting posture, clear verbal guidance, and / or visual target cues. To confirm that the patient has made maximum effort and to obtain a reliable reference value, the system records the maximum effective displacement generated in multiple repeated tests and calculates its deviation from the average displacement. Typically, when the displacement amplitude of several consecutive tests tends to stabilize, and the therapist confirms that the patient has made their best effort based on subjective feedback, this stable maximum displacement value can be set as an individualized value. .

[0104] Based on this, the strength of the original intent It can be calculated using the following formula: , in, The original intention strength is a dimensionless value. Theoretically, when achieve At that time, the intensity is 1. To ensure that the intention intensity value is within a stable and reasonable range in subsequent control mappings, it is usually necessary to perform amplitude limiting, forcibly restricting it to a closed interval between 0 and 1: , Value after amplitude limiting This refers to the final strength of the original intent, which will be hereinafter referred to as [insert value hereafter]. Furthermore, as a consistency process, when the original intent direction is encoded... When the value is 0, meaning it is determined to have no obvious directional intent, regardless of the calculated original intensity value, it should be... Setting it to 0 means that the intent strength is also zero when there is no directional intent.

[0105] This step, through normalization, transforms the physical displacement amplitude into an intensity index characterizing the relative degree of effort. This indicator enables subsequent control modules to adjust the magnitude of the assist force linearly or non-linearly based on a standardized intensity value. For example, it can be an adaptive control strategy where a stronger intention results in a weaker assist force, or vice versa.

[0106] S4.4: Perform time smoothing and anti-jitter processing on the intent direction and intensity to generate the final actual intent vector.

[0107] Directly use the original intent direction encoding and the strength of the original intention When a patient's movement approaches the threshold for judgment, it may cause frequent jumps in the direction of intent or drastic fluctuations in intent intensity between adjacent sampling cycles. To improve the stability and continuity of intent output, it is necessary to smooth the direction and intensity on the time axis.

[0108] Strength of original intention Smoothing is performed to obtain the final intent strength. Secondly, the intended direction is smoothed. Direction smoothing focuses on reducing frequent, non-physiological flips between opposite directions caused by slight signal fluctuations. A direction retention threshold is set. Its value ranges from 0 to 1. This threshold is used to maintain the intention direction from the previous moment when the original intention strength is low, in order to avoid frequent non-physiological direction flips caused by slight signal fluctuations. Its specific value depends on the intention strength. The smoothing effect and the system's requirements for direction switching stability are considered when setting the smoothing rule. The smoothing rule prioritizes maintaining the original direction: when a new direction is determined based on the original data... Final direction from the previous moment Different, and the current original intent strength Less than the direction retention threshold If the current strength is insufficient to support a reliable change in direction, then the direction from the previous moment is maintained; otherwise, the original direction is adopted. This rule can be expressed as: like and ,but , otherwise, , in, This is the final smoothed intention direction encoding at time t. Finally, the smoothed intention direction encoding... With Intent Strength Combined, they form the actual intent vector at the current moment: .

[0109] This step smooths and anti-jitters the intensity and direction in the time domain, so that the output intention vector I(t) can reflect the patient's stable and coherent intention state over a longer time scale, avoiding control command oscillations caused by instantaneous signal fluctuations.

[0110] In the technical solution of this embodiment, the active contribution quantities dispersed across multiple degrees of freedom are weighted and fused to a preset stretching direction to generate an equivalent active displacement scalar. Then, a clear motion intention direction is determined based on a preset threshold, and the displacement amplitude is normalized by comparing it with an individual's maximum capability reference value to obtain a quantified intention intensity. Finally, time smoothing and anti-jitter processing are applied to the direction and intensity to form a stable and coherent actual intention vector. This transforms the underlying kinematic separation results into high-level, clear, and directly usable semantic information for control decisions.

[0111] S5: Based on the patient's motion intention vector, generate instructions for controlling the auxiliary torque of the rehabilitation robotic arm.

[0112] like Figure 2 As shown, based on the final intention vector I(t) generated in step S4, which accurately reflects the patient's true rehabilitation intention, a series of mapping, calculation, and smoothing processes are used to generate safe and dynamically adaptable control commands that can directly drive the rehabilitation robotic arm actuator. This enables the auxiliary force and direction to be adjusted smoothly and in real time according to the patient's intention, thereby forming a stable, coordinated, and highly personalized human-machine collaboration during training. This is specifically achieved through the following sub-steps.

[0113] S5.1: Calculate the auxiliary gain coefficient based on the intent strength through linear mapping and amplitude limiting.

[0114] Intent intensity contained in intent vector I(t) , is a normalized quantity ranging from 0 to 1, quantifying the patient's level of active participation in the current direction of intent. To transform this abstract level of participation into a specific parameter that can be quantified as the magnitude of control, a mapping relationship from intent intensity to the assist gain coefficient needs to be constructed. The assist gain coefficient G(t) will serve as the core scaling factor for subsequent calculations of the specific assist force.

