Adaptive control method and system for dynamic pose tracking of back massaging robots

Abstract

The application discloses a dynamic posture tracking adaptive control method and system of a back massage robot, and relates to the field of intelligent rehabilitation robot control, and comprises the following steps: S1, acquiring multi-modal posture sensing data of a human back; S2, information fusion is performed on the multi-modal posture sensing data; S3, a dynamic parameterized model related to the center line of the human spine is constructed; S4, a short-time dynamic prediction model of the human posture is established based on the dynamic parameterized model and the human posture state vector at the current and historical moments and optimal control point parameters; S5, the motion and force control of a massage execution mechanism are adjusted according to the predicted posture result, and a complete closed-loop adaptive control mechanism is formed through feedback and updating. The application can accurately and continuously track massage points in the dynamic changes of human respiration fluctuations and trunk movements, effectively avoids the massage deviation problem caused by posture changes, and greatly improves the stability, robustness and anti-interference ability of posture sensing modeling in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent rehabilitation robot control. More specifically, this invention relates to an adaptive control method and system for dynamic posture tracking of a back massage robot. Background Technology

[0002] With the integration of automated upgrades and robot control technology in massage equipment, massage equipment has evolved from manual mechanics to electric and intelligent systems. However, the core challenge remains "how machines can adapt to dynamic changes in the human body." This need has driven the exploration of technologies such as multimodal perception and dynamic modeling in this field.

[0003] Existing technologies have also attempted to improve the machine by using multi-sensor combinations, simple force feedback control, and static spinal modeling, in order to make the machine adapt to the dynamic changes of the human body. For example, a back massage robot control system and method with patent number CN110308682B is a multi-sensor loose combination massage technology, that is, it collects data by using multiple sensors such as gyroscopes, accelerometers, and infrared sensors. However, this technology does not achieve deep fusion of multimodal data, the perception dimension is fragmented, it cannot simultaneously acquire the three-dimensional shape of the back, changes in trunk posture, and details of contact force, and it lacks a posture prediction mechanism, and can only passively respond to posture changes, resulting in insufficient safety. For example, a massage robot and its massage method based on a dual-electrode massage head, patent number CN118986704A, is only a static scanning massage technology. It generates a two-dimensional trajectory of the back through a single scan by a depth camera. During the massage, it only relies on a force sensor to fine-tune the contact force. It does not establish a dynamic model of the human spine, has a single feedback mechanism, cannot adapt to natural movements such as breathing and turning over, is prone to trajectory deviation, and has poor robustness. This shows that existing technologies have not formed a complete adaptive solution and are difficult to cope with complex dynamic postures; simple force feedback control does not combine the dynamic adjustment of stiffness with the spinal morphology, and the force adjustment lacks specificity; static spinal modeling cannot be updated in real time with posture changes, resulting in a lag in trajectory adjustment. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0005] To achieve these objectives and other advantages of the present invention, an adaptive control method for dynamic posture tracking of a back massage robot is provided, comprising: S1. During the massage, multimodal posture perception data of the human back is obtained through the human posture perception unit. S2. Perform information fusion on multimodal posture perception data to obtain a unified human posture state; S3. Construct a dynamic parameterized model related to the centerline of the human spine based on human posture. S4. Based on the dynamic parameterization model and the human posture state vector X at the current and historical moments ( t ) Optimal control point parameters Establish a short-time dynamic prediction model for human posture; S5. Adjust the motion and force control of the massage actuator based on the predicted posture results of the short-time dynamic prediction model, and feed back the massage execution results to S2, S3, and S4 to continuously update the human body posture state, dynamic parameterization model, and short-time dynamic prediction model, so as to form a complete closed-loop adaptive control mechanism.

[0006] Preferably, in S1, the multimodal pose perception data includes: three-dimensional point cloud data of the human back. D ( t ), human torso posture perception data q( t Massage contact force sensing data F ( t ) 。

[0007] Preferably, in S2, the information fusion processing flow includes: S20. Multimodal feature extraction is achieved by preprocessing the multimodal pose perception data; S21. Define various features at time t using the following formula. t Fusion confidence: In the above formula, , For the first k The preset weight coefficients of class features, and satisfying and , For the first k Class features at time t The reliability indicators, and exp(∙) is an exponential function. Indicates the first k Smoothing adjustment coefficient of class features For the first k The amount of change in class features between adjacent time points; S22. Based on the fusion confidence levels corresponding to various features, construct the following human pose state vector X( t ): In the above formula, p ( t ) indicates and D (t The relevant three-dimensional position vector, To p ( t The rate of change of the position of key areas on the back was obtained by performing a time difference operation. i ( t )and q ( t The relevant attitude direction vector, f ( t ) for and F ( t Related force perception state quantities, To and p ( t ), , i ( t ), f ( t The related vector set.

[0008] Preferably, in S3, the original spinal centerline dynamic parameterization model C ( s,t The following formula represents the continuous parametric curve of the human spine over time: In the above formula, C ( s , t ) is through time t Normalized arc length parameter s Three-dimensional spatial parameter curves are used to characterize the dynamic spatial morphology of the spine. B i ( s To ensure the continuity and smoothness of the spinal curve, the third-order B-spline basis function is used. M for B The number of control points for a spline curve u i ( t (time) t The lower B-spline curve i The three-dimensional coordinates of each control point, and u i ( t The control points are dynamically adjusted using the following formula to ensure that the spinal model continuously follows changes in human posture: In the above formula, For the first i The initial three-dimensional coordinates of the control points The time step between two consecutive data acquisitions. This represents the attitude change in the attitude direction fused between adjacent time steps. for The corresponding incremental rotation matrix, l i In the local coordinate system of the human body, the first i Each control point has a preset fixed offset vector relative to the spinal reference point.

