A real-time analysis method and system for lower limb motion data used in rehabilitation guidance

Abstract

The present invention discloses a real-time analysis method and system for lower limb motion data applied to rehabilitation guidance. The method calls a personalized model based on user information, converts real-time motion data into historical motion feature vectors, inputs them into a real-time prediction module, and generates multiple frames of future motion feature vectors through encoding and decoding. The deviation value of each future motion feature vector from a preset threshold range is calculated to obtain a motion deviation value, and the motion risk ratio of each future motion feature vector is calculated. The motion risk probability is obtained through weighted fusion, and the device parameter adjustment value and posture adjustment information are generated through reverse mapping. Finally, the training stage is advanced based on the risk probability assessment, and correction information is synchronously displayed and the device operating parameters are adjusted. This method constructs a dynamic model through pre-training and combines it with a lightweight LSTM to achieve real-time prediction of future motion features, thereby improving the predictability and personalized adaptation capabilities of rehabilitation guidance.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a real-time analysis method and system for lower limb motion data applied to rehabilitation guidance. Background Art

[0002] Traditional methods rely on real-time data to detect deviations and are unable to predict potential risks in motion trajectories, resulting in a time lag between equipment parameter adjustment and posture correction, causing delayed risk warnings and a lack of foresight in corrective measures. Summary of the Invention

[0003] In view of the above problems, the present invention provides a real-time analysis method and system for lower limb motion data applied to rehabilitation guidance, which solves the problem of delayed risk warning in traditional rehabilitation training.

[0004] To achieve the above objectives, in a first aspect, the present application provides a real-time analysis method for lower limb motion data for rehabilitation guidance, comprising:

[0005] Obtain user information and call the personalized model corresponding to the current user based on the user information. The personalized model is configured as a group-individual adversarial pre-training model and includes a real-time prediction module.

[0006] The current user's real-time motion data is obtained at a preset frequency and input into the personalized model to obtain the user's motion risk probability and correction information. The correction information includes device parameter adjustment values ​​and posture adjustment information, including:

[0007] The real-time motion data is converted into historical motion feature vectors of multiple frames arranged in chronological order and input into the real-time prediction module, which is configured to be built based on a lightweight LSTM network model;

[0008] The real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of multiple frames to obtain the future motion feature vectors of multiple frames arranged in chronological order;

[0009] Obtaining a preset threshold range, calculating the deviation value between each future motion feature vector and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order;

[0010] The motion risk ratio of each future motion feature vector is calculated based on the motion deviation value, and multiple motion risk ratios are weightedly fused to obtain the motion risk probability;

[0011] and, reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user;

[0012] Based on the risk probability of the movement, it is determined whether to execute the next stage of the training step, and the correction information is displayed and the corresponding equipment operating parameters are adjusted.

[0013] Furthermore, the motion risk ratio of each future motion feature vector is calculated based on the motion deviation value, and multiple motion risk ratios are weighted and fused to obtain the motion risk probability, including:

[0014] Obtaining a preset joint weight distribution table, which stores weight coefficients for each lower limb joint at different rehabilitation stages. The weight coefficients are positively correlated with the importance of the lower limb joints in rehabilitation training;

[0015] According to the user's current rehabilitation stage, the corresponding weight coefficient set is extracted from the joint weight distribution table;

[0016] For each future motion feature vector, perform matrix multiplication on the motion deviation components of each lower limb joint contained in it and the corresponding weight coefficient to obtain the joint weighted deviation value after dimension compression;

[0017] The joint weighted deviation values ​​are input into the rehabilitation risk quantification model for nonlinear normalization processing, and the motion risk ratio of each future motion feature vector is output. The rehabilitation risk quantification model is obtained through the following steps:

[0018] Obtain a preset age-parameter mapping table and call a basic parameter set of the sigmoid function that matches the user's age;

[0019] Query the stage-slope comparison table according to the current rehabilitation stage to obtain the corresponding slope adjustment coefficient;

[0020] The basic parameter group and the slope adjustment coefficient are fused through a one-dimensional convolutional layer to generate a user-adaptive normalization function, namely the rehabilitation risk quantification model;

[0021] The sliding time window method is used to perform temporal fusion of the motion risk proportions of multiple frames to obtain the motion risk probability.

