Sitting rehabilitation training action recognition and control system and method based on artificial intelligence

Abstract

The invention relates to the technical field of physiotherapy devices, and particularly discloses a seat rehabilitation training action recognition and control system and method based on artificial intelligence, and the system is provided with an action segment detection module, an artificial intelligence action recognition module and a computer training correction module. The action segment detection module automatically identifies and intercepts a training action segment set containing effective motion information; the artificial intelligence action recognition module outputs structured information such as action types, action stages and action amplitudes, and estimates an active participation degree coefficient of the user in combination with the same multi-source data set; the computer training correction module is used for generating control target parameters of the sitting rehabilitation training equipment by a controller based on the structured information and the active participation degree coefficient, and carrying out online correction on equipment output in the training process, so that the sitting rehabilitation training has the capability of adaptively adjusting assistance and rhythm while executing a correction action; therefore, action recognition and fine control of the on-machine training process are completed.

Description

Technical Field

[0001] This invention relates to the field of physiotherapy device technology, specifically to an artificial intelligence-based sitting rehabilitation training movement recognition and control system and method. Background Technology

[0002] The core of existing seated rehabilitation training processes involves using multi-source sensors, including inertial measurement units, pressure sensing pads, electromyography sensors, or visual cameras, installed on key parts of the patient's body (such as shoulders, waist, upper limbs, and seat cushions), to collect real-time data on posture, acceleration, center of pressure, and muscle activity. The control unit first filters, registers, and normalizes the raw signals to establish a baseline model of the patient's current sitting posture. Then, motion recognition algorithms (such as convolutional neural networks, long short-term memory networks, or spatiotemporal graph convolutional models) are used to identify and segment the training movements, determining the movement type, amplitude, and completion level. Based on the recognition results and preset training goals, the system provides feedback and control adjustments to the patient's movement status, including providing visual or verbal cues or using electric / pneumatic assist mechanisms for correction. Finally, the core data related to movement execution, balance status, and system control during the training process are stored in a database for subsequent progress assessment and optimization of individualized rehabilitation strategies.

[0003] For example, Chinese invention patent CN107019617B discloses a seated balance rehabilitation training robot and its control method. It includes a seat structure, a four-bar linkage structure, a main shaft transmission structure, a main shaft tilt angle sensor, and a crank angle sensor. The seat structure includes a support armrest, a headrest, a backrest, and a support cushion. The four-bar linkage structure includes a rocker arm, a reducer bracket, a crank, a sliding connecting rod, a stepper reducer, and a stepper motor. The main shaft transmission structure mainly includes a transmission main shaft, a main shaft support, a retaining ring, a torsion spring, and a main shaft positioning block. The main shaft tilt angle sensor records and measures the patient's tilt angle, providing real-time feedback to control the motor speed and direction. The backrest and armrests, along with other auxiliary systems, push the patient to straighten them. The angle sensor provides real-time feedback to control the motor to reverse and reset. The rocker arm disengages from the sliding connecting rod, returning the motor crank to its zero position.

[0004] For example, Chinese invention patent CN112370306B discloses a sitting-assisted rehabilitation training device for hemiplegic patients. The device, in the technical field of rehabilitation training devices for hemiplegic patients, includes: a base; a back support set on the upper side of the base; a standing-sitting assist mechanism, height-adjustable on the back support, including a healthy arm support component and an affected shoulder pressing component; a trunk splint movably set on both sides of the back support; an arm movement mechanism set on one side of the back support, capable of moving the patient's affected arm; and a leg movement mechanism slidably set on the base. It can help the patient sit down and shift their center of gravity towards the affected side while seated, and can also provide activity training for the patient's trunk and limbs.

[0005] Existing technologies have relatively limited sensing dimensions, primarily focusing on posture-related information acquisition (such as acquiring tilt angles via gyroscopes), and their control logic largely relies on rigid, fixed threshold rules to determine the patient's sitting tilt angle and drive actuators such as motors or cylinders for correction. However, in the actual application of sitting rehabilitation training, the patient's movements are not merely simple postural changes; they often involve alternation between active participation and passive assistance, and require multi-point force coupling and coordinated control of the trunk and limbs.

[0006] In such complex scenarios, existing technologies typically rely on relatively singular perception methods at the recognition level, which may fail to comprehensively capture changes in the patient's center of gravity, muscle responses, and differences in intent. Therefore, in some cases, it may be difficult to accurately determine whether the patient is active or passive. Such limitations at the recognition level can lead to deficiencies at the control level, specifically manifesting as uneven torque distribution during movement correction, misjudgment of execution timing, and even problems such as oscillation or overvoltage when multiple actuators are synchronized, thereby affecting the effectiveness and safety of training. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based sitting rehabilitation training movement recognition and control system and method, which can effectively solve the problems mentioned in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides an artificial intelligence-based sitting rehabilitation training movement recognition and control system, comprising: a movement segment detection module, used for users to perform on-the-machine training on a sitting rehabilitation training device, extracting multi-source datasets of rehabilitation training in real time according to an adapted sampling frequency to detect movement segments, thereby obtaining a set of training movement segments; an artificial intelligence movement recognition module, used to input the set of training movement segments into the artificial intelligence, outputting structured information of each training movement segment, and simultaneously outputting the user's active participation coefficient based on the multi-source dataset of rehabilitation training; and an on-the-machine training correction module, used for the controller to generate control targets for the sitting rehabilitation training device based on the structured information of each training movement segment and the user's active participation coefficient, correcting the user's on-the-machine training process, and completing the recognition and control of sitting rehabilitation training.

[0009] The second aspect of this invention provides an artificial intelligence-based method for recognizing and controlling seated rehabilitation training movements, comprising: a user performing on-the-machine training on a seated rehabilitation training device; extracting multi-source datasets of rehabilitation training in real time according to an adapted sampling frequency to detect movement segments, thereby obtaining a set of training movement segments; inputting the set of training movement segments into the artificial intelligence, outputting structured information of each training movement segment; simultaneously, based on the multi-source datasets of rehabilitation training, the artificial intelligence outputting a coefficient of the user's active participation; and based on the structured information of each training movement segment and the user's active participation coefficient, a controller generating a control target for the seated rehabilitation training device, correcting the user's on-the-machine training process, and completing the recognition and control of seated rehabilitation training.

[0010] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) This invention provides an artificial intelligence-based sitting rehabilitation training movement recognition and control system and method. The movement segment detection module is used to extract multi-source datasets of rehabilitation training in real time according to the appropriate sampling frequency when the user is training on the sitting rehabilitation training device. It segments continuous signals such as posture, center of gravity and airbag pressure, automatically identifies and extracts training movement segment sets containing effective motion information, thereby reducing the interference of static and noise data on subsequent processing while retaining key movement details. The artificial intelligence movement recognition module inputs the training movement segment set into a pre-trained artificial intelligence model and outputs "movement type, movement stage, movement amplitude". The system uses structured information such as "degree" and estimates the user's active participation coefficient by combining the same multi-source dataset, transforming the raw sensor data into high-level features that can directly drive control decisions. The on-machine training correction module, based on the structured information of each training action segment and the user's active participation coefficient, generates control target parameters for the seated rehabilitation training device (including the target pressure of the three-way airbag and the target displacement of the clamping component) by the controller. During the training process, it corrects the device output online, enabling the seated rehabilitation training to adaptively adjust assistance and rhythm while performing corrective actions, thereby completing the action recognition and fine control of the on-machine training process.