[0009] Preferably, in S3, the dynamic parameterization model is implemented through a comprehensive energy function. E Minimizing E is the optimization objective. The optimal control point parameters will be obtained by solving E using the gradient descent method. Substitution C ( s , t Then, the constrained optimization of the dynamic parameterized model is achieved through the following formula: In the above formula, C ( s , t ) * For the constrained optimized dynamic parameterized prediction model of the spinal centerline, the comprehensive energy function E is characterized by the following equation: In the above formula, E data To ensure that the model accurately reflects the current morphology of the human spine, and , N This represents the total number of point cloud sampling points. p i ( t () indicates time t Lower back surface i The three-dimensional spatial coordinates of each sampling point s i For the first i The specific values ​​of the normalized arc length parameter corresponding to the spinal centerline for each back point cloud data point are taken. . E smooth To ensure the smoothness constraint term maintains the continuous and smooth morphological characteristics of the spinal curve, and , For the spinal curve C ( s , t The second derivative with respect to the normalized arc length parameter s, E bio To avoid generating physiological consistency constraints on spinal morphology that exceed the physiological limits of the human body, and , For a moment t Below is a dynamic parametric model of the original spinal centerline.C ( s , t In the normalized arc length parameter s The local curvature at that point This is the maximum curvature threshold allowed by the physiological structure of the human spine. l 1. l 2 are the smoothing constraint terms. E smooth Physiological consistency constraint E bio The corresponding weighting coefficients.

[0010] Preferably, in S4, the short-time dynamic prediction model is characterized by the following formula: In the above formula, C ( s , t +1) * The mapping function is a dynamic parameterized prediction model of the spinal centerline at time t+1. f (∙) represents the state transition mapping based on the first-order Markov property, which is then used to predict the attitude state through extended Kalman filtering. With the current optimal control point This is mapped to a dynamic parameterized prediction model of the spinal centerline at the next moment. for t The predicted human posture state at time +1, and during the prediction process, if the comprehensive energy function corresponding to the candidate spine model of the predicted posture... E If the value exceeds the preset threshold, the predicted state will be... Make corrections to ensure The prediction results are consistent E Minimize the constraints.

[0011] Preferably, in S5, the local curvature of the spine is calculated using the following formula. Construct a variable stiffness impedance control model that adaptively adjusts with curvature. F d At the same time, combined with the comprehensive energy function E The safety threshold is determined in advance, and the massage trajectory adjustment amount and contact force control parameters are corrected in advance, thereby adjusting the motion and force control of the massage actuator: In the above formula, , These are the dynamic parameterized prediction models for the spinal centerline. For normalized arc length parameter s The first and second derivatives, Satisfying S3 Ebio Physiological threshold of constraint ; Variable stiffness impedance control model F d It is characterized by the following formula: In the above formula, H Here is the damping coefficient matrix. To follow the local curvature of the spine An adaptively adjusted stiffness coefficient matrix, where the stiffness coefficients vary with... The increase is exponentially decayed when , And the comprehensive energy function E Reduce stiffness coefficient when exceeding preset safety threshold , x d Let S4 be the human pose position vector predicted by S4, and , x This is the actual position vector of the massage head. , These are the first derivatives of the desired position and the actual position, respectively.

[0012] An adaptive control system for human dynamic posture tracking includes: Human posture sensing unit, which is used to acquire multimodal posture sensing data of the human back during the massage process; The posture information fusion unit is used to fuse the multimodal posture data output by the human posture perception unit. The spinal dynamic modeling unit is used to construct a dynamic parametric model of the human spinal centerline based on the fused human posture state. The posture prediction and control unit is used to predict the trend of human posture changes and generate control commands based on the prediction results. A massage actuator that executes control commands generated by the posture prediction and control unit; The closed-loop feedback and safety control unit is used to feed back the execution results of the massage actuator to the posture information fusion unit, the spinal dynamic modeling unit, and the posture prediction and control unit, continuously updating the human posture state, dynamic parameterization model, and short-time dynamic prediction model to build a complete closed-loop adaptive control mechanism; at the same time, it sets multi-dimensional safety constraints and executes safety protection strategies when an anomaly is triggered. The human posture sensing unit includes: A visual perception submodule for acquiring three-dimensional morphological information of the human back; An inertial posture perception submodule used to collect information on changes in the posture of the human torso during massage. A force sensing submodule is installed on the massage actuator to obtain force feedback information during the contact between the massage head and the back of the human body.

[0013] The present invention has at least the following beneficial effects: Firstly, compared with existing technologies, this invention has significant advantages in human dynamic posture tracking and modeling. By integrating multimodal posture perception data of the human back and combining it with an adaptive weight adjustment mechanism, it can continuously update the posture state and the dynamic parameterized prediction model of the spinal centerline in dynamic changes such as human breathing fluctuations and trunk movement. It can accurately and continuously track massage points, effectively avoid massage deviation caused by posture changes, and at the same time greatly improve the stability, robustness and anti-interference ability of posture perception modeling in complex scenarios.

[0014] Secondly, compared to solutions that treat the back as a rigid plane or rely on a single sensor, the dynamic parameterized spinal model constructed in this invention conforms to the physiological characteristics of the human body and the natural morphological laws of the spine, significantly improving the matching degree between the massage trajectory and the actual structure of the back.

[0015] Thirdly, this invention further optimizes the massage effect and system reliability through dynamic posture prediction and closed-loop control mechanisms. Based on short-term prediction trends of posture data, the trajectory and contact force of the massage actuator are adjusted in advance to compensate for control lag. At the same time, the trajectory and contact force parameters are adjusted in a coordinated manner to avoid excessive local pressure, thereby improving massage comfort and safety.