[0022] Furthermore, the motion deviation value is reversely mapped to the device parameter matrix to obtain multiple device parameter adjustment values ​​including:

[0023] Build a device parameter mapping relationship library, including:

[0024] Obtain the preset rehabilitation stage-power response curve and establish a nonlinear mapping relationship between the motion deviation value and the device power adjustment amount at different rehabilitation stages;

[0025] Configure the motion mode-speed comparison matrix to define the reference speed and allowable adjustment range corresponding to each phase of the gait cycle;

[0026] Loading a safety constraint parameter table to store the damping adjustment threshold boundary based on the user's physiological characteristics;

[0027] Decompose the motion deviation value according to the device parameter mapping relationship library, including:

[0028] According to the current rehabilitation stage index to the rehabilitation stage-power response curve, the time domain integral of the motion deviation value is input into the power conversion module, and the power adjustment amount adapted to the stage is output;

[0029] Analyze the frequency domain characteristic components of the motion deviation value and match the motion mode-speed comparison matrix to obtain the reference speed correction coefficient;

[0030] The spatial distribution characteristics of the motion deviation value and the user's physiological parameters are input into the damping regulator, and the compliant damping adjustment amount is generated in combination with the safety constraint parameter table;

[0031] Output device parameter adjustment values, which include power adjustment amount, reference speed correction coefficient, and compliance damping adjustment amount.

[0032] Furthermore, the motion deviation value is reversely mapped to the user posture matrix to obtain the current user posture adjustment information including:

[0033] Construct the user pose matrix, including:

[0034] Obtain a preset joint-posture mapping table and establish the corresponding relationship between the motion deviation components of each lower limb joint and the ideal posture parameters;

[0035] Configure the rehabilitation stage-correction intensity comparison table to define the maximum posture adjustment range allowed at different rehabilitation stages;

[0036] Load the user's physiological-posture constraint table and store the personalized posture safety range based on the user's height and weight;

[0037] Decompose the motion deviation value according to the user pose matrix, including:

[0038] Decompose the motion deviation value into each joint motion plane, and obtain the initial posture correction value through the joint-posture mapping table;

[0039] According to the current rehabilitation stage, query the rehabilitation stage-correction intensity comparison table, perform intensity standardization on the initial posture correction amount, and obtain the correction intensity coefficient;

[0040] The correction intensity coefficient, initial posture correction amount and user physiological parameters are input into the posture optimizer, and the safe posture adjustment instruction and the final posture correction amount are generated in combination with the user physiological-posture constraint table;

[0041] Output posture adjustment information, which includes correction strength coefficient, initial posture correction amount, final posture correction amount and safe posture adjustment instruction.

[0042] Furthermore, the preset threshold range is configured to be initialized by the following steps during the user's first training:

[0043] Obtaining user information, and matching a first preset number of group samples in a database according to the user information, where matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity;

[0044] Obtain a basic adversarial model and iteratively train it using group samples, including:

[0045] The projected gradient descent method is used to generate disturbance samples that meet the preset disturbance range;

[0046] Repair the disturbed sample according to the preset constraint conditions, which are biomechanical constraint conditions;

[0047] The population samples and the perturbation samples are mixed and input into the basic adversarial model and iterative training is performed until a first preset training accuracy threshold is met, indicating that the basic adversarial model training is completed;

[0048] Input the group samples into the trained basic adversarial model and output the sample threshold range for each group sample;

[0049] Arrange multiple sample threshold ranges in order of the degree of matching between group samples and user information, and select the sample threshold range with the highest matching degree as the preset threshold range.

[0050] Furthermore, the personalized model also includes a threshold adjustment module, and the method further includes:

[0051] When the user is not training for the first time, the preset threshold range is configured to be updated through the threshold adjustment module, including:

[0052] After the user finishes exercising, obtain the user's current batch of real-time exercise data, record it as the latest exercise data, and obtain the historical exercise data of each batch within a preset historical period;

[0053] Generate personal adversarial samples based on the latest and historical motion data;

[0054] Acquire the latest user information of the user, and match a second preset number of latest group samples in the database according to the latest user information, where the matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity;

[0055] The latest group sample and the individual adversarial sample are mixed and input into the threshold adjustment module and iteratively trained until the second preset training accuracy threshold is met, indicating that the threshold adjustment module is updated;

[0056] The latest motion data is input into the updated threshold adjustment module to obtain the updated preset threshold range.

[0057] Furthermore, based on the latest motion data and historical motion data within a preset historical period, generating personal adversarial samples includes:

[0058] Extracting significant feature information from the latest motion data, the significant feature information includes the first motion frame with the highest motion completion, the second motion frame with the largest motion deviation, and the third motion frame with clinical typicality;

[0059] In the dormant state, semantic feature reconstruction is performed on the salient feature information, including:

[0060] The pre-set diffusion model is used to decouple the significant feature information and obtain the kinematic parameters, which include joint angles, joint torque distribution, motion velocity curves, and muscle force patterns.