[0115] Set the lower limit of the auxiliary gain. With upper limit Both are dimensionless coefficients, with typical values ​​ranging from 0.2 to 1.0. The baseline assistance ratio corresponds to an intention intensity of 0, which is the basic assistance provided by the system when the patient has no active intention at all; This corresponds to the maximum assistance ratio when the intention intensity is 1, which is the upper limit of the assistance provided by the system when the patient shows maximum effort. and The specific value needs to be individually set under the guidance of a therapist, based on the goals of the rehabilitation training model (such as assisted training or resistance training) and the patient's individual motor ability. This is determined according to the intensity of intent. The auxiliary gain coefficient G(t) is obtained through mapping and then restricted to a preset range. Inside.

[0116] This step will reflect the intensity of the patient's subjective effort and intention. This is transformed into a gain coefficient G(t) that can be directly used for mechanical calibration. This mapping provides a core quantitative weight for implementing control strategies such as proportionally decreasing the system's assist force as the patient's intention strengthens, or proportionally increasing the assist force as the patient's intention strengthens in certain training modes. The gain coefficient G(t) lays the foundation for subsequently generating target assist quantities with definite physical dimensions.

[0117] S5.2: Combine the intended direction with the auxiliary gain coefficient and the reference maximum auxiliary torque to generate the target joint auxiliary torque.

[0118] After obtaining the auxiliary gain coefficient G(t), it is necessary to combine it with the intended direction. This generates target joint auxiliary quantities with specific physical meaning and directionality. Intent direction encoding. The value can be -1, 0 or 1, representing an active counter-intention opposite to the direction of stretching, an intention with no obvious direction, and an active cooperative intention along the direction of stretching, respectively.

[0119] Set a reference maximum auxiliary torque The unit is Newton-meters (Nm). This value is selected comprehensively based on the physical driving capability of the rehabilitation robotic arm, the strength of the joint structure, and the upper limit requirements to ensure patient safety, such as through equipment calibration or clinical safety guidelines. A typical range is between 2 and 20 Nm. Target joint assist torque. Calculated by combining the gain factor, direction encoding, and maximum reference torque: .

[0120] This formula is the core embodiment of rehabilitation control, realizing the rehabilitation principle of dynamically matching human-computer interaction torque with the patient's real-time intentions. Specifically, the direction of intention... The direction of the assisting force is determined by the normalized intention intensity, which in turn determines the magnitude of the assisting force. This allows the robotic arm's output torque to continuously and smoothly adapt to the patient's real-time effort and intention direction, achieving true on-demand assistance. This avoids training interruptions due to the patient's lack of strength and provides appropriate challenges when the patient intentionally exerts force, thereby optimizing the neural remodeling process and improving the initiative and safety of rehabilitation training.

[0121] According to the above formula, when hour, A positive value indicates that the system should apply a positive auxiliary torque along the direction of the training stretch to help or assist the patient in completing the movement.

[0122] when hour, A negative value indicates that the system should apply torque in the opposite direction of traction to provide appropriate resistance or braking against the patient's opposing force.

[0123] when hour, A calculation result of 0 indicates that the system does not apply additional active driving torque to the joint, and usually only maintains the necessary gravity compensation or basic support mode.

[0124] This step completes the crucial transformation from abstract intent information and gain coefficients to concrete physical control quantities. The generated... It is a target joint auxiliary torque command with a clear size, direction and unit. It directly specifies which joint the rehabilitation robotic arm should be in, in which direction and how much torque to output, thus providing a clear mechanical target for the execution of the underlying actuator.

[0125] S5.3: Smoothly filter the target auxiliary torque and convert it into actuator commands, while applying safety limits to complete dynamic control output.

[0126] Target joint auxiliary torque This is the ideal command value calculated directly based on the current intent. To avoid sudden changes in this value between adjacent control cycles that could impact actuators such as motors and reducers, or cause patient discomfort, it needs to be smoothed over time to generate the actual output joint torque command. At the same time, this torque command also needs to be converted into a physical signal that the actuator can directly receive and execute.

[0127] Target auxiliary torque Smoothing is performed to obtain the joint torque command to be output. .

[0128] Subsequently, the torque command is converted into a control signal to drive the actuator.

[0129] To ensure absolute safety and comfort during training, and to prevent the output torque from exceeding the limits of soft tissue tolerance or mechanical structure due to any calculation errors or abnormal conditions, the final torque command must be adjusted. Apply safety limits: , in, This is the final torque command after safety limiting; The minimum allowable torque is a negative value used to limit the maximum braking force that the system can provide, typically ranging from -20 Nm. The maximum safe torque limit must be determined jointly based on the patient's joint's maximum safe torque capacity and the rehabilitation robotic arm's physical drive limit, following clinical safety guidelines. The actual current commands issued to the underlying driver should be based on... The conversion yields the result.