[0016] Fourth, the closed-loop feedback mechanism of this invention can continuously correct the posture state, model and control parameters, ensuring long-term operational reliability. Moreover, all the technologies used are based on existing hardware and do not require special customized equipment. It has strong engineering feasibility and can be extended to various massage devices, thus possessing high industrialization value.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the adaptive control method for dynamic posture tracking of the back massage robot of the present invention. Figure 2 This is a schematic diagram of the composition of the adaptive control system for human dynamic posture tracking of the back massage robot of the present invention. Figure 3 This diagram illustrates a comparison of the average trajectory tracking error of the method of the present invention and the traditional control method of first static scanning and then single force reaction when running continuously for 10 minutes in a dynamic scene. Figure 4This diagram illustrates a comparison of the contact force fluctuations of the method of this invention and the traditional control method of first static scanning and then single force reaction when running continuously for 10 minutes in a dynamic scenario. Figure 5 This diagram illustrates a comparison of the response delay of the method of this invention and the traditional control method of first static scanning and then single force reaction when running continuously for 10 minutes in a dynamic scenario. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0020] An adaptive control method for dynamic posture tracking of a back massage robot, which utilizes multimodal posture perception, spinal dynamic modeling, and short-term prediction mechanisms, enables the massage robot to maintain the accuracy and safety of massage position and intensity even during natural human movement. Figure 1 As shown, the specific steps include: Step 1: Acquire multimodal pose perception data of the human back During the massage process, multiple sensors installed on the massage robot body or external support are used to synchronously collect the posture information of the human back, thereby obtaining multimodal posture raw perception data.

[0021] The multimodal attitude sensing data includes: The 3D shape perception data of the back is obtained by a depth vision sensor, which acquires 3D point cloud data of the human back. D ( t ),and ,in, Indicates the time t of the lower back surface. i The three-dimensional spatial coordinates of each sampling point.

[0022] Human torso posture perception data is acquired by an inertial measurement unit. q ( t (Among them, human torso posture perception data) q ( t ) is a quaternion, and Used to characterize changes in the spatial orientation of the human torso.

[0023] Massage contact force sensing data is collected by a force sensor located at the massage head, which gathers data on the massage contact force between the massage head and the user's back. F ( t (i.e., normal contact force).

[0024] After time synchronization and coordinate system processing, the aforementioned multimodal sensing data constitutes the original human posture sensing set at time t, providing the basic input for subsequent posture fusion.

[0025] Step two: Perform multimodal pose information fusion to form a unified human pose state. Feature extraction and fusion processing are performed on the raw multimodal pose perception data obtained in step one to construct X( t A unified description of human posture.

[0026] (1) Multimodal feature extraction From point cloud data X ( t By segmenting the point cloud of the human back region from the image, and calculating the average coordinates of all points in this region, the geometric center of the key back region, i.e., the three-dimensional position vector, is obtained. p ( t ): in, N This represents the total number of point clouds in the region.

[0027] Then to p ( t The rate of change of position of key areas on the back was obtained by performing time difference operations. This is used to characterize the velocity and direction of movement of the human back in three-dimensional space. Its calculation formula is... p ( t -1) is the three-dimensional position vector of the key back region at the previous time t−1. It is the time step between two consecutive data collections.

[0028] The posture quaternion of human torso posture perception data Convert to rotation matrix R (q), and then obtain Euler angles through ZYX sequential decomposition. This constitutes the attitude direction vector. , representing the three-dimensional orientation of the human torso.

[0029] Massage contact force sensing data F ( t After normalization, the force perception state quantity is obtained. f ( t ); in, For the modulus of contact force, F min and Fmax These are the minimum and maximum ranges of the force sensor, respectively, after normalization. .

[0030] (2) Calculation of fusion confidence Based on the continuity of visual data, the smoothness of posture changes, and the stability of contact force, adaptive weights are assigned to various features, defining the first... k The confidence level of a class feature is: In the pose fusion scenario of this patent, there are a total of 4 types of features: p ( t ), , i ( t ), f ( t ).

[0031] in, For the first k The fusion confidence of class features at time t satisfies ; For the first k The preset weight coefficients of class features reflect the prior importance of the features and satisfy the following conditions: and ; For the first k The reliability index of the class feature at time t. The value range is [0,1]. The closer the value is to 1, the more stable and reliable the current data is; For the first k The inter-frame variation of the feature class at adjacent time points, specifically for the four feature classes in this patent, takes the following form: For the first k The smoothing adjustment coefficient of the class feature is a constant with a value greater than 0. It is used to control the degree of influence of feature variation on reliability index. Its specific value is determined according to the actual scenario and experimental results. The sum of the weighted reliability of the four types of features is used to normalize the confidence level.

[0032] The confidence levels corresponding to the four types of features in this invention are as follows: , , , .

[0033] (3) Unified construction of human posture state Based on the above weights, the features of each modality are fused to construct a unified human pose state vector: The human body's pose state is represented by a 10-dimensional column vector, where: , , All three are 3-dimensional vectors. It is a 1-dimensional state variable.

[0034] As a unified expression of human posture at time t, it integrates the real-time reliability and prior importance of multimodal data, and can serve as the sole input for subsequent spinal modeling and posture prediction.

[0035] Step 3: Construct a dynamic parametric model of the human spine's centerline. Based on the human posture state obtained in step two X ( t This study models the human spine structure by abstracting the spine as a continuous parametric curve that changes over time, and constructs a dynamic parametric model of the original spinal centerline. C ( s , t Its expression is: in, C ( s , t This is a dynamic parameterized model of the original spinal centerline, a three-dimensional spatial vector function that outputs parameters of the spine at different times t and different normalized arc lengths. The three-dimensional coordinates at the location directly depict the dynamic spatial morphology of the spine.

[0036] B i ( s ) is a 3rd-order B-spline basis function used to ensure the continuity and smoothness of the spinal curve, satisfying the physiological curvature characteristics of the human spine.

[0037] M represents the number of control points for the B-spline curve, which is set according to the physiological structural characteristics of the human spine.