[0011] (2) This scheme does not mechanically apply corrective force based solely on the magnitude of postural deviation or center of gravity shift. Instead, it uses the active participation coefficient output by the artificial intelligence action recognition module to characterize the proportion of the user's contribution to the current training action through autonomous force exertion. Based on this, the on-machine training correction module dynamically scales the additional pressure and displacement assistance applied by the device, so that users with high active participation have more opportunities to practice independently, while users with low active participation receive more safety support. This enhances the activation and reconstruction of neuromuscular pathways and improves long-term rehabilitation effects while meeting safety requirements.

[0012] (3) This solution reuses the same set of rehabilitation training multi-source datasets in the three stages of motion segment detection, artificial intelligence motion recognition and on-machine training correction. Physical quantities such as IMU posture, seat pressure distribution, airbag pressure and clamp displacement are used not only for motion segment start and end determination, but also for motion semantic recognition and active participation inference. Finally, it directly serves the generation and closed-loop correction of control targets, avoiding the waste of resources by repeatedly arranging sensors or repeatedly collecting data for different functional modules. Through the integrated reuse of multiple variables in the detection, recognition and control links, the overall consistency and robustness of the system are improved, and the hardware cost and data processing overhead are reduced, which facilitates the engineering implementation of the equipment.

[0013] (4) Traditional seated rehabilitation training devices often rely on a single sensor value (e.g., based solely on trunk tilt angle or single-point pressure) to trigger fixed control logic, making it difficult to simultaneously consider the accuracy of motion recognition, the degree of active user participation, and the fine adjustment of the corrective torque distribution. In contrast, this solution uses a motion segment detection module to structure the continuous training process into motion segments, and an artificial intelligence motion recognition module to output structured motion information and active participation coefficients. Then, the on-machine training correction module generates individualized control targets for the three-dimensional airbag and clamping mechanism, forming an integrated closed loop of detection-recognition-correction. Compared with existing processes based solely on a single threshold or single-channel feedback, this solution can achieve higher resolution motion recognition, finer mechanical control, and more adaptive rehabilitation strategies that better suit the patient's condition while ensuring training safety. Attached Figure Description

[0014] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the system module connections of the present invention.

[0016] Figure 2 This is a schematic diagram of the method steps of the present invention.

[0017] Figure 3 This is a schematic diagram of the static center of gravity offset of the seat plate sensor array.

[0018] Figure 4 This is a schematic diagram representing the angular velocity curve of motion.

[0019] Figure 5 This is a flowchart for action segment detection. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] It should be explained that the implementation process of this invention is based on a seated rehabilitation training device. This device mainly consists of a hollow square base, a seat board inside the base, and multiple sets of adjustable clamping devices distributed on the front and left and right sides of the human body. These include a first support rod / first sliding rod / first locking element / first adjusting rod and its end, and a second clamping element; and a second support rod / second sliding rod / second locking element / second adjusting rod and its end. Some embodiments also include a third clamping element composed of multi-stage connecting rods and an airbag assembly, forming a three-sided clamping structure through multi-directional force application. The core working principle of the device is: after the user enters the base in a seated position, the clamping elements are adjusted vertically, horizontally, and backwards to conform to the curved parts of the spine, applying directional corrective forces from the left and right sides and the front of the spine, thereby compressing and correcting the curvature deformity of the spine. Simultaneously, the sliding and locking rods achieve precise positioning to adapt to different user body shapes and lesion sites, realizing multi-dimensional corrective training under seated conditions.

[0023] Reference Figure 1 As shown, the first aspect of this invention provides an artificial intelligence-based seated rehabilitation training movement recognition and control system, including: a movement segment detection module, an artificial intelligence movement recognition module, an on-machine training correction module, and a rehabilitation training management library. The rehabilitation training management library is used to store preset values ​​for various parameters.

[0024] The preset relationships stored in the aforementioned rehabilitation training management library include, but are not limited to, preset, matching, and mapping relationships. Specifically, in this embodiment, the method for obtaining these preset relationships takes the extraction of the rated computing power and sampling frequency correction parameters of the edge processor belonging to the sitting rehabilitation training device as an example. On one hand, theoretical modeling and simulation analysis can be performed on the mechanical structure, actuator, sensing unit, and edge processor of the sitting rehabilitation training device to derive the functional relationship or piecewise mapping relationship between parameters such as control bandwidth, sampling frequency, valve opening, and target pressure. On the other hand, clinical rehabilitation experience and a large amount of historical training data can be combined to divide and optimize the key parameter range using offline statistical analysis or machine learning methods, forming a parameter comparison table adapted to different patient types, training modes, and equipment configurations. Based on this, the mapping relationship is corrected and converged through on-machine testing and calibration, ultimately solidifying it into preset relationship entries in the rehabilitation training management library. In this embodiment, the rated computing power and sampling frequency correction parameter of the edge processor to which the seated rehabilitation training device belongs are extracted as an example. The mapping result between the rated computing power and the sampling frequency correction parameter is written into the management library. In actual application, the same mechanism can also be used to construct various preset relationships such as control bandwidth-sampling frequency adjustment factor and movement amplitude-target pressure difference.

[0025] The motion segment detection module is connected to the artificial intelligence motion recognition module, which in turn is connected to the on-machine training correction module. All three modules—motion segment detection, artificial intelligence motion recognition, and on-machine training correction—are connected to the rehabilitation training management database.

[0026] The motion segment detection module is used by users to perform on-the-job training on a seated rehabilitation training device. It extracts multi-source datasets of rehabilitation training in real time according to the appropriate sampling frequency to detect motion segments and obtain a set of training motion segments.

[0027] The training system constructs a multi-source dataset including posture, force, and mechanism position. IMU (Inertial Measurement Unit) modules are deployed in the chest, back, and pelvic regions to collect real-time Euler angles or quaternion posture information and angular velocity for the corresponding areas. A seat pressure sensor array, specifically pressure sensor units, outputs the user's pressure distribution, center of gravity X / Y coordinates, and overall force level while seated. The first, second, and third airbags inside each clamping component transmit real-time airbag pressure via pressure gauges and A / D conversion. Simultaneously, displacement encoders on each adjustment rod and support rod continuously provide linear or angular displacement readings, specifically displacement sensor units, reflecting the current physical position and motion state of the clamping components, thus forming multi-dimensional input features describing the user's state and the device's actions.