[0061] Construct physical constraint information and personal constraint information based on motion mechanics parameters;

[0062] Generate variant samples similar to salient feature information based on historical movement data, physical constraint information, and personal constraint information;

[0063] The mutated samples are inserted into the historical motion data according to the execution sequence of the first action frame, the second action frame, and the third action frame in the latest action data to obtain personal adversarial samples.

[0064] Furthermore, based on the latest motion data and historical motion data within a preset historical period, generating personal adversarial samples includes:

[0065] Extracting first abnormal feature information from the latest motion data, the first abnormal feature information including joint motion trajectory deviation features, muscle activation abnormality features, and motion timing disorder features;

[0066] Mapping the first anomaly feature information to the federated shared latent space for similarity matching includes:

[0067] Calculate the latent space projection similarity with the second abnormal feature information of other users in the database;

[0068] Filter matching cases whose latent space projection similarity exceeds the preset similarity threshold, and obtain the desensitization correction strategy associated with the matching cases;

[0069] The first abnormal feature information is fused and reconstructed according to the desensitization correction strategy to generate a resonance adversarial sample after the first abnormal feature information is corrected;

[0070] The generated resonance adversarial sample is mixed with the historical motion data and the latest motion data to obtain the personal adversarial sample.

[0071] Furthermore, based on the latest motion data and historical motion data within a preset historical period, generating personal adversarial samples includes:

[0072] Extracting first motion feature information from the latest motion data, the first motion feature information including the user's habitual motion pattern features, typical incorrect motion features, and biomechanically reasonable motion features;

[0073] generating second motion characteristic information based on the biomechanical constraint condition, wherein the second motion characteristic information is configured to meet clinical standards but has a preset difference from the first motion characteristic information;

[0074] Add an explainability label to each second action feature information, where the explainability label contains specific quantitative indicators required for action improvement;

[0075] Inputting the first motion feature information and the second motion feature information into the personalized model respectively, obtaining a first risk assessment result corresponding to the first motion feature information and a second risk assessment result corresponding to the second motion feature information;

[0076] generating an acceptance index based on the second risk assessment result, and generating a stubbornness index based on the first risk assessment result;

[0077] The second action feature information is filtered according to the acceptance index and stubbornness index, and recorded as a valid adversarial sample. It is then mixed with the historical motion data and the latest motion data to obtain a personal adversarial sample.

[0078] In a second aspect, the present invention further provides a real-time analysis system for lower limb motion data used in rehabilitation guidance, which is applicable to the real-time analysis method of the first aspect, and the system includes a data acquisition module and a logic processing module;

[0079] The data acquisition module is used to obtain user information. The data acquisition module is also used to obtain the real-time motion data of the current user according to a preset frequency;

[0080] The logic processing module is used to call the personalized model corresponding to the current user based on user information. The personalized model is configured as a group adversarial pre-training model and includes a real-time prediction module.

[0081] The logic processing module is also used to input real-time motion data into the personalized model to obtain the user's motion risk probability and correction information. The correction information includes device parameter adjustment values ​​and posture adjustment information, including:

[0082] The real-time motion data is converted into historical motion feature vectors of multiple frames arranged in chronological order and input into the real-time prediction module, which is configured to be built based on a lightweight LSTM network model;

[0083] The real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of multiple frames to obtain the future motion feature vectors of multiple frames arranged in chronological order;

[0084] Obtaining a preset threshold range, calculating the deviation value between each future motion feature vector and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order;

[0085] The motion risk ratio of each future motion feature vector is calculated based on the motion deviation value, and multiple motion risk ratios are weightedly fused to obtain the motion risk probability;

[0086] and, reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user;

[0087] Based on the risk probability of the movement, it is determined whether to execute the next stage of the training step, and the correction information is displayed and the corresponding equipment operating parameters are adjusted.

[0088] Different from the existing technology, the above technical solution provides a real-time analysis method and system for lower limb motion data applied to rehabilitation guidance. The method calls a personalized model based on user information, converts real-time motion data into historical motion feature vectors and inputs them into a real-time prediction module. After encoding and decoding, multiple frames of future motion feature vectors are generated. The deviation value of each future motion feature vector from the preset threshold range is calculated to obtain the motion deviation value, and the motion risk ratio of each future motion feature vector is calculated. The motion risk probability is obtained through weighted fusion, and the device parameter adjustment value and posture adjustment information are generated through reverse mapping. Finally, the training stage is advanced based on the risk probability assessment, and correction information is displayed synchronously and the device operating parameters are adjusted. This technical solution constructs a dynamic model through pre-training, and combines lightweight LSTM to achieve real-time prediction of future motion characteristics, thereby improving the predictability and personalized adaptation capabilities of rehabilitation guidance.