[0038] u i ( t ) represents the first B-spline curve at time t. i The three-dimensional coordinates of each control point are determined by a unified human posture. X ( t Initialize, and follow X ( tThe dynamic adjustment is based on real-time updates, ensuring that the dynamic parameterized model of the spinal centerline remains highly consistent with the current real-time posture of the human body in terms of overall position and orientation.

[0039] Control Points u i ( t Initialization and dynamic updates of ) (1) Initialization calculation based on X ( t Initialize the B-spline control points to ensure the initial position and orientation of the spine model are perfectly aligned with the current state of the human body. The calculation formula is as follows: in, For the first i Initial three-dimensional coordinates of each control point; To be based on the fusion attitude direction X θ ( t The resulting three-dimensional rotation matrix represents the overall spatial orientation of the human torso. l i In the local coordinate system of the human body, the first i Each control point has a preset fixed offset vector relative to the spinal reference point, which is evenly distributed along the physiological direction of the spine and is preset according to the anatomical characteristics of the human spine.

[0040] (2) Real-time dynamic updates Follow X ( t The real-time changes of the spinal model are used to dynamically adjust the control points, ensuring that the spinal model continuously follows changes in human posture. The calculation formula is as follows: in, The time step between two consecutive data acquisitions; This represents the attitude change in the attitude direction fused between adjacent time points; This is the incremental rotation matrix corresponding to the posture change, used to correct the overall orientation of the spine model.

[0041] To ensure that the constructed dynamic parametric model of the spinal centerline accurately fits the point cloud data of the human back while strictly adhering to the physiological structural characteristics of the human spine, and to avoid problems such as spatial bending and morphological distortion, a comprehensive energy function is introduced to constrain and optimize the model. Its expression is as follows: In the above formula, the constraint weight coefficient l 1. l2 are both constants greater than 0, and are the weight coefficients of the smoothing constraint term and the physiological consistency constraint term, respectively. They are used to balance the influence of each constraint on the model. Their specific values ​​are determined by debugging based on the actual scenario and experimental results.

[0042] Data Consistency Items E data The formula used to constrain the fit between the spinal model and the actual point cloud data of the human back, ensuring that the model accurately reflects the current spinal morphology, is as follows: In the above formula, N This represents the total number of point cloud sampling points. p i ( t () indicates time t Lower back surface i The three-dimensional spatial coordinates of each sampling point s i For the first i The specific values ​​of the normalized arc length parameter corresponding to the spinal centerline for each back point cloud data point are taken. .

[0043] Smoothing constraint E smooth The formula used to suppress drastic bending and discontinuous changes in the spinal centerline in space, thus maintaining a continuous and smooth morphological characteristic of the spinal curve, is as follows: In the above formula, For the spinal curve C ( s , t For the normalized arc length parameter s The second derivative of .

[0044] Physiological consistency constraint E bio The formula used to constrain the spatial shape of the spinal centerline to conform to the natural physiological curvature of the human spine and avoid generating a spinal shape that exceeds the physiological limits of the human body is as follows: In the above formula, For time t, the original dynamic parameterized model of the spinal centerline C ( s , t In the normalized arc length parameter s The local curvature at that point The maximum curvature threshold allowed by the physiological structure of the human spine is set based on human anatomical characteristics; Indicates only the local curvature When this term is non-zero, it will increase the overall energy. E This is to punish model shapes that exceed physiological limits; if the curvature is within physiological limits and This item is 0, so there is no penalty.

[0045] Based on the comprehensive energy function E Minimize as the optimization objective, and solve using the gradient descent method. E The optimal solution yields the optimal control point parameters of the B-spline curve. Substitute it into the dynamic parameterization model expression of the spinal centerline. The original control point parameters were replaced to update the dynamic parameterized model of the spinal centerline, resulting in a constrained optimized dynamic parameterized model of the spinal centerline. C ( s , t ) * To achieve precise depiction of the dynamic spatial morphology of the human spine.

[0046] Step 4: Dynamically predict human posture Based on the comprehensive energy function in step three E Constrained optimization of the optimal dynamic parameterized model of the spinal centerline and the current and historical human posture. X ( t ) and optimal control point parameters A short-time dynamic prediction model for human posture is established. Within the short-time prediction window, assuming that changes in human posture satisfy the first-order Markov property, a posture state transition model is then constructed: Where A is the attitude state transition matrix, which represents the evolution of the attitude state from time t to t+1; g ( t ) represents zero-mean Gaussian process noise, used to model uncertainties in attitude changes.

[0047] Using an extended Kalman filter to predict the pose state, we obtain the predicted human pose state at future time points: During the prediction process, a comprehensive energy function is introduced simultaneously. E The constraint logic is: if the energy of the candidate spine model corresponding to the predicted posture is... E If the value exceeds the preset threshold, the predicted state will be... Make corrections to ensure the prediction results are consistent. E Minimize the constraints.

[0048] The corrected predicted attitude state Mapped to S3 E The optimal dynamic parameterized model of the spinal centerline after constraint optimization is used to obtain the dynamic parameterized prediction model of the spinal centerline (hereinafter referred to as the prediction model): Predictive Model C ( s , t +1) * The comprehensive energy function must be satisfied. E The constraints ensure that the dots conform to the human back, are smooth and continuous, and conform to physiological characteristics.

[0049] Step 5: Adjust the massage trajectory and contact force in real time based on the predicted posture results. Based on the prediction model obtained in step S4 C ( s , t +1) * Calculate the local curvature of the spine: in, , These are the dynamic parameterized prediction models for the spinal centerline. For normalized arc length parameter s The first and second derivatives, and curvature Must meet S3 E bio Physiological threshold of constraint This ensures that the curvature calculation conforms to the anatomical characteristics of the human spine.

