Ankle pump exercise state recognition method based on multi-modal data acquisition

Abstract

The application discloses an ankle pump motion state recognition method based on multi-modal data collection, comprising the following steps: synchronously collecting multi-source sensing data of a patient in an ankle pump motion process through a sensor arranged on an ankle wearing assembly, and extracting motion characteristic parameters from each modal data; performing space-time domain fusion on the motion characteristic parameters extracted from each modal to construct a joint motion characteristic vector reflecting full-dimensional information of the ankle pump motion; inputting the joint motion characteristic vector into a pre-constructed ankle pump motion recognition model; the ankle pump motion recognition model outputs an action category recognition result of the current ankle pump motion; generating a corresponding motion control instruction according to the action category recognition result; generating training record data and outputting the training record data to an external device or a cloud platform, which improves the compliance and effectiveness of patient training, reduces the risk of lower extremity deep venous thrombosis complications, and promotes the ankle function rehabilitation, and has significant clinical application value.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, specifically to a method for identifying ankle pump motion state based on multimodal data acquisition. Background Technology

[0002] Ankle pump exercises refer to the rhythmic contraction and relaxation of the calf muscles through active or passive flexion, extension, and circumduction of the ankle joint. This muscle pump action promotes blood and lymph return from the lower limbs and is one of the simplest and most economical basic preventive measures for preventing deep vein thrombosis (DVT), reducing limb swelling, and promoting ankle joint function recovery. It is widely applicable to high-risk groups for DVT, such as those who have undergone orthopedic surgery, those who are bedridden for a long time, pregnant women, and the elderly. Currently, clinical monitoring and evaluation of ankle pump exercises mainly rely on the following technical approaches: First, monitoring methods based on a single sensor, such as using only a posture sensor to detect ankle joint movement angle -3, using only a pressure sensor to detect plantar force, or using a vibration sensor placed under the ankle to collect movement signals; Second, rehabilitation movement recognition methods based on computer vision, which use cameras to collect movement images and use posture estimation algorithms to extract joint movement parameters; Third, some solutions attempt to introduce the concept of personalized training prescriptions, matching preset ranges of movement parameters based on the patient's basic information. (1) Existing ankle pump motion monitoring devices are mostly single sensor solutions. Although posture sensors can detect changes in joint angles, they cannot obtain the force distribution and force application state of the ankle. Although pressure sensors can sense contact force, they cannot accurately measure the range of motion and trajectory. Although vision solutions can capture the trajectory, they are easily affected by occlusion and lighting conditions. Each single modal data can only reflect some dimensions of ankle pump motion information and cannot achieve comprehensive synchronous acquisition of motion angle, speed, amplitude, force distribution and trajectory, resulting in insufficient recognition of compensatory movements. (2) When formulating training programs, existing technologies often adopt fixed and uniform exercise frequency, amplitude and duration standards, which fail to fully consider individual differences such as patient age, weight, muscle strength, postoperative days and thrombosis risk level, lack personalized adaptive control capabilities, and cannot dynamically adjust training parameters according to individual patient differences. (3) Most existing technologies can only output the results of motion state determination and cannot actively intervene and forcibly correct non-standard movements when they are detected. In actual training, patients often lack real-time guidance and correction, resulting in repeated compensatory movements, which greatly reduces the effectiveness of training and cannot truly guarantee the standardization and safety of ankle pump exercise. Summary of the Invention

[0003] The purpose of this invention is to provide an ankle pump motion state recognition method based on multimodal data acquisition, in order to solve the technical problems in the prior art, such as difficulty in fully capturing the fine features of ankle pump motion, lack of personalized adaptive control capability, inability to dynamically adjust training parameters according to individual patient differences, and lack of closed-loop error correction mechanism.

[0004] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: This invention provides a method for identifying ankle pump motion state based on multimodal data acquisition, comprising the following steps: The sensor on the ankle wearing component synchronously collects multi-source sensor data of the patient during the ankle pump exercise. The multi-source sensor data is preprocessed and motion feature parameters are extracted from each modal data. The motion feature parameters include at least the angle parameter, frequency parameter, amplitude parameter, pressure distribution parameter and motion trajectory parameter of the ankle pump exercise. The motion feature parameters extracted from each modality are fused in the spatiotemporal domain to construct a joint motion feature vector that reflects the full-dimensional information of ankle pump movement. The joint motion feature vector is then input into a pre-constructed ankle pump movement recognition model, which is trained based on standard ankle pump movement samples and compensatory error movement samples and has a built-in standard movement judgment threshold and compensatory movement feature library. The ankle pump motion recognition model outputs the action category recognition result of the current ankle pump motion by comparing real-time motion features with the judgment threshold and feature library. The recognition result includes at least the classification judgment of standard ankle pump motion and compensatory error motion. Based on the action category recognition result, a corresponding motion control command is generated. The control command includes at least one or more of the motion resistance parameter, motion speed parameter and motion stroke parameter output by the ankle pump motion actuator, so as to make the ankle pump motion state approach the standard motion mode. Record multi-source sensor data, motion feature parameters, action recognition results and control parameters at each stage of the training process, generate training record data, and output the training record data to external devices or cloud platforms.

[0005] As a preferred embodiment of the present invention, the multi-source sensing data includes: The multi-source sensing data includes at least angle sensing data reflecting changes in ankle joint angle, pressure sensing data reflecting the force state of the foot and ankle, posture sensing data reflecting the spatial orientation of the foot and ankle, and image sensing data reflecting the movement trajectory of the ankle joint. The multi-source sensor data is time-stamped and aligned so that the data of each modality are under the same time reference. The angle sensing data is continuously collected by an angle sensor at a first sampling frequency during the real-time angle values ​​of the ankle joint during dorsiflexion and plantarflexion. The pressure sensing data is obtained by multiple pressure sensing units distributed on the inner side of the ankle wearing component, which collect the dynamic pressure distribution value of the contact surface between the ankle and the wearing component at a second sampling frequency. The attitude sensing data is collected by an attitude sensor, which captures the acceleration and angular velocity values ​​of the ankle in three-dimensional space. The image sensing data is a sequence of continuous image frames acquired by a miniature vision camera during ankle joint movement. The raw data for each modality were subjected to sliding window filtering for noise reduction, baseline drift correction, and outlier removal.

[0006] As a preferred embodiment of the present invention, the method for extracting and fusing motion feature parameters includes: The peak dorsiflexion angle, peak plantarflexion angle, and duration of a single movement cycle are extracted from the angle sensing data. Extract the pressure peak value and pressure center of gravity offset trajectory of each area of ​​the foot and ankle from the pressure sensing data; Extract the peak acceleration and angular velocity variation curve of the ankle movement from the posture sensing data; Extract the pixel coordinate sequence of key points of the ankle joint from the image sensing data and convert it into a spatial motion trajectory; The motion feature parameters extracted from each modality are aligned along the time axis according to a unified timestamp to form a time series feature set; The time series feature set is segmented using a sliding time window with a preset duration and a preset step size, and the features of multiple time nodes within each time window are combined into a time series feature matrix. Calculate the cross-correlation features between different modal features, wherein the cross-correlation features include at least angle-pressure temporal correlation features, attitude-trajectory spatial consistency features, and angle-angular velocity phase synchronization features; The original features are combined with the cross-correlation features to generate a joint motion feature vector. The joint motion feature vector is then normalized and dimensionality-reduced before being input into the ankle pump motion recognition model.