[0028] It should be explained that the aforementioned sitting rehabilitation training equipment is generally composed of two main layers: the first is the functional unit for data monitoring and decision-making, such as the inertial measurement unit and pressure sensing unit mentioned above; the second is the basic execution layer for equipment operation, including mechanical and pneumatic devices such as the seat board, support rods, and clamping components. Under the control of the rehabilitation training management unit, these devices complete specific support, clamping, and correction actions. The data monitoring unit provides real-time quantitative representation of the user's state and the device's state, and the execution device implements physical intervention based on the monitoring and analysis results, thereby forming a closed-loop sitting rehabilitation training system of monitoring-decision-execution.

[0029] Specifically, user-based on-the-job training using seated rehabilitation training equipment also includes: The seated rehabilitation training equipment includes an inertial measurement unit, a pressure sensing unit, a displacement sensing unit, and a rehabilitation training management unit.

[0030] The rehabilitation training management unit is pre-configured with a set of predetermined training movements and the current training sequence, specifically the rehabilitation training management library.

[0031] The user's sitting data is extracted using an inertial measurement unit and a pressure sensing unit. The user's sitting data includes the user's static torso angle and the user's static maximum pressure.

[0032] It's important to clarify that the user's static trunk angle actually corresponds to the attitude parameters output by the IMU, specifically the Euler angles calculated by the chest / back / pelvic IMU (or Euler angles converted from quaternions), not angular velocity. Specifically, the pitch and / or roll angles, which reflect the trunk's deflection relative to the direction of gravity, are generally chosen as indicators of the static trunk angle. The static trunk angle is the angle of deflection of the trunk's longitudinal axis relative to the device's coordinate system or the vertical direction of gravity when the user is sitting stably and without active movement. It can be defined as: the flexion / extension angle (pitch angle) in the sagittal plane, or the left / right tilt angle (roll angle) in the coronal plane. Alternatively, it can be combined into a comprehensive deflection angle to characterize how many degrees the user's upper body has deviated from an upright sitting posture in a static baseline state.

[0033] The user's static torso angle and the user's static maximum pressure are respectively denoted as the user's static torso reference angle and the user's static reference center of gravity. The static maximum pressure detected by the seat pressure sensor array is denoted as the static reference center of gravity.

[0034] Furthermore, the rehabilitation training multi-source dataset is extracted in real time according to the appropriate sampling frequency for action segment detection. The specific process is as follows: Multi-source rehabilitation training data is collected based on an adaptive sampling frequency and arranged sequentially according to the collection time sequence to form a multi-source rehabilitation training dataset. The multi-source rehabilitation training dataset is then divided according to a set sliding time window to obtain the multi-source rehabilitation training data of the user under each sliding time window.

[0035] The multi-source data for rehabilitation training can include, but is not limited to, the following categories: postural data, such as Euler angles / quaternions, angular velocity, and linear acceleration output from IMUs deployed on the chest, back, pelvis, and limbs; force and contact data, such as pressure distribution on the surface of the seat, backrest, and clamping components, X / Y coordinates of the center of gravity, total force, and real-time pressure of the airbags inside each clamping component; position and kinematic data, such as displacement, velocity, and locking position of each adjusting rod, support rod, slide rail, and electric actuator; physiological and muscle strength data, such as the degree of muscle activation reflected by surface electromyography (EMG), and safety monitoring indicators such as heart rate and blood oxygen saturation when necessary; and environmental and equipment status data, such as equipment operating mode, valve opening, control output commands, and abnormal alarm signs, thereby forming a three-dimensional depiction of the user's status and equipment actions on the same time axis.

[0036] The specific process for setting the sliding time window is as follows: Based on the appropriate sampling frequency, a time window correction factor is obtained by matching with the rehabilitation training management library. The baseline sliding time window corresponding to the current predetermined training action is extracted and multiplied by the time window correction factor. The baseline sliding time window can be extracted from the configuration record of the predetermined training action set to obtain the set sliding time window.

[0037] Based on multi-source data of user rehabilitation training under each sliding time window, the real-time posture changes and real-time center of gravity shifts of users under each sliding time window are calculated.

[0038] The real-time attitude changes of the user under each sliding time window are obtained by calculating the L2 norm of the user's real-time angular velocity under each sliding time window. The real-time center of gravity movement of the user under each sliding time window is obtained by calculating the difference in center of gravity movement at each adjacent time point.

[0039] The real-time posture change and the user's real-time center of gravity movement are both dynamic change values ​​calculated with reference to the aforementioned baseline state. This yields the offset of the user's current posture relative to the baseline posture and the movement of the center of gravity relative to the baseline center of gravity, which are used to characterize the user's real-time body tilt direction, amplitude, and sitting posture stability.

[0040] The real-time posture changes and center of gravity shifts of the user under each sliding time window are compared with the predefined posture change thresholds and center of gravity shift thresholds in the rehabilitation training management library under this training sequence to detect motion segments. Based on the detection results, it is determined whether to mark the sliding time window as a training motion segment, thus obtaining a set of training motion segments.

[0041] The specific process for detecting the above-mentioned action segments is as follows: The real-time attitude change of the user under each sliding time window is compared with the attitude change threshold, and the real-time center of gravity movement of the user under each sliding time window is compared with the center of gravity movement threshold. When the user's real-time attitude change is greater than the attitude change threshold and the user's real-time center of gravity movement is greater than the center of gravity movement threshold, the duration of the greater state is recorded as the candidate motion duration.

[0042] The candidate motion duration is compared with a first threshold. When the candidate motion duration exceeds the first threshold, the time period is marked as a training motion. When the user's real-time posture change falls below the threshold and remains below the second threshold of the candidate motion duration, the training motion is marked as ended, thus obtaining a training motion segment. The training motion segment set is obtained by traversing each sliding time window.

[0043] The specific motion segment detection process is as follows: Figure 5 As shown, Figure 5The flowchart for motion segment detection illustrates how the training system divides the entire training process into many small sliding time windows. Each window calculates two things: the magnitude of the posture change and the magnitude of the center of gravity shift within that segment. These two quantities are then compared to pre-set thresholds for posture change and center of gravity shift. Only when both the posture change and center of gravity shift exceed the threshold within a given time window is the system considered to have performed a real movement, and the duration of this exceedance is accumulated and recorded as a candidate motion duration. If this candidate motion duration exceeds the first duration threshold (e.g., at least 0.5 seconds of continuous movement), the system recognizes the movement as officially started and marks this segment as a training motion. If the detected posture change then falls back below the threshold and remains stationary for more than the second duration threshold (e.g., 0.3 seconds or more without meeting the threshold), the system determines that the training motion has ended. The entire segment from the start to the end of the movement constitutes a training motion segment. The system scans the entire timeline from beginning to end in this way, and then obtains a set of training action segments, which is the collection of all effective action segments in this training.

[0044] For example, with a sampling period of 0.1s, the threshold for posture change is set to 5°, the threshold for center of gravity shift is set to 1cm, and the first threshold for candidate motion duration is 0.5s and the second threshold is 0.3s. If a user completes a significant side bend within a time period of 0.3 to 0.9s, and the posture change is greater than 5° and the center of gravity shift is greater than 1cm in multiple consecutive time windows, the system starts accumulating candidate motion durations when both thresholds are met simultaneously for the first time. When the accumulated time exceeds 0.5s, the interval from approximately 0.3s is marked as the start of the training action. Subsequently, the user returns to a neutral position. After 1.3s, the posture change and center of gravity shift remain below the thresholds in multiple time windows. When the accumulated stillness time exceeds 0.3s, the system determines that the training action ends around 1.3s, thus obtaining a training action segment of approximately 0.3 to 1.3s.