[0050] To maintain safety under different curvatures, a variable stiffness impedance control model is established. F d : in, Here is the stiffness coefficient matrix, which varies with the predicted curvature. Increases and decays exponentially, when , And the comprehensive energy function E Reduce stiffness coefficient when exceeding preset safety threshold To prevent applying excessive force to areas where the spine protrudes or is close to its physiological limits; H This is the damping coefficient matrix, used to suppress system oscillations and ensure a smooth massage process; Using the attitude prediction results from S4, combined with the comprehensive energy function E The constraint logic enables feedforward-oriented early correction, rather than the passive response adjustment of traditional solutions: Trajectory Adjustment: Based on predicted changes in spinal position, the target trajectory of the massage head is corrected in advance. x d This ensures that the massage head remains in contact with the target massage area as the body breathes and the body moves slightly, avoiding problems such as the massage head being suspended or deviating from its trajectory; Force-controlled adjustment: Adjusts the stiffness coefficient in real time based on predicted local curvature of the spine. Simultaneously correct the desired contact force F d If the comprehensive energy function corresponding to the predicted spinal model is... E If the stiffness coefficient is exceeded, the stiffness coefficient will be further reduced. Avoid contact force exceeding physiological safety limits.

[0051] The adjustment process is based on the optimal dynamic parameterized model of the spinal centerline optimized by the comprehensive energy function E in S3, and uses the short-time posture prediction results in S4 as feedforward to compensate for the response lag problem of the traditional control scheme, and finally realizes the adaptive and safe closed-loop control of massage trajectory and contact force.

[0052] Using the attitude prediction results from S4, combined with the comprehensive energy function E The constraint logic pre-corrects the massage trajectory position and contact force parameters: if the comprehensive energy corresponding to the predicted spinal model is... E If the stiffness coefficient is exceeded, the stiffness coefficient will be further reduced. This ensures that the massage head remains in contact with the target massage area as the body moves and changes posture, while avoiding exceeding the physiological safety range.

[0053] This adjustment process is based on S3. E Based on the constrained optimal spinal model, and with the predicted posture of S4 as feedforward, adaptive and safe control of massage trajectory and contact force is achieved.

[0054] Step 6: Form a closed-loop adaptive control and continuously update it. The massage execution results in step five, including actual contact force, massage trajectory deviation, and real-time perceived changes in human posture, as well as the deviation between the actual posture and the predicted posture, are fed back to the point cloud preprocessing and posture perception module in step two, the dynamic parameterization model optimization of the spinal centerline in step three, and the human posture dynamic prediction module in step four. This continuously updates the human posture state, spinal model parameters, and prediction model, forming a complete closed-loop adaptive control mechanism.

[0055] In the third module, the integrated energy function is recalculated based on the actual attitude and point cloud data from the feedback. The optimal control point is updated iteratively using the gradient descent method. The updated dynamic parameterization of the spinal centerline is obtained. To ensure that the model always closely matches the real-time state of the human body and meets the requirements. E Minimize constraints; in module S4, based on the updated spine model and actual posture deviation, correct the posture state transition matrix A and noise covariance parameters to optimize the accuracy of the short-time dynamic prediction model; in module S2, adjust the thresholds of point cloud preprocessing and posture perception according to feedback deviation to improve the reliability of input data.

[0056] When the deviation between the magnitude of posture change, the actual contact force and the expected force is detected, or the comprehensive energy corresponding to the spinal model is detected. E When the preset safety threshold is exceeded, the system automatically reduces the massage intensity, adjusts the trajectory, or triggers a safety rollback, while simultaneously updating... E Physiological threshold Energy safety thresholds are set to ensure safe use.

[0057] This invention integrates multimodal perception, dynamic modeling, and predictive control technologies into the field of massage, overcoming the shortcomings of traditional technologies that are singular, static, and passive. By using dynamic spinal modeling, posture prediction, and closed-loop control, it achieves the goal of machines actively adapting to the dynamic changes of the human body, thus making up for the lack of precise control technology for massage devices under natural movement conditions.

[0058] The adaptive control for human dynamic posture tracking in this invention achieves both human dynamic posture tracking and adaptive massage control through the coordinated operation of various units, such as... Figure 2 As shown, it includes: 1. Human posture sensing unit: The human posture sensing unit is used to acquire multimodal posture sensing data of the human back during the massage process. It is the data input terminal of the entire system and includes: ①Visual Perception Submodule It is used to collect three-dimensional morphological information of the human back, obtain back point cloud data and depth information, so as to reflect the spatial contour features of the human back.

[0059] ②Inertial Attitude Sensing Submodule It is used to collect information on the posture changes of the human torso during the massage process, and to obtain posture angle, angular velocity and posture quaternion data to characterize the overall posture change trend of the human body.

[0060] ③ Force perception submodule It is installed on the massage actuator to obtain force feedback information during the contact process between the massage head and the back of the human body, so as to reflect the contact state and force conditions.

[0061] The human posture perception unit outputs multi-source heterogeneous perception data to the posture fusion processing unit through time synchronization and coordinate system unified processing.

[0062] 2. Attitude information fusion unit: The posture information fusion unit is used to fuse the multimodal posture data output by the human posture perception unit. In its specific implementation, it constructs a unified human posture state description model, maps posture information from different dimensions such as vision, inertia, and force perception to a unified posture state space, and adaptively adjusts the fusion weights based on the real-time reliability of various perception data, thereby outputting a stable and continuous 10-dimensional human posture state vector. X ( t This provides a unique standardized input for subsequent dynamic modeling of the spine.

[0063] 3. Spinal dynamic modeling unit: The spinal dynamic modeling unit is used to construct a dynamic parametric model of the human spine centerline based on the fused human posture state. In its specific implementation, the human spine is abstracted as a time-varying 3rd-order B-spline continuous parametric curve, based on the human posture state. X ( t The control points were initialized and dynamically updated; a comprehensive energy function was introduced during the modeling process. E With energy minimization as the optimization objective, the optimal control point parameters of the B-spline curve are solved by gradient descent, generating a dynamic parameterized model of the spinal centerline that simultaneously satisfies point cloud fit, spatial smoothness, and physiological structural consistency.