[0089] The above-mentioned records related to the content of the invention are only an overview of the technical solution of this application. In order to enable ordinary technicians in this field to understand the technical solution of this application more clearly, and then implement it according to the text of the specification and the contents recorded in the drawings, and to make the above-mentioned purposes and other purposes, features and advantages of this application easier to understand, the following is an explanation in combination with the specific implementation methods and drawings of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The drawings are only used to illustrate the principles, implementation methods, applications, features and effects of the specific embodiments of the present invention and other related contents, and are not to be considered as limiting the present application.

[0091] In the drawings of the specification:

[0092] Figure 1 A method step diagram of steps S101 to S107 of the real-time analysis method described in the specific embodiment;

[0093] Figure 2 A method step diagram of steps S201 to S205 of the real-time analysis method described in the specific embodiment;

[0094] Figure 3 This is a method step diagram of steps S301 to S303 of the real-time analysis method described in the specific embodiment;

[0095] Figure 4 This is a method step diagram of steps S401 to S403 of the real-time analysis method described in the specific embodiment;

[0096] Figure 5 Schematic diagram of the structure of the real-time analysis system described in the specific implementation method.

[0097] The reference numerals in the above drawings are described as follows:

[0098] 1. Real-time analysis system;

[0099] 11. Data acquisition module;

[0100] 12. Logic processing module. DETAILED DESCRIPTION

[0101] In order to explain in detail the possible application scenarios, technical principles, specific solutions that can be implemented, and the purpose and effects of this application, the following is a detailed description of the specific embodiments listed in conjunction with the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this application and are therefore only examples and are not intended to limit the scope of protection of this application.

[0102] References to "embodiments" herein mean that the specific features, structures, or characteristics described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the word "embodiment" in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or relevance to other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the various technical features mentioned in the embodiments can be combined in any manner to form a corresponding implementable technical solution.

[0103] Unless otherwise defined, the technical terms used herein have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms herein is only for describing specific embodiments and is not intended to limit this application.

[0104] In the description of this application, the term "and / or" is used to describe a logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and both A and B exist. In addition, the character " / " in this document generally indicates that the objects before and after are in a logical "or" relationship.

[0105] In this application, terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, priority or sequence relationship between these entities or operations.

[0106] Without further limitations, in this application, the words "include", "comprise", "have" or other similar open-ended expressions used in sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product that includes the elements, so that the process, method or product that includes a series of elements may include not only those defined elements, but also other elements that are not explicitly listed, or also include elements inherent to such process, method or product.

[0107] Consistent with the understanding in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceed" are understood to exclude the number itself; expressions such as "above," "below," and "within" are understood to include the number itself. Furthermore, in the description of the embodiments of this application, "multiple" means two or more (including two), and similar expressions related to "multiple," such as "multiple groups" and "multiple times," are also understood in this manner, unless otherwise specifically defined.

[0108] In the description of the embodiments of the present application, the space-related expressions used, such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicate the orientation or position relationship based on the orientation or position relationship shown in the specific embodiments or drawings, and are only for the convenience of describing the specific embodiments of the present application or facilitating the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, it should not be understood as a limitation on the embodiments of the present application.

[0109] The processor described in the embodiments of the present application can be implemented by hardware, firmware, software or a combination thereof, and can use a circuit, a single or multiple application-specific integrated circuits (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, at least one of a microprocessor, and also includes other physical, biological or chemical structures that can achieve similar or equivalent functions to the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of the present application, or any combination of the steps mentioned therein.

[0110] The computer program involved in the embodiment can be stored in a computer device readable storage medium, which includes but is not limited to a disk, a tape, a magnetic card, a floppy disk, a flash memory, an optical disc, an optical card, a read-only memory (ROM), a random access memory (RAM), an erasable programmable ROM (EPROM) and an electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical or chemical structures that can achieve similar or equivalent functions to the storage media listed above, such as DNA, RNA, protein and other units with information storage capabilities. In a specific embodiment, the storage medium involved can be one of the above-mentioned media types or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiment can be stored in a single medium in a centralized manner or in a distributed manner on multiple media. The memory containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built into the device or connected to the device involved in the embodiment as an external device or part of an external device. In some embodiments, a memory having a computer-readable storage medium is deployed locally; in other embodiments, a solution of deploying the memory away from the processor may also be adopted, such as a network-attached memory accessed via an RF circuit or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), etc., or a suitable combination thereof, as long as the computer device can access the memory. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form, or can be designed as training data, which can be integrated and reorganized through model training and implicitly stored in the parameter state of a deep neural network or other machine learning model.