[0007] As a preferred embodiment of the present invention, the method for constructing the ankle pump motion recognition model includes: Several clinical standard ankle pump exercise data and several compensatory error exercise data were collected as training sample sets. The compensatory error exercises were classified and labeled according to the compensation type. The compensation type includes at least four categories: foot and ankle compensation, hip joint compensation, trunk compensation, and insufficient range of motion. Each type of compensation sample was labeled with the corresponding compensation category label. The standard ankle pump exercise samples were labeled with the standard action label. A joint motion feature vector is extracted from each training sample, and the corresponding key feature discrimination parameters and feature threshold ranges are extracted for each type of compensation to form a compensation action feature library. The compensation action feature library contains the mapping relationship between each compensation type and the key feature discrimination parameters. Using the labeled joint motion feature vector as input and the action category label as output, a classification and recognition network model is trained to obtain the initial ankle pump motion recognition model.

[0008] As a preferred embodiment of the present invention, the recognition process of the ankle pump motion recognition model includes: The joint motion feature vector generated by real-time acquisition and fusion is input into the ankle pump motion recognition model. The ankle pump motion recognition model first compares the real-time motion features with the built-in standard action judgment threshold, and outputs a binary classification judgment result on whether the current motion is a standard ankle pump action. If it is determined to be a non-standard action, the real-time motion features are further matched with the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library, and the compensation type identification result of the current motion is output. The ankle pump motion recognition model calculates the confidence score corresponding to the action category recognition result while outputting the action category recognition result; When the confidence score is higher than the preset high confidence threshold, the final recognition result is directly output; When the confidence score is between a preset low confidence threshold and a high confidence threshold, a voting mechanism for recognition results of multiple consecutive frames within a sliding time window is triggered, and the voting result is used as the final recognition output. When the confidence score is lower than the preset low confidence threshold, the movement is judged as a suspicious action, triggering a re-acquisition and recognition process; The marked suspicious action data and the misjudged data confirmed by medical staff are stored in the sample set. The ankle pump movement recognition model is periodically updated through incremental learning to update the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library.

[0009] As a preferred embodiment of the present invention, the confidence threshold is set according to a standardized ankle pump exercise movement specification threshold library, which includes reference threshold ranges for standard ankle pump movements at different rehabilitation stages. The reference threshold range includes at least the standard ankle dorsiflexion angle threshold range, the standard plantarflexion angle threshold range, the standard movement rate threshold range, the single standard movement frequency threshold range, and the standard force distribution reference interval of the foot and ankle.

[0010] In a preferred embodiment of the present invention, the ankle pump motion recognition model incorporates an AI adaptive control algorithm, which generates corresponding motion control commands. The AI ​​adaptive control algorithm includes: The patient's basic vital signs data and real-time motion state recognition results are used to construct a design feature vector, which is then input into a pre-trained adaptive control network model. The adaptive regulation network model adopts a multilayer perceptron structure, and its training samples include exercise training data and corresponding effective training parameters of patients of different ages, weights, injury degrees and rehabilitation stages. The adaptive control network model takes maximizing training effectiveness and minimizing compensation risk as its objective function, and outputs a personalized ankle pump exercise training program that matches the patient's current state, including the movement amplitude threshold, training frequency, single training duration, and rest interval duration. The training program is dynamically updated as the patient's rehabilitation progresses and their real-time movement status changes.

[0011] As a preferred embodiment of the present invention, the AI ​​adaptive control algorithm generates corresponding motion control commands based on the action category recognition results, and is equipped with a closed-loop intelligent error correction intervention mechanism, the closed-loop intelligent error correction intervention mechanism including: The real-time motion state recognition results are compared item by item with the standard parameters set in the personalized ankle pump exercise training program to calculate the deviation value between the actual motion state and the standard program. When the deviation value exceeds the preset tolerance range, it is determined to be a deviation condition, and a graded error correction prompt instruction is generated according to the deviation type and degree. The hierarchical error correction prompts include at least a first-level voice prompt error correction instruction and a second-level visual interface error correction instruction. The first-level voice prompt correction instruction is used to provide specific corrective action guidance via voice broadcast when a slight deviation in the operating condition is detected; The secondary visual interface error correction command is used to simultaneously display an animated illustration of the deviation location and a demonstration of the correct action on the interactive interface when a moderate or higher level of deviation is detected. Continuously collect multimodal motion data after error correction intervention and repeat the above steps until the deviation value falls back to the preset tolerance range.

[0012] As a preferred embodiment of the present invention, the patient's rehabilitation effect is dynamically monitored based on the action category recognition result, and the ankle pump movement state is brought closer to the standard movement pattern through the control command, including: Record the total number of movements, the number of standard movements, and the number of deviation movements in a single training session for the patient, and calculate the movement attainment rate; Record the ratio of the actual training time of a single patient session to the training time set in the protocol, and calculate training compliance; The trends of the patients’ achievement rate and range of motion during training over several consecutive days were statistically analyzed, and the progress of ankle joint function rehabilitation was assessed based on these trends. When the rate of achieving the target for the movement within a set number of consecutive days shows an upward trend and the range of movement gradually approaches the standard threshold range, a good rehabilitation progress assessment result is output. When the rate of achieving the target for the exercise shows a downward trend or stagnates within a set number of consecutive days, the system outputs a slow recovery progress assessment result and triggers a process to readjust the training program.

[0013] In a preferred embodiment of the present invention, during the dynamic monitoring of the rehabilitation effect of a patient, a risk warning is provided to indicate the ankle pump movement status. The risk warning method includes: The relevant feature parameters reflecting the blood circulation status of the lower limbs are extracted from real-time motion data. The relevant feature parameters include at least the peak angular velocity of ankle pump movement, motion frequency stability index and foot and ankle force recovery curve slope. The associated feature parameters are input into a preset risk assessment model, which has built-in risk thresholds for lower extremity deep vein thrombosis and limb swelling. When the associated feature parameters exceed the risk discrimination threshold, it is determined that there is a risk of complications such as deep vein thrombosis or limb swelling in the lower extremities. An early warning signal containing the risk level and risk type is output and the early warning signal is pushed to the medical terminal simultaneously.