[0064] The dynamic parameterized model of the human spine centerline output by the spinal dynamic modeling unit serves as an important basis for subsequent posture prediction and massage trajectory planning.

[0065] 4. Attitude Prediction and Control Unit The posture prediction and control unit is used to predict the trend of human posture changes and generate control commands based on the prediction results, including: ① Attitude dynamic prediction submodule Based on current and historical human posture states and optimal spinal model parameters, a posture state transition model satisfying first-order Markov properties is constructed. An extended Kalman filter is used to predict the short-term posture change trend of the human body. A comprehensive energy function is simultaneously introduced during the prediction process. E The constraint logic corrects candidate predicted postures that exceed the energy threshold, and finally outputs a posture prediction result that conforms to human physiological constraints and a dynamic parameterized model of the predicted spinal centerline. ② Adaptive Control Submodule Based on human posture prediction results and a dynamically parameterized model of the spinal centerline optimized by energy constraints, the local curvature of the spine is calculated in real time. A variable stiffness impedance control model that adaptively adjusts with curvature is constructed, and a comprehensive energy function is also incorporated. EThe safety threshold is determined in advance, and the massage trajectory adjustment amount and contact force control parameters are corrected to guide the motion and force control of the massage actuator.

[0066] 5. Massage execution mechanism The massage actuator is used to execute control commands generated by the posture prediction and control unit. The massage actuator includes a massage head, a drive device and an execution control interface. It is used to perform massage operations on the back of the human body according to the adjusted trajectory, and continuously collect contact force and actual position data during the massage process, and output them to the closed-loop feedback and safety control unit.

[0067] 6. Closed-loop feedback and safety control unit The closed-loop feedback and safety control unit is used to construct the system's closed-loop adaptive control mechanism. In specific implementation, the closed-loop feedback and safety control unit receives the massage execution results and real-time sensing information, and feeds back the deviation between the actual posture and the predicted posture, and the deviation between the actual contact force and the expected force, to the posture information fusion unit, the spinal dynamic modeling unit, and the posture prediction and control unit, continuously correcting the fusion weights and the comprehensive energy function. E The relevant parameters, optimal control points, attitude state transition matrix, and control parameters are determined; simultaneously, multi-dimensional safety constraints are set, and when abnormal attitude, abnormal contact force, or the comprehensive energy corresponding to the spine model is detected... E When the preset safety threshold is exceeded, the massage intensity will be automatically reduced, the trajectory will be adjusted, or a safety retraction will be triggered to ensure safe use.

[0068] Example: The adaptive control method for back massage based on human dynamic posture tracking is implemented as follows: 1. Acquisition of human back posture data Input hardware and parameters: A depth vision sensor acquires a depth stream at a resolution of 640×480 and a frame rate of 30Hz to obtain 3D point cloud data of the back; an IMU (Inertial Measurement Unit) acquires attitude quaternions with a sampling rate of 100Hz to obtain torso posture changes; a six-dimensional force sensor at the end of the massage head acquires normal contact force data with a sampling rate of 500Hz. All data are time-synchronized and processed with a world coordinate system to form a multimodal raw perception dataset.

[0069] 2. Human posture fusion processing The S2 adaptive confidence-weighted fusion algorithm is used to calculate the 10-dimensional unified human pose state at time t. X ( t The core parameters are as follows: Back center reference point p ( t ) is in the world coordinate system (320 mm 240 mm 500 mm ); Rate of change in back position For (0) mm / s 0 mm / s -80 mm / s ), used to simulate the downward trend of respiration; Forward tilt of the torso i ( t The angle is 3.5°, with no lateral flexion or torsion; Current normalized contact force f ( t The value is 0.4, corresponding to an actual contact force of 18N, and the force sensor range is 5-45N.

[0070] 3. Dynamic modeling of the spinal centerline Based on the fused human posture state X ( t A dynamic parametric model of the spine's centerline using a third-order B-spline was constructed, and the number of control points was set. M =5, complete control point initialization and dynamic update; introduce comprehensive energy function. E Set weight coefficients l 1 = 0.6 l 2 = 0.8, physiological maximum curvature threshold The optimal control point is obtained iteratively by using the gradient descent method with the goal of minimizing energy. The coordinates of the core control points are: .

[0071] Finally, an energy-constrained optimized prediction model is generated. C ( s , t ) * The curve smoothly fits the shape of the back spinal groove, while satisfying the requirements of fit, smoothness and physiological constraints.

[0072] 4. Dynamic prediction of human posture Based on the optimal spine model and historical posture sequences, a first-order Markov posture state transition model is constructed. Short-time prediction is performed using an extended Kalman filter with a prediction window of 50ms, corresponding to the next time step. t +1; During the prediction process, the comprehensive energy corresponding to the candidate predicted attitude is simultaneously verified. E After confirming that it is within the safety threshold, the final prediction result is output: The center point of the back will shift 4mm in the negative Z-axis direction, corresponding to the downward movement during breathing; the local curvature of the thoracic spine. It will be made from 0.02mm -1 Increased to 0.0212mm -1 Increased by approximately 6%; predictive models were generated simultaneously.C ( s , t +1) * .

[0073] 5. Massage trajectory and contact force adaptive adjustment Based on prediction model C ( s , t +1) * Calculate the local curvature of the spine and construct a variable stiffness impedance control model: Trajectory correction: The target position of the massage head is automatically moved down 4mm along the negative Z-axis to track the predicted breathing depth in advance and prevent the massage head from being suspended in the air; Force control adjustment: Due to the detection of increased local curvature of the spine, approaching the thoracic vertebral protrusion, the resistance stiffness coefficient... The target contact force is automatically and exponentially reduced, smoothly decreasing from 18N to 16N, while simultaneously verifying the overall energy of the prediction model. E Once the parameters are within the safe threshold, the control parameters are finally locked.