[0045] The system does not directly determine actions from single-frame data. Instead, it uses a sliding time window to dynamically observe posture changes and center of gravity shifts. When the change continuously exceeds an amplitude threshold and remains above a preset first threshold for candidate motion duration, the system determines that the user has entered the initial stage of the training action. When the change falls below the threshold and its duration exceeds a second threshold for candidate motion duration, the system determines that the training action has ended. The system automatically records the start and end times t_start and t_end and extracts the multi-source feature sequences within this interval as input for subsequent action recognition models.

[0046] Specifically, the process for adapting the sampling frequency is as follows: The real-time control bandwidth of the sitting rehabilitation training device is extracted. The real-time control bandwidth can be extracted from the execution record of the sitting rehabilitation training device. It is matched with the sampling frequency adjustment factor corresponding to each predefined real-time control bandwidth interval to determine the interval to which the real-time control bandwidth of the sitting rehabilitation training device belongs, and the sampling frequency adjustment factor corresponding to the interval is obtained.

[0047] In this implementation, the real-time control bandwidth of the device under typical rehabilitation conditions is extracted to characterize the highest effective response frequency that the system needs to reliably track under closed-loop control. This real-time control bandwidth is then categorized and matched with multiple predefined real-time control bandwidth intervals, and a corresponding sampling frequency adjustment factor is pre-set for each bandwidth interval. This ensures that the sampling frequency is always a multiple of the control bandwidth, thus satisfying the Nyquist sampling theorem and a certain oversampling margin. Based on this, the higher the control bandwidth, the faster the attitude changes the system needs to track, and the higher the requirement for real-time control. Therefore, a larger sampling frequency adjustment factor is matched to ensure a sufficiently fine sampling period, enabling the closed-loop control to promptly perceive and respond to changes in user attitude and center of gravity. Conversely, when the control bandwidth is low and the training process mainly involves slow attitude adjustments and steady-state maintenance, a smaller sampling frequency adjustment factor can be matched to reduce computational and communication burdens. Through the above mapping relationship between bandwidth and sampling frequency, the adaptive setting of the sampling frequency under different control bandwidth conditions is ensured to have a clear physical basis and control theory support.

[0048] Extract the baseline sampling frequency corresponding to the current predetermined training action. The baseline sampling frequency can be extracted from the configuration record of the predetermined training action set. Multiply it by the sampling frequency adjustment factor to obtain and record it as the theoretical lower limit of the sampling frequency.

[0049] Extract the rated computing power of the edge processor to which the seated rehabilitation training equipment belongs. The rated computing power can be extracted from the usage log of the edge processor. The sampling frequency correction parameter is obtained by association with the rehabilitation training management library and multiplied by the theoretical lower limit of the sampling frequency to obtain and record the adaptive sampling frequency.

[0050] In this embodiment, the rated computing power of the edge processor to which the seated rehabilitation training device belongs is extracted in order to constrain the total amount of computation that can be completed in a given unit of time for sensor data preprocessing, motion recognition inference and control command issuance, thereby providing a hardware boundary for the upper limit of the sampling frequency: when the sampling frequency is increased, the number of data frames that need to be processed in a unit of time increases proportionally, and each frame of data is often accompanied by fixed computational overhead such as filtering, feature extraction, AI inference and control quantity calculation. If the sampling frequency is set too high and exceeds the computing power per second that the edge processor's rated computing power can bear, it will lead to the accumulation of single frame processing delay, control loop lag or even data packet loss, which will damage real-time performance and stability. Therefore, after determining the theoretical lower limit of the sampling frequency based on the control bandwidth, a sampling frequency correction parameter is obtained by further combining the rated computing power of the edge processor. This parameter is used to scale the theoretical lower limit upwards or downwards: when the rated computing power margin is large, the correction parameter can be slightly greater than 1, so that the adaptive sampling frequency is appropriately increased to enrich the timing information while meeting the control requirements; when the rated computing power is close to saturation or some computing power needs to be reserved for other tasks, the correction parameter is less than 1, so that the theoretical lower limit is appropriately lowered to the adaptive sampling frequency. This achieves a balance between control bandwidth and computing resources, ensuring that the entire identification and control link can still operate stably in real time under the computing power constraints of the edge processor.

[0051] The AI ​​action recognition module is used to input the training action fragment set into the AI ​​and output the structured information of each training action fragment. At the same time, based on the multi-source dataset of rehabilitation training, the AI ​​outputs the user's active participation coefficient.

[0052] Furthermore, the structured information of each training action segment is output, and the specific analysis process is as follows: Extract the user's maximum torso tilt angle and static deviation from center of gravity during the current training action segment. The maximum torso tilt angle is obtained by monitoring the inertial measurement unit, and the static deviation from center of gravity is obtained by monitoring the pressure through the pressure sensing unit.

[0053] The maximum tilt angle of the user's torso is obtained by subtracting the user's maximum torso tilt angle from the user's static torso reference angle. The user's static deviation from the center of gravity is compared with the user's static reference center of gravity to obtain the user's static center of gravity offset.

[0054] In this embodiment, the above deviation comparison first obtains the user's static reference centroid coordinates (X). ref Y ref Specifically, in the geometric coordinate system of the seat plate sensor array, the row and column indices of the static reference center of gravity in the sensor array are converted into specific center of gravity coordinate values, and the center of gravity coordinates (X, Y) corresponding to the user's static deviation state are obtained under the current set posture of the device. off Y offThen, the static offset center of gravity is compared with the static reference center of gravity in the same coordinate system. The user's static center of gravity offset is characterized by calculating the straight-line distance between the two on the X / Y coordinate plane of the center of gravity. That is, the static center of gravity offset is determined by... This is determined to reflect the degree of deviation of the user's overall seated center of gravity from the ideal reference position in a static state using a dimensionally unified distance index, which is manifested as follows: Figure 3 As shown, Figure 3 This is a schematic diagram of the static center of gravity offset of the seat plate sensor array. The horizontal and vertical axes represent the X and Y coordinates of the seat plate sensor array, respectively, showing the position of the static reference center of gravity and the static offset center of gravity in the geometric coordinate system.

[0055] The structured information for each training exercise segment includes the type of exercise, the training order, the maximum tilt angle of the user's torso, the user's static center of gravity offset, and the confidence level of each exercise segment.