[0014] Compared with the prior art, the present invention has the following advantages: This invention utilizes multimodal synchronous acquisition of ankle pump motion data via angle sensors, pressure sensors, posture sensors, and a miniature vision camera. Combined with deep spatiotemporal fusion and cross-modal cross-correlation feature calculation, it comprehensively captures motion angle, force distribution, spatial pose, and motion trajectory information, significantly improving the accuracy of compensatory motion recognition. A standard-compensation dual-path fine-grained discrimination model and a compensatory motion feature library are constructed, enabling accurate classification and recognition of four types of compensation: ankle-foot compensation, hip joint compensation, trunk compensation, and insufficient range of motion. Furthermore, an AI adaptive control algorithm integrates patient basic vital signs, real-time motion status, and historical training data to dynamically generate personalized training programs, overcoming the limitations of traditional methods. The limitations of fixed training modes are overcome; the closed-loop error correction intervention mechanism upgrades passive monitoring to active intervention, and the three-level confidence level differential processing and incremental learning mechanism continuously optimize model performance; at the same time, training data is fully recorded and uploaded to the cloud to generate rehabilitation reports, realizing digital and traceable management of rehabilitation nursing, effectively reducing nursing labor costs, and realizing the technological leap of ankle pump rehabilitation training from experience to intelligence, and from standardization to personalization. It has significant clinical application value in improving training compliance and effectiveness, reducing the risk of complications of deep vein thrombosis in the lower extremities, and promoting ankle joint functional rehabilitation. Moreover, the system has a simple structure, is suitable for various bedridden departments, and has good prospects for industrialization and promotion. Attached Figure Description

[0015] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0016] Figure 1 A flowchart of an ankle pump motion state recognition method based on multimodal data acquisition provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this invention provides a method for identifying ankle pump motion state based on multimodal data acquisition, including the following steps: The sensor on the ankle wearing component synchronously collects multi-source sensor data of the patient during the ankle pump exercise. The multi-source sensor data is preprocessed and motion feature parameters are extracted from each modal data. The motion feature parameters include at least the angle parameter, frequency parameter, amplitude parameter, pressure distribution parameter and motion trajectory parameter of the ankle pump exercise. The motion feature parameters extracted from each modality are fused in the spatiotemporal domain to construct a joint motion feature vector that reflects the full-dimensional information of ankle pump movement. The joint motion feature vector is then input into a pre-constructed ankle pump movement recognition model, which is trained based on standard ankle pump movement samples and compensatory error movement samples and has a built-in standard movement judgment threshold and compensatory movement feature library. In this embodiment, angle sensors, pressure sensors, posture sensors, and miniature vision cameras installed on the ankle-foot wearing component simultaneously collect four-dimensional data: ankle joint movement angle, ankle force distribution, spatial pose, and movement trajectory. This overcomes the limitation of a single sensor only being able to acquire local information. By synchronizing and aligning the timestamps of each modality's data, interpolating and resampling, and preprocessing with sliding window filtering, accurate matching of multi-source data in the spatiotemporal dimension is ensured. Furthermore, by extracting the motion feature parameters of each modality and calculating cross-correlation features such as angle-pressure temporal correlation, posture-trajectory spatial consistency, and angle-angular velocity phase synchronization, the original features and cross-correlation features are fused to generate a joint motion feature vector. This enables the recognition model to comprehensively utilize multi-dimensional information for judgment, effectively solving the problem of insufficient recognition accuracy of single-modality data in complex clinical scenarios such as occlusion, signal drift, and individual differences, and significantly improving the detection rate of compensatory movements.

[0019] In this embodiment, instead of simply splicing the data from each modality, a time-series feature matrix is ​​constructed through time axis alignment and sliding window segmentation. Furthermore, cross-modal cross-correlation features such as angle-pressure temporal correlation, attitude-trajectory spatial consistency, and angle-angular velocity phase synchronization are calculated. The original features and deep correlation features are fused together to generate a joint motion feature vector. This spatiotemporal deep fusion strategy enables the recognition model to comprehensively utilize the collaborative and complementary information between each modality, effectively solving the problem of insufficient recognition accuracy of single-modal data in complex clinical scenarios such as signal drift, occlusion interference, and individual differences. It significantly improves the detection rate of compensatory movements. The recognition accuracy of this invention for standard ankle pump movements reaches 94.6%, and the average recognition accuracy for four common compensatory movements reaches 89.2%, which has significant advantages compared to single-sensor solutions.

[0020] The ankle pump motion recognition model outputs the action category recognition result of the current ankle pump motion by comparing real-time motion features with the judgment threshold and feature library. The recognition result includes at least the classification judgment of standard ankle pump motion and compensatory error motion. In this embodiment, the constructed ankle pump motion recognition model not only outputs a binary classification of standard movements and compensatory movements, but also further performs fine-grained classification of compensatory types. By establishing a compensatory movement feature library containing key feature discrimination parameters and feature threshold ranges for each compensatory type, the model can accurately identify specific compensatory patterns after judging non-standard movements. At the same time, through a strategy combining confidence scoring and multi-frame voting mechanism, the interference of single-frame misjudgment on the overall recognition result is effectively avoided. For low-confidence suspicious movements, a re-acquisition and recognition and periodic incremental learning update mechanism are adopted to continuously optimize the model during clinical use and continuously improve its adaptability to specific patient groups.

[0021] Based on the action category recognition result, a corresponding motion control command is generated. The control command includes at least one or more of the motion resistance parameter, motion speed parameter and motion stroke parameter output by the ankle pump motion actuator, so as to make the ankle pump motion state approach the standard motion mode. Record multi-source sensor data, motion feature parameters, action recognition results and control parameters at each stage of the training process, generate training record data, and output the training record data to external devices or cloud platforms.

[0022] In this embodiment, an AI adaptive control algorithm is used to combine basic vital signs data such as patient age, weight, ankle joint injury severity, and rehabilitation stage with real-time movement state recognition results. These data are then input into a pre-trained multilayer perceptron adaptive control network model. With the objective function of maximizing training effectiveness and minimizing compensatory risk, the model dynamically outputs personalized movement amplitude thresholds, training frequency, single training duration, and rest intervals that match the patient's current state. This fully considers individual differences among patients in terms of injury severity, tolerance, and rehabilitation progress, effectively avoiding problems such as decreased compliance or poor rehabilitation effects caused by mismatched training intensity. Furthermore, the training plan is dynamically iterated and updated as the patient's rehabilitation stage progresses and real-time movement state changes, achieving truly adaptive intelligent rehabilitation.

[0023] The multi-source sensing data includes: The multi-source sensing data includes at least angle sensing data reflecting changes in ankle joint angle, pressure sensing data reflecting the force state of the foot and ankle, posture sensing data reflecting the spatial orientation of the foot and ankle, and image sensing data reflecting the movement trajectory of the ankle joint. In this embodiment, angle sensors, pressure sensors, posture sensors, and a miniature vision camera installed on the ankle-foot wearing component simultaneously collect data on ankle joint angle changes, ankle force distribution, ankle spatial posture, and ankle motion trajectory images. The angle sensor accurately measures real-time angle changes during ankle dorsiflexion and plantarflexion, providing the most direct quantitative indicators of motion amplitude. Multiple pressure sensing units distributed on the inner side of the wearing component simultaneously collect dynamic pressure distribution values ​​between the ankle and the wearing component, reflecting the force application status and force uniformity in different areas of the patient's ankle. The posture sensor captures spatial posture changes of the ankle through triaxial acceleration and angular velocity data, providing key criteria for identifying trunk compensation and limb displacement. The miniature vision camera collects continuous image frame sequences, reconstructing the true spatial motion trajectory of the ankle joint through image data. These four types of sensor data complement and cross-validate each other from different physical dimensions, jointly constituting a comprehensive data acquisition system for ankle pump movement, laying a solid data foundation for subsequent accurate identification and intelligent control.