[0074] 6. Closed-loop feedback and continuous updates During the massage process, the actual positional deviation, contact force deviation, and real-time posture changes are fed back to the front-end units: The spinal dynamic modeling unit recalculates the comprehensive energy function based on real-time point cloud data. E Iteratively update the optimal control points and correct the spinal model; The attitude prediction unit corrects the state transition matrix parameters based on the actual attitude deviation to optimize prediction accuracy; The pose fusion unit fine-tunes the feature confidence weights based on feedback deviations, thereby improving the stability of the input data.

[0075] After 10 minutes of continuous testing, this embodiment achieved an average trajectory tracking error as low as 3.2mm under normal user breathing and torso micro-movement scenarios, with contact force fluctuations controlled within a specified range. The system response latency is less than 60ms, which significantly improves the continuity and comfort of the massage.

[0076] Furthermore, to verify the effectiveness of the method described in this invention, a comparative experimental environment was constructed in this embodiment. An adult male with a height of 175cm and a weight of 70kg was selected as the test subject, and back massage tests were conducted using different control methods under the following two typical dynamic scenarios: Typical scenario 1: Normal breathing, frequency 0.2-0.3Hz, fluctuation range 3-5mm; Typical scenario 2: Slight trunk movement, simulating adjusting sleeping posture, displacement 5-10mm; Control Method 1: This method employs a traditional approach of static scanning followed by a single force response. A static 3D scan is performed before the massage begins to generate a fixed trajectory. During the massage, the force is detected solely by a force sensor, and simple PID adjustments are made when the force exceeds a threshold. This method lacks dynamic spinal modeling, energy constraints, and posture prediction mechanisms.

[0077] Control Method Two: Employing the full-process control method of this invention, enabling multimodal adaptive fusion with integrated energy function. E Constrained spinal dynamic modeling, extended Kalman filter attitude prediction and full-link closed-loop control; After running continuously for 10 minutes in the two typical dynamic scenarios described above, the core evaluation index data collected by the present invention and the comparative example are shown in Table 1: Table 1: Comparison of Core Evaluation Indicators During the 10-minute run in the aforementioned dynamic scenario, from Figure 3 It can be seen that existing technologies, due to their use of static fixed trajectories, cannot detect the 4mm Z-axis fluctuation caused by human respiration in advance. The robot can only passively adjust after abnormal contact force, leading to the massage head becoming suspended or excessively compressed, with an average error as high as 7.6mm. In contrast, this invention, by constructing a posture state transition model, predicts the displacement of the human back 50ms in advance, while also relying on a comprehensive energy function... E The constrained dynamic spine model ensures that the predicted trajectory conforms to the real-time spinal morphology of the human body. The predicted value is used as a feedforward signal input to the controller to achieve synchronization, which greatly reduces the tracking error. This reveals the mechanism by which the present invention improves the accuracy of trajectory tracking.

[0078] During the 10-minute run in the aforementioned dynamic scenario, such as Figure 4 As shown, through contact force stability and comfort analysis, it can be seen that existing technologies fail to recognize changes in the geometric characteristics of the spine when passing over bony prominences, maintaining constant stiffness and causing the contact force to surge instantaneously to 24N, resulting in significant pain for the user. In contrast, this invention, through a comprehensive energy function... E It ensures that the spinal curvature conforms to physiological constraints from the modeling stage, and adjusts the stiffness coefficient in real time based on the predicted curvature to achieve adaptive tactile feedback; at the same time, it uses comprehensive energy... E As a safety verification indicator, it avoids force output exceeding physiological limits and controls force fluctuations within a certain range. Within a safe and comfortable range.

[0079] Furthermore, such as Figure 5As shown, in the test, when the subject made a slight movement of the torso, the existing technology, due to its lengthy data processing link and lack of parallel optimization mechanism, resulted in a significant increase in system response latency (150ms-225ms) and violent fluctuations, making it difficult to meet the real-time requirements. However, the present invention significantly reduced the processing time through an efficient data fusion architecture and parallel computing strategy, keeping the system response latency stably controlled at a low level (about 60ms). Combined with a closed-loop feedback mechanism, it ensured that the system could maintain extremely low lag and high real-time performance under dynamic interference.

[0080] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0081] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. An adaptive control method for dynamic posture tracking of a back massage robot, characterized in that, include: S1. During the massage, multimodal posture perception data of the human back is obtained through the human posture perception unit. S2. Perform information fusion on multimodal posture perception data to obtain a unified human posture state; S3. Construct a dynamic parameterized model related to the centerline of the human spine based on human posture. S4. Based on the dynamic parameterization model and the human posture state vector X at the current and historical moments ( t ) Optimal control point parameters Establish a short-time dynamic prediction model for human posture; S5. Adjust the motion and force control of the massage actuator based on the predicted posture results of the short-time dynamic prediction model, and feed back the massage execution results to S2, S3, and S4 to continuously update the human body posture state, dynamic parameterization model, and short-time dynamic prediction model, so as to form a complete closed-loop adaptive control mechanism.

2. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 1, characterized in that, In S1, the multimodal pose perception data includes: three-dimensional point cloud data of the human back. D ( t Human torso posture perception data q ( t Massage contact force sensing data F ( t ) 。 3. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 1, characterized in that, In S2, the information fusion processing flow includes: S20. Multimodal feature extraction is achieved by preprocessing the multimodal pose perception data; S21. Define various features at time using the following formula. t Fusion confidence: In the above formula, , For the first k The preset weight coefficients of class features, and satisfying and , For the first k Class features at time t The reliability indicators, and exp(∙) is an exponential function. Indicates the first k Smoothing adjustment coefficient of class features For the first k The amount of change in class features between adjacent time points; S22. Based on the fusion confidence levels corresponding to various features, construct the following human pose state vector X( t ): In the above formula, p ( t ) indicates and D ( t The relevant three-dimensional position vector, To p ( t The rate of change of the position of key areas on the back was obtained by performing a time difference operation. θ ( t )and q ( t The relevant attitude direction vector, f ( t ) for and F ( t Related force perception state quantities, To and p ( t ), , θ ( t ), f ( t The related vector set.

4. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 1, characterized in that, In S3, the original dynamic parameterized model of the spinal centerline. C ( s,t The following formula represents the continuous parametric curve of the human spine over time: In the above formula, C ( s , t ) is through time t Normalized arc length parameter s Three-dimensional spatial parameter curves are used to characterize the dynamic spatial morphology of the spine. B i ( s To ensure the continuity and smoothness of the spinal curve, the third-order B-spline basis function is used. M for B The number of control points for a spline curve u i ( t (time) t The lower B-spline curve i The three-dimensional coordinates of each control point, and u i ( t The control points are dynamically adjusted using the following formula to ensure that the spinal model continuously follows changes in human posture: In the above formula, For the first i The initial three-dimensional coordinates of the control points The time step between two consecutive data acquisitions. This represents the attitude change in the attitude direction fused between adjacent time steps. for The corresponding incremental rotation matrix, l i In the local coordinate system of the human body, the first i Each control point has a preset fixed offset vector relative to the spinal reference point.

5. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 4, characterized in that, In S3, the dynamic parameterization model takes minimizing the comprehensive energy function E as the optimization objective and is solved using the gradient descent method. E Obtain the optimal control point parameters Substitution C ( s , t Then, the constrained optimization of the dynamic parameterized model is achieved through the following formula: In the above formula, C ( s , t ) * For the constrained optimized dynamic parameterized prediction model of the spinal centerline, the comprehensive energy function E is characterized by the following equation: In the above formula, E data To ensure that the model accurately reflects the current morphology of the human spine, and , N This represents the total number of point cloud sampling points. p i ( t () indicates time t Lower back surface i The three-dimensional spatial coordinates of each sampling point s i For the first i The specific values ​​of the normalized arc length parameter corresponding to the spinal centerline for each back point cloud data point are taken. ; E smooth To ensure the smoothness constraint term maintains the continuous and smooth morphological characteristics of the spinal curve, and , For the spinal curve C ( s , t For the normalized arc length parameter s The second derivative, E bio To avoid generating physiological consistency constraints on spinal morphology that exceed the physiological limits of the human body, and , For a moment t Below is a dynamic parametric model of the original spinal centerline. C ( s , t In the normalized arc length parameter s Local curvature at that point This is the maximum curvature threshold allowed by the physiological structure of the human spine. λ 1. λ 2 are the smoothing constraint terms. E smooth Physiological consistency constraint E bio The corresponding weighting coefficients.

6. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 1, characterized in that, In S4, the short-time dynamic prediction model is characterized by the following equation: In the above formula, C ( s , t +1) * The mapping function is a dynamic parameterized prediction model of the spinal centerline at time t+1. f (∙) represents the state transition mapping based on the first-order Markov property, which is then used to predict the attitude state through extended Kalman filtering. With the current optimal control point This is mapped to a dynamic parameterized prediction model of the spinal centerline at the next moment. for t The predicted human posture state at time +1, and during the prediction process, if the comprehensive energy function corresponding to the candidate spine model of the predicted posture... E If the value exceeds the preset threshold, the predicted state will be... Make corrections to ensure The prediction results are consistent E Minimize the constraints.

7. The adaptive control method for dynamic posture tracking of the back massage robot as described in claim 1, characterized in that, In S5, the local curvature of the spine is calculated using the following formula. Construct a variable stiffness impedance control model that adaptively adjusts with curvature. F d At the same time, combined with the comprehensive energy function E The safety threshold is determined in advance, and the massage trajectory adjustment amount and contact force control parameters are corrected in advance, thereby adjusting the motion and force control of the massage actuator: In the above formula, , These are the dynamic parameterized prediction models for the spinal centerline. For normalized arc length parameter s The first and second derivatives, Satisfying S3 E bio Physiological threshold of constraint ; Variable stiffness impedance control model F d It is characterized by the following formula: In the above formula, H Here is the damping coefficient matrix. To follow the local curvature of the spine An adaptively adjusted stiffness coefficient matrix, where the stiffness coefficients vary with... The increase is exponentially decayed when , And the comprehensive energy function E Reduce stiffness coefficient when exceeding preset safety threshold , x d Let S4 be the human pose position vector predicted by S4, and , x The actual position vector of the massage head. , These are the first derivatives of the desired position and the actual position, respectively.

8. An adaptive control system for dynamic posture tracking of a back massage robot, applied in the adaptive control method for dynamic posture tracking of a back massage robot as described in any one of claims 1 to 7, characterized in that, include: Human posture sensing unit, which is used to acquire multimodal posture sensing data of the human back during the massage process; The posture information fusion unit is used to fuse the multimodal posture data output by the human posture perception unit. The spinal dynamic modeling unit is used to construct a dynamic parametric model of the human spinal centerline based on the fused human posture state. The posture prediction and control unit is used to predict the trend of human posture changes and generate control commands based on the prediction results. A massage actuator that executes control commands generated by the posture prediction and control unit; The closed-loop feedback and safety control unit is used to feed back the execution results of the massage actuator to the posture information fusion unit, the spinal dynamic modeling unit, and the posture prediction and control unit, continuously updating the human posture state, dynamic parameterization model, and short-time dynamic prediction model to build a complete closed-loop adaptive control mechanism; at the same time, it sets multi-dimensional safety constraints and executes safety protection strategies when an anomaly is triggered. The human posture sensing unit includes: A visual perception submodule for acquiring three-dimensional morphological information of the human back; An inertial posture perception submodule used to collect information on changes in the posture of the human torso during massage. A force sensing submodule is installed on the massage actuator to obtain force feedback information during the contact between the massage head and the back of the human body.

Images