[0056] The confidence score of a training action segment is a quantitative indicator used to characterize the reliability of a detected and segmented action segment belonging to a certain target training action. Specifically, when an AI action recognition model makes a category determination for a segment (such as right lateral curvature correction, forward flexion stretching, etc.), it will simultaneously output a value between 0 and 1 or between 0 and 100% to indicate the reliability of the determination under the current multi-source data conditions: the higher the confidence score, the higher the matching degree of the segment's posture trajectory, center of gravity transfer path, velocity features, etc. with the corresponding action template, the better the data quality (such as completeness, noise level), and the more "confident" the model is of its classification result; the lower the confidence score, the more likely the segment may have non-standard movements, insufficient amplitude, large interference, or be ambiguous between multiple movements. In subsequent statistical analysis of training effects, generation of rehabilitation reports, or triggering adaptive control, this can be used for screening, downweighting, or review.

[0057] Specifically, artificial intelligence outputs a user's active participation coefficient, and the specific analysis process is as follows: Artificial intelligence analyzes multi-source datasets of rehabilitation training by calling a scoring model, thereby outputting the user's active participation coefficient. The scoring model calls active participation constraints to accumulate active participation scores to obtain the active participation coefficient.

[0058] In this embodiment, the artificial intelligence module calls a pre-trained scoring model to analyze the multi-source dataset of rehabilitation training (including posture trajectory, center of gravity shift, airbag pressure change, clamp displacement, and electromyography signals when necessary). Within each training action segment, it calculates the active participation score by time window. Under the premise of satisfying preset active participation constraints, it accumulates and normalizes the scores of each window, ultimately outputting the user active participation coefficient corresponding to that segment. The active participation score characterizes the contribution ratio of the user's own muscle strength or spontaneous control to the overall movement at a certain moment or within a short time window. For example, when the trunk movement direction is consistent with the direction of the device's force application, and the user's posture changes and center of gravity shifts precede or significantly exceed the portion explained by the device's actuation, the active participation score for that time window is high. Conversely, if the movement is mainly passively propelled by the device, and the user's own posture adjustment is delayed or minimal, the active participation score is low. By accumulating and averaging the active participation scores of each time window within the entire action segment, the active participation coefficient, reflecting the user's degree of autonomous participation during the overall training process, can be obtained.

[0059] Active participation in constraints, specifically: 1) The user's posture deviation time point is greater than the pressure deviation time point of the seated rehabilitation training device. The user's posture deviation time point is the time when the user's torso tilt angle exceeds the corresponding baseline value. The pressure deviation time point is the time when the pressure of each airbag in the device (the pressure gauge reading inside the first / second / third clamping component) changes. The above statement that the posture deviation time point is greater than the pressure deviation time point indicates a temporal sequence, meaning the user's posture deviation time point precedes the device's pressure deviation time point.

[0060] 2) The user's motion angular velocity curve conforms to the preset active motion template curve.

[0061] In this embodiment, the user's angular velocity curve is a time-series curve obtained by sequentially connecting angular velocity data collected in real time by inertial measurement units deployed on the chest, back, or pelvis along the time axis, as specifically as follows: Figure 4 As shown, Figure 4This diagram illustrates the representation of motion angular velocity curves, where the horizontal axis represents time and the vertical axis represents the angular velocity of the torso around the corresponding axis. It characterizes the complete velocity change process of a user during a seated training movement, from initiation, acceleration, reaching peak velocity, to deceleration and return to the starting position. The preset active movement template curve is the ideal active movement velocity trajectory obtained through a pre-defined process. During actual recognition, the motion angular velocity curve calculated from the user's current training movement segment is registered and its amplitude normalized on the time axis with the template curve for the corresponding movement type. By comparing the similarity of the two curves in features such as peak time, acceleration / deceleration slope, and duration, if the similarity is higher than a preset threshold, the user's movement process is determined to conform to the preset active movement template and can be used for subsequent active participation assessment and training quality evaluation.

[0062] 3) The user's static center of gravity deviation time point is greater than the time point of displacement of the clamping components of the seated rehabilitation training device. The static center of gravity deviation time point refers to the time point when the static center of gravity deviates from the corresponding reference value. The clamping component displacement time point refers to the time point when the clamping components begin to displace.

[0063] In this embodiment, the controller dynamically scales the airbag target pressure based on the user's active participation coefficient output by the recognition layer. When the active participation is high, indicating that the user can complete most of the corrective movements by their own strength, the system automatically reduces the additional pressure applied by the device, so that the clamping component only provides mild guidance and necessary safety support, thereby avoiding excessive device intervention. Conversely, when the active participation is low and the user's posture control ability is insufficient, the additional pressure is appropriately increased to compensate for their lack of muscle strength and stability. Through the above adaptive adjustment mechanism, the device maximizes the encouragement of the user's active participation in the training process while ensuring safety, thereby strengthening the recovery of neuromuscular control ability.

[0064] In a preferred embodiment of the present invention, the corrective pressure applied by the control module to the clamping component is always defined as the positive force along the direction in which the patient converges from the current offset posture to the target sitting posture. The magnitude of the positive force is adaptively adjusted according to the active participation coefficient output by the recognition layer: when the active participation is low, the system increases the positive auxiliary force to ensure the posture correction effect; when the active participation is high, the system correspondingly reduces the positive auxiliary force, so that the device mainly plays a role in mild guidance and safety protection, without forming a reverse resistance with the patient's voluntary force. Only in the advanced resistance training mode, and when the safety threshold condition is met, can the control module apply a limited resistance force in the opposite direction to the patient's voluntary force to achieve targeted strength training.

[0065] During training, the controller continuously monitors the user's status in real time based on IMU posture data, seat pressure distribution, and airbag pressure sensor signals. When it detects an abnormal change in the movement trajectory within a short period of time, a sharp shift in the center of gravity, and the user is not currently in the predetermined training action triggering phase or has not received a corresponding training instruction, the controller identifies this state as a suspected "unexpected imbalance" event and triggers safety protection procedures such as decompression, limit, or alarm. At the same time, the controller also compares the matching relationship between airbag pressure changes and posture changes. When the internal pressure of the airbag rises rapidly within a unit of time while the user's torso posture and center of gravity shift only change slightly or remain basically unchanged, the controller identifies this state as a possible "over-clamping" risk and promptly limits the subsequent pressure increase or performs a decompression operation to avoid discomfort or even secondary injury caused by abnormal states.

[0066] The on-machine training correction module is used to generate control targets for the seated rehabilitation training equipment based on the structured information of each training action segment and the user's active participation coefficient, correct the user's on-machine training process, and complete the identification and control of seated rehabilitation training.

[0067] Furthermore, the controller generates control objectives for the seated rehabilitation training equipment and corrects the user's on-machine training process. The specific analysis process is as follows: The seated rehabilitation training device includes various clamping components and various adjustment rods / support rods, and each clamping component contains an internal airbag.

[0068] The structured information of each training action segment and the user's active participation coefficient are input into the AI ​​controller. The controller outputs the control target pressure of each clamping component and the control target displacement of each clamping component along the corresponding adjustment rod through the preset control target mapping set and the set pressure limit scaling factor.