[0024] The multi-source sensor data is time-stamped and aligned so that the data of each modality are under the same time reference. In this embodiment, the 100Hz angle sensing data, 50Hz pressure sensing data, 100Hz attitude sensing data, and 30fps image sensing data with different sampling frequencies are synchronized and aligned with a unified timestamp, so that all modal data are under the same time reference. This solves the problem of data asynchrony caused by different sampling frequencies, start-up time differences, and clock drift of multi-source heterogeneous sensors. It ensures that the angle value, pressure distribution value, attitude data, and image frame collected at each moment can accurately correspond to the same instant of motion, avoiding feature fusion distortion and action misjudgment caused by time misalignment. It provides multimodal data input with accurate time alignment for subsequent spatiotemporal fusion and accurate recognition.

[0025] The angle sensing data is continuously collected by an angle sensor at a first sampling frequency during the real-time angle values ​​of the ankle joint during dorsiflexion and plantarflexion. The pressure sensing data is obtained by multiple pressure sensing units distributed on the inner side of the ankle wearing component, which collect the dynamic pressure distribution value of the contact surface between the ankle and the wearing component at a second sampling frequency. The attitude sensing data is collected by an attitude sensor, which captures the acceleration and angular velocity values ​​of the ankle in three-dimensional space. The image sensing data is a sequence of continuous image frames acquired by a miniature vision camera during ankle joint movement. The raw data for each modality were subjected to sliding window filtering for noise reduction, baseline drift correction, and outlier removal.

[0026] In this embodiment, the raw data of each modality are subjected to sliding window filtering for noise reduction. Sliding window filtering effectively removes high-frequency random noise caused by patient limb tremors, bed vibrations, cable disturbances, etc., by applying mean or median filtering to the sampling points within a preset window length, while retaining the effective low-frequency components of ankle pump movement. Compared with traditional low-pass filters with fixed cutoff frequencies, sliding window filtering has better adaptability and real-time performance, and can adapt to the differences in movement frequencies of different patients. Baseline drift correction addresses the baseline offset caused by temperature changes, loosening of the sensor, or sensor zero drift during long-term wear. It restores the true zero reference through polynomial fitting or high-pass filtering methods, preventing the accumulation of baseline drift from causing measurement deviations in key parameters such as angle and pressure, and ensuring the accuracy and consistency of long-term monitoring data.

[0027] The methods for extracting and fusing motion feature parameters include: The peak dorsiflexion angle, peak plantarflexion angle, and duration of a single movement cycle are extracted from the angle sensing data. Extract the pressure peak value and pressure center of gravity offset trajectory of each area of ​​the foot and ankle from the pressure sensing data; Extract the peak acceleration and angular velocity variation curve of the ankle movement from the posture sensing data; Extract the pixel coordinate sequence of key points of the ankle joint from the image sensing data and convert it into a spatial motion trajectory; The motion feature parameters extracted from each modality are aligned along the time axis according to a unified timestamp to form a time series feature set; The time series feature set is segmented using a sliding time window with a preset duration and a preset step size, and the features of multiple time nodes within each time window are combined into a time series feature matrix. In this embodiment, the motion feature parameters extracted from each modality are aligned along the time axis according to a unified timestamp, forming a time series feature set containing all feature parameters at each time node. This ensures that angle, pressure, attitude, and trajectory features accurately correspond to the same instant of motion in the time dimension. A sliding time window with a preset duration of 2 seconds and a preset step size of 0.5 seconds is used to segment the time series feature set. The features of multiple time nodes within each time window are combined into a time series feature matrix. The row vectors of this matrix represent the feature parameters of each modality at the same instant, and the column vectors represent the time evolution sequence of the same feature parameter within the time window. This sliding window segmentation strategy allows the feature matrix to simultaneously retain the multi-attribute distribution of each modality parameter in the spatial dimension and the dynamic evolution law in the time dimension. This provides a joint feature input with both spatiotemporal correlation structure for the subsequent recognition model, enabling the model to make comprehensive judgments based on the spatiotemporal features within a complete motion cycle. This avoids the one-sidedness and instability of relying solely on single-frame features or instantaneous values ​​for judgment.

[0028] Calculate the cross-correlation features between different modal features, wherein the cross-correlation features include at least angle-pressure temporal correlation features, attitude-trajectory spatial consistency features, and angle-angular velocity phase synchronization features; In this embodiment, the angle-pressure temporal correlation feature is used to reflect the temporal linkage between changes in ankle joint movement angle and changes in foot and ankle force. If the angle increases but the force does not increase synchronously, it suggests that there may be insufficient force or compensation. The posture-trajectory spatial consistency feature is used to describe the degree of matching between changes in foot and ankle spatial posture and actual movement trajectory. If the posture data shows inversion of the ankle but the trajectory shows that the movement direction is deviated outward, it suggests that there is an abnormal spatial posture. The angle-angular velocity phase synchronization feature is used to reflect the phase relationship between the joint angle and its rate of change. If the angle has reached its peak but the angular velocity is still at a high level, it suggests that the movement rhythm is abnormal or there is a pause. These cross-correlation features are not a simple repetition of the original features of each modality, but reveal deep-level motion coordination and motion quality information that cannot be obtained by single-modal independent analysis. They are an effective supplement and enhancement to single-dimensional features.

[0029] The original features are combined with the cross-correlation features to generate a joint motion feature vector. The joint motion feature vector is then normalized and dimensionality-reduced before being input into the ankle pump motion recognition model.

[0030] In this embodiment, the original features and cross-correlation features are combined to generate a joint motion feature vector. This vector contains both the motion parameter information of each modality and the collaborative discrimination information between cross-modalities. For example, when a single angle data shows that the motion amplitude is up to standard, if the angle-pressure correlation feature shows that the angle is achieved under abnormally reduced force, it indicates that there may be compensation or insufficient force. The model can make a more accurate comprehensive judgment based on this. The joint motion feature vector comprehensively describes the complete information of ankle pump movement from three levels: original parameters of multimodalities, temporal evolution law and cross-modal correlation relationship. It effectively solves the technical problems of insufficient single modal feature information and underutilization of cross-modal information, and significantly improves the recognition model's ability to discriminate complex clinical scenarios.