[0130] This step achieves a complete and reliable mapping from high-level intent to low-level action by smoothing and limiting the target auxiliary torque and converting it into a physical quantity that can directly drive the actuator.

[0131] In the technical solution of this embodiment, based on the final intent vector, the normalized intent intensity is mapped to an auxiliary gain coefficient, and a target joint auxiliary torque command with specific magnitude and direction is generated by combining the intent direction. This command is then smoothed, filtered, and converted into a physical signal for the drive stage, while a safety limit is applied, thereby achieving dynamic, real-time, and smooth control of the auxiliary torque output by the rehabilitation robotic arm. This step completes a closed loop from patient intent recognition to precise auxiliary execution, enabling the auxiliary force to adapt to the patient's real-time intent and effort level, achieving personalized and intelligent human-machine collaboration.

[0132] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions stored in the memory to execute the methods provided in the above embodiments.

[0133] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0135] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0137] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A dynamic control method for upper limb rehabilitation based on three-dimensional visual intent recognition, characterized in that, The method includes: S1. Simultaneously collect three-dimensional visual observation data of the patient's upper limb and the positional data of the actuator of the rehabilitation robotic arm, construct an upper limb posture vector containing the kinematic state of the upper limb, and associate it with the external stretching displacement increment. S2. Based on the upper limb posture vector and the external stretching displacement increment, generate a passive posture reference vector that represents the theoretical posture of the upper limb under pure passive stretching conditions. S3. Calculate the residual between the actual observed upper limb posture vector and the passive posture reference vector, and separate the patient's active movement contribution from the residual; S4. Based on the active motion contribution, generate a patient motion intention vector containing direction and intensity information; S5. Based on the patient's motion intention vector, generate instructions for controlling the auxiliary torque of the rehabilitation robotic arm.

2. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 1, characterized in that, S1 includes: The three-dimensional visual observation data and the actuator pose data are converted to the same coordinate system; Extract key points representing the position of upper limb joints from the three-dimensional visual observation data; Calculate the joint angles and distal displacements based on the key points to form the upper limb posture vector; The displacement increment along the preset stretching direction is calculated based on the actuator pose data and used as the external stretching displacement increment; The upper limb posture vector at the same moment is correlated with the external stretching displacement increment to form a synchronous data pair.

3. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 1, characterized in that, S2 includes: Based on the upper limb posture vector and the external stretching displacement increment, calculate the proportion of change in the equivalent length of the upper limb caused by a unit stretch; Based on the aforementioned ratio and the aforementioned external stretching displacement increment, the passive angle change of each joint is estimated. The passive attitude reference vector is generated based on the passive angle change.

4. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 3, characterized in that, The estimation of the passive angle change of each joint includes: Based on the aforementioned ratio and the aforementioned external stretching displacement increment, according to Calculate the passive angle change representing the joint. ,in, C(t) represents the percentage change in the equivalent length of the upper limb caused by the unit stretch, and C(t) represents the increment of the external stretch displacement. The equivalent arm length representing the joint is pre-calibrated.

5. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 1, characterized in that, S3 includes: Calculate the residuals of the observed upper limb posture vector and the passive posture reference vector in each degree of freedom; Set a judgment threshold for each component of the residual, and regard the residual component whose magnitude exceeds the judgment threshold as an effective active contribution; The residual components considered as effective active contributions are subjected to time-series smoothing to form a stable active motion contribution.

6. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 5, characterized in that, Setting a judgment threshold for each component of the residual includes: Different judgment thresholds are set for joint angle residuals and end displacement residuals, respectively; Specifically, for the end displacement residual, the average amplitude of the residual displacement in each direction of the three-dimensional space is calculated, and the average amplitude is compared with the preset end amplitude threshold to determine whether there is an effective active contribution at the end.

7. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 1, characterized in that, S4 includes: The contributions of multiple joints in the active motion contribution are merged into a scalar displacement along the stretching direction. The direction of the intended motion is determined based on the sign and magnitude of the scalar displacement. The normalized intention intensity is calculated based on the magnitude of the scalar displacement and the preset reference value. The intention direction and intention intensity are subjected to anti-shaking processing to form the patient's motion intention vector.

8. The method for dynamic control of upper limb rehabilitation based on three-dimensional visual intent recognition according to claim 1, characterized in that, S5 includes: Map the intensity information in the patient's motion intention vector to an auxiliary gain coefficient; The target auxiliary torque is calculated by combining the auxiliary gain coefficient, the intended direction information, and the reference maximum auxiliary torque. The target auxiliary torque is smoothed, filtered, and limited for safety to generate the final control command.

9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the upper limb rehabilitation dynamic control method based on three-dimensional visual intent recognition as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, It stores a computer program, which, when executed, implements the upper limb rehabilitation dynamic control method based on three-dimensional visual intent recognition according to any one of claims 1-8.

Images