[0069] In this embodiment, the controller has a pre-set control target mapping set to guide the distribution relationship between airbag pressure and clamping component displacement. This mapping set can be organized as a rule table combined with interpolation calculation. For example, when the training mode is right-side bend correction—correction stage, the target relationship stored in the mapping set specifies the target pressure P of the right-side clamping component. R It should be higher than the target pressure P of the left clamping member. L And require the pressure P of the back clamping component. back Synchronous lifting is used to form an effective three-dimensional corrective torque; for different degrees of lateral bending deviation, the mapping set drives interpolation calculations through the pressure difference data corresponding to the amplitude range, so that the larger the lateral bending amplitude, the larger the target pressure difference ΔP=P is automatically generated. R - P LThis enables the controller to adaptively determine the combined force application mode of the three-way airbags based on the currently identified lesion characteristics and offset, thereby achieving graded correction control with force and direction proportional.

[0070] The controller first retrieves the basic target parameters corresponding to the action under standard operating conditions from the preset control target mapping set according to index items such as action type, deviation level, and lesion side / healthy side. These parameters include the baseline target pressure distribution relationship of each clamping component and the baseline target displacement of each adjusting rod / support rod. Then, it combines the current user active participation coefficient and the set pressure limit scaling factor to scale and individualize the above-mentioned baseline target pressure. On the one hand, it ensures that the pressure of the three-way airbag does not exceed the preset safety limit. On the other hand, it appropriately lowers the additional pressure of the equipment when the active participation is high and moderately increases the support and correction force when the active participation is low. Finally, it outputs the control target pressure of the airbag inside each clamping component and the control target displacement of each clamping component along the corresponding adjusting rod direction, which serve as the input for subsequent closed-loop control and advance trigger calculation.

[0071] For example, in a training segment during the right lateral curvature correction phase, structured information indicates that the current curvature is moderate, the lesion is on the right side, and the active participation coefficient is 0.7. At this time, the baseline relationship given by the control target mapping set is: the target pressure of the right clamp is higher than that of the left, and the target pressure of the back clamp is slightly increased, with a corresponding baseline pressure difference ΔP_base of 8 kPa. Based on this, the controller sets the pressure limit scaling factor to 0.8 according to the high active participation and the patient's current safe pressure limit, and scales ΔP_base to obtain an actual target pressure difference of approximately 6.4 kPa. It also scales the baseline target displacement of each clamp along the horizontal and anterior-posterior directions accordingly, so that the right clamp generates a moderate corrective force without excessively replacing the patient's own force. This outputs a set of control target pressures and target displacements for the right clamp, left clamp, and back clamp, which are used to drive the device to implement individualized and intensity-adaptive corrective control within this action segment.

[0072] Based on the control target pressure of each clamping element and the control target displacement of each clamping element along the corresponding adjusting rod, calculate the advance trigger time of each clamping element and the ideal valve opening sequence of the internal air bladder of each clamping element.

[0073] In this embodiment, the controller first uses a pre-calibrated dynamic characteristic model of displacement-force-pressure and valve opening-airbag pressure change rate to estimate the response time required for each clamping component to transition from its current pressure and position to the target pressure and target displacement, based on the target pressure and target displacement of each clamping component. Specifically, given the expected effective time of the target action phase, the controller back-calculates the time required for the airbag inside each clamping component to reach the target pressure based on the airbag inflation / deflation time constant, pipeline resistance, and current pressure difference. This time is then differentiated from the expected effective time to obtain the advance trigger time corresponding to each clamping component. Based on this, the controller discretizes and interpolates the airbag pressure target trajectory according to the allowable pressure rise slope and displacement change constraints within each time step. Combined with the airbag model, the controller solves for the valve opening control quantity corresponding to each time slice in the entire action phase, thereby forming the ideal valve opening sequence of the airbag inside each clamping component, so as to realize that the clamping component pressure and displacement approach the control target on schedule and smoothly on the time axis.

[0074] Based on the advance trigger time and the ideal valve opening sequence, configure the correction process for user on-machine training.

[0075] Specifically, the pressure limit scaling factor is set, and the specific analysis process is as follows: Extract basic user characteristic data and evaluate the user's basic characteristic representation coefficient.

[0076] The user's basic feature representation coefficients are matched with the pressure limit scaling factors corresponding to each predefined range of basic feature representation coefficients to determine the specific range of the user's basic feature representation coefficients and obtain the pressure limit scaling factor corresponding to that range.

[0077] The user's basic characteristic representation coefficient and the pressure limit scaling factor can be set as a monotonically positive correlation: when the basic characteristic representation coefficient is low, it indicates that after comprehensively considering bone density, muscle strength, fatigue state and body fat, the user's safe carrying capacity and suitable training intensity are low, and the corresponding pressure limit scaling factor value is small. This is used to lower the maximum and minimum values ​​of the target pressure mapping interval in the control target mapping set as a whole, narrowing and reducing the target pressure mapping adaptation interval. As the basic characteristic representation coefficient increases, it indicates that the user has a higher margin in terms of structural stability and active force exertion capacity. The corresponding pressure limit scaling factor can gradually approach or be slightly higher than 1, so that the target pressure mapping adaptation interval can be moderately widened without exceeding the safety limit, allowing the controller to distribute corrective force at a higher target pressure level.

[0078] Based on the pressure limit scaling factor, multiply it by the maximum and minimum values ​​of the target pressure mapping interval in the control target mapping set to obtain and record the target pressure mapping adaptation interval.

[0079] The aforementioned target pressure mapping adaptation range, on the one hand, scales the upper and lower limits of the target pressure originally set for standard users or standard working conditions according to the current user's tolerance (such as body size, basic characteristics, past records, and degree of active participation), forming a usable target pressure range that is neither overloaded nor lacks sufficient correction effect. On the other hand, when the controller calculates the specific control target pressure of each clamping component based on the action structure information and the active participation coefficient, all calculation results are restricted and clipped within this target pressure mapping adaptation range. Even if the theoretical value given by the mapping relationship or scoring model is too high or too low, it will be pulled back to a safe and reasonable range by this range, thereby ensuring that the target pressure generated under different action segments and different training stages inherits the mechanical design logic of the control target mapping set and always meets the individualized safety limit requirements.

[0080] The user's basic characteristic data includes muscle fatigue level, bone mineral density, average body fat percentage, and grip strength measurement at the start of the exercise training cycle. Muscle fatigue level can be obtained through electromyography (EMG) signal analysis. Bone mineral density can be measured using methods such as DEXA scanning. Body fat percentage is obtained using a body fat scale or magnetic resonance imaging (MRI). Grip strength measurement is obtained using a grip strength meter.

[0081] The aforementioned assessment of the user's basic characteristic representation coefficients specifically involves preprocessing the user's muscle fatigue level, bone density, average body fat percentage, and grip strength measurements. This preprocessing includes normalization and de-unitization. The results of this preprocessing are then used for a joint evaluation.

[0082] In the formula, Here, Bn is the user's basic characteristic representation coefficient, Gn is the user's bone mineral density, Fn is the user's grip strength measurement, Pn is the user's muscle fatigue level, w1 is the weight parameter corresponding to bone mineral density predefined in the rehabilitation training management library, w2 is the weight parameter corresponding to grip strength measurement predefined in the rehabilitation training management library, w3 is the weight parameter corresponding to muscle fatigue predefined in the rehabilitation training management library, and w4 is the weight parameter corresponding to average body fat percentage predefined in the rehabilitation training management library.