[0031] The method for constructing the ankle pump motion recognition model includes: Several clinical standard ankle pump exercise data and several compensatory error exercise data were collected as training sample sets. The compensatory error exercises were classified and labeled according to the compensation type. The compensation type includes at least four categories: foot and ankle compensation, hip joint compensation, trunk compensation, and insufficient range of motion. Each type of compensation sample was labeled with the corresponding compensation category label. The standard ankle pump exercise samples were labeled with the standard action label. In this embodiment, by labeling and training multiple types of compensatory samples, the model can not only determine whether the action is standard, but also accurately identify the specific type of compensation, which significantly improves the model's coverage and recognition capabilities for complex clinical compensatory scenarios.

[0032] A joint motion feature vector is extracted from each training sample, and the corresponding key feature discrimination parameters and feature threshold ranges are extracted for each type of compensation to form a compensation action feature library. The compensation action feature library contains the mapping relationship between each compensation type and the key feature discrimination parameters. In this embodiment, by constructing a compensatory action feature library, clinical kinematic knowledge is organically combined with a data-driven model, enabling the recognition process to have both high accuracy and strong interpretability. This facilitates clinical medical staff to understand and trust the model's judgment criteria, while also providing an operable parameterized interface for subsequent clinical manual verification and correction of the feature library.

[0033] Using the labeled joint motion feature vector as input and the action category label as output, a classification and recognition network model is trained to obtain the initial ankle pump motion recognition model.

[0034] In this embodiment, the regularized matching of the compensatory feature library is used as an auxiliary discrimination criterion, which greatly reduces the model's dependence on the amount of labeled samples, improves the model's generalization ability and stability under the condition of limited labeled data, and facilitates the rapid updating and expansion of the compensatory feature library without retraining the entire network.

[0035] The recognition process of the ankle pump motion recognition model includes: The joint motion feature vector generated by real-time acquisition and fusion is input into the ankle pump motion recognition model. The ankle pump motion recognition model first compares the real-time motion features with the built-in standard action judgment threshold, and outputs a binary classification judgment result on whether the current motion is a standard ankle pump action. If it is determined to be a non-standard action, the real-time motion features are further matched with the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library, and the compensation type identification result of the current motion is output. The ankle pump motion recognition model calculates the confidence score corresponding to the action category recognition result while outputting the action category recognition result; When the confidence score is higher than the preset high confidence threshold, the final recognition result is directly output; When the confidence score is between a preset low confidence threshold and a high confidence threshold, a voting mechanism for recognition results of multiple consecutive frames within a sliding time window is triggered, and the voting result is used as the final recognition output. When the confidence score is lower than the preset low confidence threshold, the movement is judged as a suspicious action, triggering a re-acquisition and recognition process; The marked suspicious action data and the misjudged data confirmed by medical staff are stored in the sample set. The ankle pump movement recognition model is periodically updated through incremental learning to update the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library.

[0036] In this embodiment, the data processing procedure and recognition process of the ankle pump motion recognition model are as follows: Multimodal raw data acquisition and preprocessing: The ankle-foot wearing component is fixed to the affected ankle. After the device is activated, the angle sensor continuously acquires real-time angle values ​​during ankle dorsiflexion and plantar flexion at a sampling frequency of 100Hz; 16 thin-film pressure sensing units simultaneously acquire dynamic pressure distribution values ​​at a sampling frequency of 50Hz between the ankle and the wearing component; the posture sensor acquires acceleration and angular velocity values ​​of the ankle in three-dimensional space at a sampling frequency of 100Hz; and the miniature vision camera acquires continuous image frame sequences during ankle joint movement at a frame rate of 30fps. Sliding window filtering was applied to the raw data of each modality for noise reduction, with a window size of 10 sampling points. Third-order spline interpolation was used to resample the modal data of different sampling frequencies to a uniform time resolution of 100Hz. Baseline drift correction and 3σ principle outlier removal were performed on each modal data. All preprocessed data were spatially matched according to the acquisition timestamp, so that the image frame, inertial data frame and pressure data frame at each moment constituted a spatiotemporally aligned multimodal preprocessed data unit. Single-dimensional motion feature parameter extraction: For the preprocessed angle sensing data, detect local extreme points in the angle time sequence signal, extract the peak dorsiflexion angle and peak plantarflexion angle within a single motion cycle, and record the time interval between two adjacent extreme points in the same direction as the duration of a single motion cycle. In this embodiment, the duration of a single motion cycle is set to 3.2 seconds. The preprocessed pressure sensing data was divided into three regions according to anatomical location: 4 units in the heel region, 6 units in the arch region, and 6 units in the forefoot region. The pressure peak value of each region during the movement cycle was extracted, and the two-dimensional coordinates of the pressure center of gravity at each moment were calculated using the weighted average method. The offset trajectory curve of the pressure center of gravity during the movement process was plotted. For the preprocessed attitude sensing data, the peak value of the triaxial acceleration composite vector during the motion cycle is extracted, and the temporal variation curve of the triaxial angular velocity composite vector is extracted. The peak value of angular velocity and the moment when angular velocity crosses zero are recorded to reflect the time node of ankle joint flexion and extension direction switching. For the preprocessed image sensing data, the image frames are input into a lightweight pose estimation network with MobileNetV2 as the backbone network to extract the pixel coordinates of the inner and outer ankle key points in the ankle joint area. The pixel coordinates are converted into three-dimensional spatial coordinates by combining the camera intrinsic parameter matrix and depth estimation, and the three-dimensional spatial motion trajectory coordinate sequence of the ankle key points in the continuous image frames is recorded. The feature parameters extracted from each modality are aligned along the timeline using a unified timestamp, forming a time series feature set where each time node contains 36 dimensions of features, including 3-dimensional angle features, 12-dimensional pressure features, 6-dimensional posture features, 12-dimensional trajectory features, and 3-dimensional kinematic derived features. The time series feature set is segmented using a sliding time window with a preset duration of 2 seconds and a preset step size of 1 second. Each time window covers 20 time nodes, forming a 20×36-dimensional time series feature matrix. The original time series feature matrix of 20×36=720 dimensions is flattened into a one-dimensional feature vector, which is then concatenated with the three types of cross-correlation features of 12 dimensions to obtain a 732-dimensional joint motion feature vector. After Z-score normalization, the vector is reduced to 64 dimensions using principal component analysis, retaining 95% of the information content of the original features. The 64-dimensional joint motion feature vector after dimensionality reduction is then input into the ankle pump motion recognition model. First, the core parameters in the real-time motion features, such as angle, speed, frequency, and force distribution, are compared with the built-in standard motion judgment thresholds item by item. If real-time monitoring shows that the peak dorsiflexion angle is 24°, the peak plantarflexion angle is 30°, the cycle duration is 2.8 seconds, the force ratio between the heel and forefoot is 1.05, and all dimensions fall within the threshold range, the model outputs a binary classification result that the movement is a standard ankle pump action. If real-time monitoring shows that the peak dorsiflexion angle of the patient is only 14° in a certain exercise cycle, and other parameters are normal, then the current exercise is determined to be a non-standard movement, triggering the subsequent compensation type identification process.

[0037] In this embodiment, by triggering a voting mechanism for consecutive multi-frame recognition results within a sliding time window, the characteristic of ankle pump movement being periodically repetitive is utilized. Temporal consistency verification is used to reduce misjudgments caused by occasional fluctuations in a single frame. The voting results are used as the final recognition output, effectively improving the recognition stability in the medium confidence region. When the confidence level is lower than the preset low confidence threshold, the recognition result is deemed unreliable, and the movement is judged as a suspicious action, triggering a re-acquisition and recognition process. This avoids the risk of misoperation that may be caused by low-quality data driving control commands.

[0038] In this embodiment, for recognition scenarios with a medium level of confidence, a voting mechanism of continuous multi-frame recognition results within a sliding time window is adopted as the final recognition basis. Ankle pump movements naturally have periodic repetitive characteristics, that is, the standard ankle pump movement consists of continuous movement phases such as dorsiflexion, plantarflexion, and return to neutral position. The movement categories have temporal consistency in adjacent movement cycles. When the recognition confidence of a single frame is insufficient, the recognition results of its preceding and following frames are fused for majority voting, which can effectively filter out single-frame recognition jumps caused by occasional factors such as muscle tremors, sensor transient noise, or image motion blur. This voting mechanism is equivalent to introducing a smooth constraint in the time dimension at the decision-making level, which significantly improves the temporal stability and overall accuracy of recognition without increasing the complexity of front-end feature extraction. In addition, the sliding window strategy ensures that the voting results are always based on the most recent continuous movement data, avoiding response delays caused by introducing historical data with too long a time span.

[0039] The confidence threshold is set according to the standardized ankle pump exercise movement specification threshold library, which contains the reference threshold range of standard ankle pump movements at different rehabilitation stages. The reference threshold range includes at least the standard ankle dorsiflexion angle threshold range, the standard plantarflexion angle threshold range, the standard movement rate threshold range, the single standard movement frequency threshold range, and the standard force distribution reference interval of the foot and ankle.

[0040] In this embodiment, the confidence score is directly correlated with the degree of matching between the patient's real-time motion characteristics and clinical standard thresholds. The closer the patient's motion characteristics match the standard range of the corresponding stage in the standardized threshold library, the higher the confidence score output by the recognition model. Conversely, when the motion characteristics deviate further from the standard range or when there are cross-stage parameter contradictions, the confidence score automatically decreases. This correlation makes the confidence score naturally clinically interpretable. Medical staff can intuitively understand the degree of deviation between the current motion quality and the standard action by looking at the confidence score, without the need for additional interpretation of the model's internal decision-making logic. This significantly improves the clinical usability of the system and the acceptance of medical staff.

[0041] The ankle pump motion recognition model has a built-in AI adaptive control algorithm, which generates corresponding motion control commands. The AI ​​adaptive control algorithm includes: The patient's basic vital signs data and real-time motion state recognition results are used to construct a design feature vector, which is then input into a pre-trained adaptive control network model. The adaptive regulation network model adopts a multilayer perceptron structure, and its training samples include exercise training data and corresponding effective training parameters of patients of different ages, weights, injury degrees and rehabilitation stages. The adaptive control network model takes maximizing training effectiveness and minimizing compensation risk as its objective function, and outputs a personalized ankle pump exercise training program that matches the patient's current state, including the movement amplitude threshold, training frequency, single training duration, and rest interval duration. The training program is dynamically updated as the patient's rehabilitation progresses and their real-time movement status changes.

[0042] In this embodiment, an AI adaptive control algorithm is used to combine basic vital signs data such as patient age, weight, ankle joint injury severity, and rehabilitation stage with real-time motion state recognition results. These data are then input into a pre-trained multilayer perceptron adaptive control network model. With the objective function of maximizing training effectiveness and minimizing compensatory risk, the model dynamically outputs personalized motion amplitude thresholds, training frequency, single training duration, and rest intervals that match the patient's current state. This fully considers individual differences among patients in terms of injury severity, tolerance, and rehabilitation progress. The model can dynamically iterate and update as the patient's rehabilitation progresses and real-time motion state changes, effectively avoiding the risk of secondary injury due to excessive training intensity or poor rehabilitation results due to insufficient training intensity. This fundamentally improves patient training compliance and rehabilitation outcomes.

[0043] In this embodiment, the AI ​​adaptive regulation algorithm relies on an adaptive regulation network model, and the specific processing flow is as follows: The adaptive control network model is a multilayer perceptron structure, which includes a 24-dimensional input layer, three hidden layers with dimensions of 64, 32, and 16 respectively, and a 4-dimensional output layer. Each hidden layer uses the ReLU activation function to introduce nonlinearity, and a Dropout layer is set after each hidden layer with a dropout rate of 0.2 to prevent overfitting. The output layer uses a linear activation function to directly output the prediction results of four continuous values. In this embodiment, the adaptive control network model is input with the following characteristics: 65 years old, 70 kg, moderate injury, acute phase, intermediate risk of thrombosis, current insufficient amplitude compensation, confidence level 0.78, moderate severity, and a target achievement rate of 72% in the last 3 training sessions. The network inference process is as follows: After standardization, the 24-dimensional input features are propagated forward layer by layer, and the output layer obtains Y_output=[20.5,12.3,2.8,2.1], which correspond to a dorsiflexion target angle of 20.5°, a training frequency of 12 times per group, a single training session duration of 2.8 minutes, and an interval rest duration of 2.1 minutes, respectively. The continuous predicted values ​​output by the network are decoded and combined with preset clinical safety boundary constraints to generate a final executable personalized ankle pump exercise training program.

[0044] The AI ​​adaptive control algorithm generates corresponding motion control commands based on the action category recognition results, and is equipped with a closed-loop intelligent error correction intervention mechanism, which includes: The real-time motion state recognition results are compared item by item with the standard parameters set in the personalized ankle pump exercise training program to calculate the deviation value between the actual motion state and the standard program. When the deviation value exceeds the preset tolerance range, it is determined to be a deviation condition, and a graded error correction prompt instruction is generated according to the deviation type and degree. The hierarchical error correction prompts include at least a first-level voice prompt error correction instruction and a second-level visual interface error correction instruction. The first-level voice prompt correction instruction is used to provide specific corrective action guidance via voice broadcast when a slight deviation in the operating condition is detected; The secondary visual interface error correction command is used to simultaneously display an animated illustration of the deviation location and a demonstration of the correct action on the interactive interface when a moderate or higher level of deviation is detected. Continuously collect multimodal motion data after error correction intervention and repeat the above steps until the deviation value falls back to the preset tolerance range.

[0045] In this embodiment, the real-time motion state recognition results are compared item by item with the standard parameters in the personalized training plan. When a deviation is detected, the first-level voice prompt error correction command and the second-level visual interface error correction command are triggered sequentially according to the type and degree of deviation. This realizes the functional leap from simple motion monitoring to active guidance and correction. By continuously collecting multimodal motion data after error correction intervention and re-executing recognition and comparison until the deviation value falls back to the preset tolerance range, a complete closed-loop control link is formed. This effectively solves the pain points of "only monitoring without intervention" or "only reminding without verification" in the prior art, and ensures the standardization and effectiveness of ankle pump training.