[0083] Higher real-time muscle fatigue indicates lower available muscle strength reserves and endurance, which should negatively impact the baseline characteristic coefficient to reduce training intensity and the risk of corrective load. Higher bone density indicates a more stable bone support structure and stronger resistance to pressure and injury, allowing for higher training intensity under safe conditions, and should positively impact the baseline characteristic coefficient. When average body fat percentage exceeds the normal range, it is often accompanied by a decrease in the proportion of effective muscle and an increase in metabolic burden, which should be considered a constraint on the tolerable corrective intensity and should negatively impact the baseline characteristic coefficient. Real-time grip strength measurement can serve as a sensitive indicator of overall muscle strength and neuromuscular control ability; a higher value generally indicates better active force and resistance, and should positively impact the baseline characteristic coefficient.

[0084] Furthermore, these parameters exhibit certain synergistic and restrictive relationships. For instance, individuals with low bone density often experience reduced muscle mass and grip strength, and are more prone to muscle fatigue. Conversely, a higher average body fat percentage may correspond to lower grip strength and faster fatigue accumulation at the same body weight. Short-term increases in muscle fatigue can also temporarily lower grip strength performance. These intrinsic connections dictate that the calculation of the basic characteristic coefficients should employ a combination of weighting and constraints to ensure that the final coefficients reflect the physical meaning of individual indicators while avoiding the underestimation of risks caused by the superposition of multiple indicators.

[0085] Furthermore, the correction process for configuring user-on-machine training is specifically analyzed as follows: The controller extracts the real-time pressure of the air bladder inside each clamping component and the real-time displacement of each clamping component in real time. It then performs difference processing on the control target pressure of each clamping component and the control target displacement of each clamping component along the corresponding adjustment rod to obtain the real-time pressure error of the air bladder inside each clamping component and the real-time displacement error of each clamping component.

[0086] The current pressure error of the air bladder inside each clamping component is compared with the previous pressure error of the air bladder inside each clamping component to obtain the real-time pressure error fluctuation parameter of the air bladder inside each clamping component, and the real-time displacement error fluctuation parameter of each clamping component is also obtained.

[0087] The real-time displacement error fluctuation parameter is obtained by comparing the current displacement error of each clamping component with the previous displacement error of each clamping component.

[0088] Based on the real-time pressure error fluctuation parameters of the airbags inside each clamping component and the real-time displacement error fluctuation parameters of each clamping component, the parameters are verified against the predefined pressure error fluctuation parameter boundary values ​​and displacement error fluctuation parameter boundary values, respectively. Based on the verification results, the correction process for user on-machine training is configured.

[0089] The verification result is expressed as follows: the real-time pressure error fluctuation parameter of the airbag inside a certain clamping component is greater than or equal to the pressure error fluctuation parameter boundary value, or the real-time pressure error fluctuation parameter of the airbag inside a certain clamping component is less than the pressure error fluctuation parameter boundary value.

[0090] The real-time displacement error fluctuation parameter of a certain clamping component is greater than or equal to the displacement error fluctuation parameter boundary value, or the real-time displacement error fluctuation parameter of a certain clamping component is less than the displacement error fluctuation parameter boundary value.

[0091] The specific steps for configuring user-initiated training are as follows: In this embodiment, after generating advance trigger times and ideal valve opening sequences for each clamping component, the controller reads the airbag pressure and clamping component displacement in real time during the training execution phase, calculates the pressure error fluctuation parameters and displacement error fluctuation parameters at adjacent sampling times, and verifies them against predefined pressure error fluctuation thresholds and displacement error fluctuation thresholds respectively. If the amplitude of a certain error fluctuation is lower than the corresponding threshold, it indicates that the error changes slowly over time and is a slow variable. In this case, the controller calls the PI loop to smoothly correct the baseline part of the ideal valve opening sequence based on the dynamic lag compensated by the advance trigger time, so that the pressure and displacement continuously approach the planned target trajectory in the overall trend. If the amplitude of the error fluctuation exceeds the threshold and exhibits rapid abrupt change characteristics, it indicates that it is a fast variable. In this case, the controller superimposes feedforward prediction and small-amplitude PD correction on the current time slice of the ideal valve opening sequence to achieve rapid suppression of sudden errors and prevent overshoot and oscillation caused by rapid inflation and deflation of the airbag, active force exerted by the user, or external disturbances. Through the chain process of early trigger compensation, ideal sequence tracking, error fluctuation classification, and PI / PD adaptive correction, the three-way gripper maintains stable, timely, and precise attitude and pressure control throughout the entire training cycle.

[0092] It should be explained that the control of the three-way airbags in this embodiment adopts a coordinated scheduling strategy. That is, when the controller generates the target pressure and valve opening sequence for each airbag, it does not pressurize or depressurize the three airbags independently and in isolation. Instead, it comprehensively considers the torso correction direction, the desired center of gravity migration path, and the current airbag pressure state, and uniformly plans the inflation and deflation sequence, pressure rise slope, and maximum pressure amplitude of the dorsal airbag and the left and right airbags. For example, when lateral correction is needed, the dorsal support pressure is first moderately increased, and then the pressure of the target side airbag is increased in stages while the pressure change of the non-target side airbag is limited. This makes the three-way airbags form an organized sequence and division of labor on the time axis and pressure axis, thereby reducing the coupling interference and posture oscillation caused by the simultaneous rapid action of multiple airbags while ensuring the reasonable distribution of the overall correction torque and user comfort.

[0093] Reference Figure 2As shown, the second aspect of the present invention provides an artificial intelligence-based method for recognizing and controlling seated rehabilitation training movements, including: a user performing on-machine training on a seated rehabilitation training device, extracting multi-source datasets of rehabilitation training in real time according to an appropriate sampling frequency to detect movement segments, and obtaining a training movement segment set.

[0094] The training action fragment set is input into the artificial intelligence, which outputs the structured information of each training action fragment. At the same time, based on the multi-source dataset of rehabilitation training, the artificial intelligence outputs the user's active participation coefficient.

[0095] Based on the structured information of each training movement segment and the user's active participation coefficient, the controller generates the control target of the seated rehabilitation training device, corrects the user's on-machine training process, and completes the identification and control of seated rehabilitation training.

[0096] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. An artificial intelligence-based seated rehabilitation training movement recognition and control system, characterized in that, include: The motion segment detection module is used by users to perform on-the-machine training on a seated rehabilitation training device. It extracts multi-source datasets of rehabilitation training in real time according to the appropriate sampling frequency to detect motion segments and obtain a set of training motion segments. The AI ​​action recognition module is used to input the training action fragment set into the AI ​​and output the structured information of each training action fragment. At the same time, based on the multi-source dataset of rehabilitation training, the AI ​​outputs the user's active participation coefficient. The on-machine training correction module is used to generate control targets for the seated rehabilitation training equipment based on the structured information of each training action segment and the user's active participation coefficient, correct the user's on-machine training process, and complete the identification and control of seated rehabilitation training.

2. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 1, characterized in that: The user's on-machine training based on the seated rehabilitation training device also includes: The seated rehabilitation training device includes an inertial measurement unit, a pressure sensing unit, a displacement sensing unit, and a rehabilitation training management unit. The rehabilitation training management unit is pre-configured with a set of predetermined training movements and the current training sequence; The user's sitting data is extracted based on the inertial measurement unit and the pressure sensing unit. The user's sitting data includes the user's static torso angle and the user's static maximum pressure. The user's static torso angle and the user's static maximum pressure are respectively denoted as the user's static torso reference angle and the user's static reference center of gravity.

3. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 1, characterized in that: The specific process for real-time extraction of multi-source rehabilitation training datasets according to an adapted sampling frequency for motion segment detection is as follows: Multi-source rehabilitation training data is collected based on the adaptive sampling frequency and arranged sequentially according to the collection time sequence to form a multi-source rehabilitation training dataset. The multi-source rehabilitation training dataset is then divided according to a set sliding time window to obtain the multi-source rehabilitation training data of the user under each sliding time window. Based on the multi-source data of the user's rehabilitation training under each sliding time window, the real-time posture changes and the real-time center of gravity movement of the user under each sliding time window are calculated. The real-time posture changes and center of gravity shifts of the user under each sliding time window are compared with predefined posture change thresholds and center of gravity shift thresholds to detect motion segments. Based on the detection results, it is determined whether to mark the sliding time window as a training motion segment, thus obtaining a training motion segment set.

4. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 3, characterized in that: The specific process for defining the adaptive sampling frequency is as follows: Extract the real-time control bandwidth of the sitting rehabilitation training device, match it with the sampling frequency adjustment factor corresponding to each predefined real-time control bandwidth interval, determine the interval to which the real-time control bandwidth of the sitting rehabilitation training device belongs, and obtain the sampling frequency adjustment factor corresponding to the interval. Extract the baseline sampling frequency corresponding to the current predetermined training action and multiply it by the sampling frequency adjustment factor to obtain and record it as the theoretical lower limit of the sampling frequency; Extract the rated computing power of the edge processor to which the seated rehabilitation training equipment belongs, correlate it with the sampling frequency correction parameter, multiply it with the theoretical lower limit of the sampling frequency, and obtain and record it as the adapted sampling frequency.

5. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 1, characterized in that: The structured information of each training action segment is output, and the specific analysis process is as follows: Extract the user's maximum torso tilt angle and static deviation from the center of gravity during the current training motion segment; The maximum tilt angle of the user's torso is obtained by subtracting the user's maximum tilt angle from the user's static torso reference angle. The user's static deviation from the center of gravity is compared with the user's static reference center of gravity to obtain the user's static center of gravity offset. The structured information of each training action segment includes the action type of each training action segment, the training order of each training action segment, the maximum tilt angle of the user's torso in each training action segment, the user's static center of gravity offset in each training action segment, and the confidence level of each training action segment.

6. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 1, characterized in that: The AI ​​outputs a user's active participation coefficient, and the specific analysis process is as follows: Artificial intelligence analyzes the multi-source dataset of rehabilitation training by calling a scoring model and outputs the user's active participation coefficient. The scoring model calls active participation constraints to accumulate active participation scores to obtain the active participation coefficient. The active participation constraint is specifically as follows: 1) The time point of user posture deviation is greater than the time point of pressure deviation of the seated rehabilitation training equipment; 2) The user's motion angular velocity curve conforms to the preset active motion template curve; 3) The time point when the user's static center of gravity deviates is greater than the time point when the clamping parts of the seated rehabilitation training device displace.

7. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 1, characterized in that: The controller generates control targets for the seated rehabilitation training device and corrects the user's on-machine training process. The specific analysis process is as follows: The seated rehabilitation training device includes various clamping components and various adjusting rods, and each clamping component includes an internal airbag. The structured information of each training action segment and the user's active participation coefficient are input into the AI ​​controller. The controller outputs the control target pressure of each clamping component and the control target displacement of each clamping component along the corresponding adjustment rod through the preset control target mapping set and the set pressure limit scaling factor. Based on the control target pressure of each clamping element and the control target displacement of each clamping element along the corresponding adjusting rod, calculate the advance trigger time of each clamping element and the ideal valve opening sequence of the internal air bladder of each clamping element. Based on the advance trigger time and the ideal valve opening sequence, configure the correction process for user on-machine training.

8. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 7, characterized in that: The specific analysis process for the set pressure limit scaling factor is as follows: Extract users' basic feature data and evaluate the user's basic feature representation coefficients; The user's basic feature representation coefficients are matched with the pressure limit scaling factor corresponding to each predefined interval of basic feature representation coefficients to determine the specific interval of the user's basic feature representation coefficients and obtain the pressure limit scaling factor corresponding to that interval. Based on the pressure limit scaling factor, multiply it by the maximum and minimum values ​​of the target pressure mapping interval in the control target mapping set to obtain and record the target pressure mapping adaptation interval; The user's basic characteristic data includes the user's muscle fatigue level, bone density, average body fat percentage, and grip strength measurement at the start of the exercise training cycle.

9. The artificial intelligence-based seated rehabilitation training movement recognition and control system according to claim 7, characterized in that: The specific analysis process for correcting the configuration of user on-machine training is as follows: The controller extracts the real-time pressure of the airbag inside each clamping component and the real-time displacement of each clamping component in real time. It then performs difference processing on the control target pressure of each clamping component and the control target displacement of each clamping component along the corresponding adjustment rod to obtain the real-time pressure error of the airbag inside each clamping component and the real-time displacement error of each clamping component. The pressure error of the air bladder inside each clamping component at the current moment is compared with the pressure error of the air bladder inside each clamping component at the previous moment to obtain the real-time pressure error fluctuation parameter of the air bladder inside each clamping component, and the real-time displacement error fluctuation parameter of each clamping component is also obtained. Based on the real-time pressure error fluctuation parameters of the airbags inside each clamping component and the real-time displacement error fluctuation parameters of each clamping component, the parameters are verified against the predefined pressure error fluctuation parameter boundary values ​​and displacement error fluctuation parameter boundary values, respectively. Based on the verification results, the correction process for user on-machine training is configured.

10. A method for recognizing and controlling seated rehabilitation training movements based on artificial intelligence, applied to the artificial intelligence-based seated rehabilitation training movement recognition and control system described in any one of claims 1-9, characterized in that: include: Users perform on-the-machine training based on the seated rehabilitation training device. The multi-source dataset of rehabilitation training is extracted in real time according to the appropriate sampling frequency to detect movement segments and obtain a set of training movement segments. The training action fragment set is input into the artificial intelligence, which outputs the structured information of each training action fragment. At the same time, based on the multi-source dataset of rehabilitation training, the artificial intelligence outputs the user's active participation coefficient. Based on the structured information of each training movement segment and the user's active participation coefficient, the controller generates the control target of the seated rehabilitation training device, corrects the user's on-machine training process, and completes the identification and control of seated rehabilitation training.

Images