[0046] Based on the action category recognition results, the patient's rehabilitation effect is dynamically monitored, and the ankle pump movement state is brought closer to the standard movement pattern through the control commands, including: Record the total number of movements, the number of standard movements, and the number of deviation movements in a single training session for the patient, and calculate the movement attainment rate; Record the ratio of the actual training time of a single patient session to the training time set in the protocol, and calculate training compliance; The trends of the patients’ achievement rate and range of motion during training over several consecutive days were statistically analyzed, and the progress of ankle joint function rehabilitation was assessed based on these trends. When the rate of achieving the target for the movement within a set number of consecutive days shows an upward trend and the range of movement gradually approaches the standard threshold range, a good rehabilitation progress assessment result is output. When the rate of achieving the target for the exercise shows a downward trend or stagnates within a set number of consecutive days, the system outputs a slow recovery progress assessment result and triggers a process to readjust the training program.

[0047] During the dynamic monitoring of the patient's rehabilitation effect, the ankle pump movement status is indicated by a risk warning. The risk warning method includes: The relevant feature parameters reflecting the blood circulation status of the lower limbs are extracted from real-time motion data. The relevant feature parameters include at least the peak angular velocity of ankle pump movement, motion frequency stability index and foot and ankle force recovery curve slope. The associated feature parameters are input into a preset risk assessment model, which has built-in risk thresholds for lower extremity deep vein thrombosis and limb swelling. When the associated feature parameters exceed the risk discrimination threshold, it is determined that there is a risk of complications such as deep vein thrombosis or limb swelling in the lower extremities. An early warning signal containing the risk level and risk type is output and the early warning signal is pushed to the medical terminal simultaneously.

[0048] In this embodiment, the total number of movements, the number of standard movements, and the number of deviation movements in a single training session are recorded in full during the training process. The movement attainment rate and training compliance are calculated in real time. The progress of ankle joint function rehabilitation is evaluated through data trend analysis over multiple days. When slow progress is detected, the training program is automatically readjusted. At the same time, by extracting blood flow circulation correlation features such as peak angular velocity, movement frequency stability index, and foot and ankle force recovery curve slope from the motion state data, the data are input into the risk assessment model to intelligently predict the risk of complications such as deep vein thrombosis and limb swelling in the lower extremities. This achieves integrated management of rehabilitation training and complication prevention, providing a data-driven and traceable decision support tool for clinical nursing.

[0049] This invention utilizes multimodal synchronous acquisition of ankle pump motion data via angle sensors, pressure sensors, posture sensors, and a miniature vision camera. Combined with deep spatiotemporal fusion and cross-modal cross-correlation feature calculation, it comprehensively captures motion angle, force distribution, spatial pose, and motion trajectory information, significantly improving the accuracy of compensatory motion recognition. A standard-compensation dual-path fine-grained discrimination model and a compensatory motion feature library are constructed, enabling accurate classification and recognition of four types of compensation: ankle-foot compensation, hip joint compensation, trunk compensation, and insufficient range of motion. Furthermore, an AI adaptive control algorithm integrates patient basic vital signs, real-time motion status, and historical training data to dynamically generate personalized training programs, overcoming the limitations of traditional methods. The limitations of fixed training modes are overcome; the closed-loop error correction intervention mechanism upgrades passive monitoring to active intervention, and the three-level confidence level differential processing and incremental learning mechanism continuously optimize model performance; at the same time, training data is fully recorded and uploaded to the cloud to generate rehabilitation reports, realizing digital and traceable management of rehabilitation nursing, effectively reducing nursing labor costs, and realizing the technological leap of ankle pump rehabilitation training from experience to intelligence, and from standardization to personalization. It has significant clinical application value in improving training compliance and effectiveness, reducing the risk of complications of deep vein thrombosis in the lower extremities, and promoting ankle joint functional rehabilitation. Moreover, the system has a simple structure, is suitable for various bedridden departments, and has good prospects for industrialization and promotion.

[0050] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for ankle pump exercise state recognition based on multi-modal data acquisition, characterized in that, Includes the following steps: The sensor on the ankle wearing component synchronously collects multi-source sensor data of the patient during the ankle pump exercise. The multi-source sensor data is preprocessed and motion feature parameters are extracted from each modal data. The motion feature parameters include at least the angle parameter, frequency parameter, amplitude parameter, pressure distribution parameter and motion trajectory parameter of the ankle pump exercise. The motion feature parameters extracted from each modality are fused in the spatiotemporal domain to construct a joint motion feature vector that reflects the full-dimensional information of ankle pump movement. The joint motion feature vector is then input into a pre-constructed ankle pump movement recognition model, which is trained based on standard ankle pump movement samples and compensatory error movement samples and has a built-in standard movement judgment threshold and compensatory movement feature library. The ankle pump motion recognition model outputs the action category recognition result of the current ankle pump motion by comparing real-time motion features with the judgment threshold and feature library. The recognition result includes at least the classification judgment of standard ankle pump motion and compensatory error motion. Based on the action category recognition result, a corresponding motion control command is generated. The control command includes at least one or more of the motion resistance parameter, motion speed parameter and motion stroke parameter output by the ankle pump motion actuator, so as to make the ankle pump motion state approach the standard motion mode. Record multi-source sensor data, motion feature parameters, action recognition results and control parameters at each stage of the training process, generate training record data, and output the training record data to external devices or cloud platforms.

2. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 1, characterized in that, The multi-source sensing data includes: The multi-source sensing data includes at least angle sensing data reflecting changes in ankle joint angle, pressure sensing data reflecting the force state of the foot and ankle, posture sensing data reflecting the spatial orientation of the foot and ankle, and image sensing data reflecting the movement trajectory of the ankle joint. The multi-source sensor data is time-stamped and aligned so that the data of each mode are under the same time reference. The angle sensing data is continuously collected by an angle sensor at a first sampling frequency during the real-time angle values ​​of the ankle joint during dorsiflexion and plantarflexion. The pressure sensing data is obtained by multiple pressure sensing units distributed on the inner side of the ankle-feet wearing component, which collect the dynamic pressure distribution value of the contact surface between the ankle and the wearing component at a second sampling frequency. The attitude sensing data is collected by an attitude sensor, which captures the acceleration and angular velocity values ​​of the ankle in three-dimensional space. The image sensing data is a sequence of continuous image frames acquired by a miniature vision camera during ankle joint movement. The raw data for each modality were subjected to sliding window filtering for noise reduction, baseline drift correction, and outlier removal.

3. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 2, characterized in that, The methods for extracting and fusing motion feature parameters include: The peak dorsiflexion angle, peak plantarflexion angle, and duration of a single movement cycle are extracted from the angle sensing data. Extract the pressure peak value and pressure center of gravity offset trajectory of each area of ​​the foot and ankle from the pressure sensing data; Extract the peak acceleration and angular velocity variation curve of the ankle movement from the posture sensing data; Extract the pixel coordinate sequence of key points of the ankle joint from the image sensing data and convert it into a spatial motion trajectory; The motion feature parameters extracted from each modality are aligned along the time axis according to a unified timestamp to form a time series feature set; The time series feature set is segmented using a sliding time window with a preset duration and a preset step size, and the features of multiple time nodes within each time window are combined into a time series feature matrix. Calculate the cross-correlation features between different modal features, wherein the cross-correlation features include at least angle-pressure temporal correlation features, attitude-trajectory spatial consistency features, and angle-angular velocity phase synchronization features; The original features are combined with the cross-correlation features to generate a joint motion feature vector. The joint motion feature vector is then normalized and dimensionality-reduced before being input into the ankle pump motion recognition model.

4. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 1, characterized in that, The method for constructing the ankle pump motion recognition model includes: Several clinical standard ankle pump exercise data and several compensatory error exercise data were collected as training sample sets. The compensatory error exercises were classified and labeled according to the compensation type. The compensation type includes at least four categories: foot and ankle compensation, hip joint compensation, trunk compensation, and insufficient range of motion. Each type of compensation sample was labeled with the corresponding compensation category label. The standard ankle pump exercise samples were labeled with the standard action label. A joint motion feature vector is extracted from each training sample, and the corresponding key feature discrimination parameters and feature threshold ranges are extracted for each type of compensation to form a compensation action feature library. The compensation action feature library contains the mapping relationship between each compensation type and the key feature discrimination parameters. Using the labeled joint motion feature vector as input and the action category label as output, a classification and recognition network model is trained to obtain the initial ankle pump motion recognition model.

5. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 1, characterized in that, The recognition process of the ankle pump motion recognition model includes: The joint motion feature vector generated by real-time acquisition and fusion is input into the ankle pump motion recognition model. The ankle pump motion recognition model first compares the real-time motion features with the built-in standard action judgment threshold, and outputs a binary classification judgment result on whether the current motion is a standard ankle pump action. If it is determined to be a non-standard action, the real-time motion features are further matched with the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library, and the compensation type identification result of the current motion is output. The ankle pump motion recognition model calculates the confidence score corresponding to the action category recognition result while outputting the action category recognition result; When the confidence score is higher than the preset high confidence threshold, the final recognition result is directly output; When the confidence score is between a preset low confidence threshold and a high confidence threshold, a voting mechanism for recognition results of multiple consecutive frames within a sliding time window is triggered, and the voting result is used as the final recognition output. When the confidence score is lower than the preset low confidence threshold, the movement is judged as a suspicious action, triggering a re-acquisition and recognition process; The marked suspicious action data and the misjudged data confirmed by medical staff are stored in the sample set. The ankle pump movement recognition model is periodically updated through incremental learning to update the feature discrimination rules and feature threshold ranges of each compensation type in the compensation action feature library.

6. The multi-modal data acquisition based ankle pump exercise status recognition method of claim 5, wherein, include: The confidence threshold is set according to the standardized ankle pump exercise movement specification threshold library, which contains the reference threshold range of standard ankle pump movements at different rehabilitation stages. The reference threshold range includes at least the standard ankle dorsiflexion angle threshold range, the standard plantarflexion angle threshold range, the standard movement rate threshold range, the single standard movement frequency threshold range, and the standard force distribution reference interval of the foot and ankle.

7. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 6, characterized in that, The ankle pump motion recognition model has a built-in AI adaptive control algorithm, which generates corresponding motion control commands. The AI ​​adaptive control algorithm includes: The patient's basic vital signs data and real-time motion state recognition results are used to construct a design feature vector, which is then input into a pre-trained adaptive control network model. The adaptive regulation network model adopts a multilayer perceptron structure, and its training samples include exercise training data and corresponding effective training parameters of patients of different ages, weights, injury degrees and rehabilitation stages. The adaptive control network model takes maximizing training effectiveness and minimizing compensation risk as its objective function, and outputs a personalized ankle pump exercise training program that matches the patient's current state, including the movement amplitude threshold, training frequency, single training duration, and rest interval duration. The training program is dynamically updated as the patient's rehabilitation progresses and their real-time movement status changes.

8. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 7, characterized in that, The AI ​​adaptive control algorithm generates corresponding motion control commands based on the action category recognition results, and is equipped with a closed-loop intelligent error correction intervention mechanism, which includes: The real-time motion state recognition results are compared item by item with the standard parameters set in the personalized ankle pump exercise training program to calculate the deviation value between the actual motion state and the standard program. When the deviation value exceeds the preset tolerance range, it is determined to be a deviation condition, and a graded error correction prompt instruction is generated according to the deviation type and degree. The hierarchical error correction prompts include at least a first-level voice prompt error correction instruction and a second-level visual interface error correction instruction. The first-level voice prompt correction instruction is used to provide specific corrective action guidance via voice broadcast when a slight deviation in the operating condition is detected; The secondary visual interface error correction command is used to simultaneously display an animated illustration of the deviation location and a demonstration of the correct action on the interactive interface when a moderate or higher level of deviation is detected. Continuously collect multimodal motion data after error correction intervention and repeat the above steps until the deviation value falls back to the preset tolerance range.

9. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 1, characterized in that, Based on the action category recognition results, the patient's rehabilitation effect is dynamically monitored, and the ankle pump movement state is brought closer to the standard movement pattern through the control commands, including: Record the total number of movements, the number of standard movements, and the number of deviation movements in a single training session for the patient, and calculate the movement attainment rate; Record the ratio of the actual training time of a single patient session to the training time set in the protocol, and calculate training compliance; The trends of the patients’ achievement rate and range of motion during training over several consecutive days were statistically analyzed, and the progress of ankle joint function rehabilitation was assessed based on these trends. When the rate of achieving the target for the movement within a set number of consecutive days shows an upward trend and the range of movement gradually approaches the standard threshold range, a good rehabilitation progress assessment result is output. When the rate of achieving the target for the exercise shows a downward trend or stagnates within a set number of consecutive days, the system outputs a slow recovery progress assessment result and triggers a process to readjust the training program.

10. The ankle pump motion state recognition method based on multimodal data acquisition according to claim 9, characterized in that, During the dynamic monitoring of the patient's rehabilitation effect, the ankle pump movement status is indicated by a risk warning. The risk warning method includes: The relevant feature parameters reflecting the blood circulation status of the lower limbs are extracted from real-time motion data. The relevant feature parameters include at least the peak angular velocity of ankle pump movement, motion frequency stability index and foot and ankle force recovery curve slope. The associated feature parameters are input into a preset risk assessment model, which has built-in risk thresholds for lower extremity deep vein thrombosis and limb swelling. When the associated feature parameters exceed the risk discrimination threshold, it is determined that there is a risk of complications such as deep vein thrombosis or limb swelling in the lower extremities. An early warning signal containing the risk level and risk type is output and the early warning signal is pushed to the medical terminal simultaneously.

Images