[0111] See also Figure 1 In a first aspect, this embodiment provides a real-time analysis method for lower limb motion data for rehabilitation guidance, comprising:

[0112] S101. Obtain user information and call a personalized model corresponding to the current user based on the user information. The personalized model is configured as a group-individual adversarial pre-training model and includes a real-time prediction module.

[0113] S102: Acquire the current user's real-time motion data at a preset frequency and input it into a personalized model to obtain the user's motion risk probability and correction information. The correction information includes device parameter adjustment values ​​and posture adjustment information, including:

[0114] S103, converting the real-time motion data into a multi-frame historical motion feature vector arranged in chronological order, and inputting the historical motion feature vector into a real-time prediction module, wherein the real-time prediction module is configured to be constructed based on a lightweight LSTM network model;

[0115] S104, the real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of the multiple frames to obtain future motion feature vectors of the multiple frames arranged in chronological order;

[0116] S105, obtaining a preset threshold range, calculating the deviation value between each future motion feature vector and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order;

[0117] S106, calculating the motion risk ratio of each future motion feature vector according to the motion deviation value, performing weighted fusion on multiple motion risk ratios to obtain a motion risk probability;

[0118] and, reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user;

[0119] S107: Evaluate whether to execute the next stage training step based on the motion risk probability, and display the correction information and adjust the corresponding equipment operating parameters.

[0120] In step S101, user information refers to multidimensional information including the user's physiological characteristics, injury type, and historical rehabilitation data, which is used to generate an initial personalized model through database matching. The group-individual adversarial pre-training model is an adversarial training framework based on group samples and individual data. By introducing perturbation samples and biomechanical constraints during the pre-training phase, the model possesses the dual characteristics of group experience and individual adaptation. The real-time prediction module is used to achieve efficient processing of time series features.

[0121] In step S102, the preset frequency is dynamically determined by the sampling capacity of the device sensor and the user's current rehabilitation stage. Real-time motion data includes lower limb joint angles, motion trajectory speed and acceleration indicators. Preferably, the device parameter adjustment value includes power adjustment amount, speed correction coefficient and damping adjustment amount, which are used to dynamically adjust the power output of the rehabilitation device; posture adjustment information includes joint correction amount and strength parameters, which are generated based on the user's physiological constraints to guide lower limb posture adjustment to avoid sports injuries. The device parameter adjustment value and posture adjustment information are generated by reverse mapping the motion deviation value to achieve synchronous optimization of device adaptation and user action.

[0122] In step S103, the lightweight LSTM network optimizes computational efficiency by reducing the number of hidden layer nodes and parameter pruning while retaining the ability to capture temporal features.

[0123] In step S104, the encoder extracts the spatiotemporal correlation of the historical motion feature vectors, and the decoder generates multiple frames of future motion feature vectors based on the attention mechanism.

[0124] In step S105, the preset threshold range is preferably initialized during initial training by matching similar group samples to generate an adversarial model. During non-initial training, the threshold adjustment module of the personalized model is updated in conjunction with individual historical data. The details are described below. When calculating the motion deviation value, the Euclidean distance can be used to quantify the degree of deviation of the future motion feature vector from the boundaries of the preset threshold range.

[0125] In step S106, the motion risk percentage of each future motion feature vector is calculated based on the motion deviation value. This can be understood by referring to the following example: For each future motion feature vector, the deviation components of the hip, knee, and other joints contained therein are multiplied by the weight coefficient corresponding to the rehabilitation stage (e.g., the hip joint has a higher weight in the early stages). Nonlinear compression is performed using an age-adaptive S-shaped function, and the motion risk percentage is output within the range of 0-1. Multiple motion risk percentages are weighted and fused, using the weight coefficients of the frames within the sliding time window (e.g., the weight of the proximal frame is increased by 20%). A temporal attention mechanism is then used to generate a comprehensive motion risk probability.

[0126] The process of obtaining the device parameter adjustment value can be understood as: decomposing the motion deviation value into power, speed and damping components. For example, the gait phase deviation triggers the speed reference matrix correction, and the power response curve converts the time domain deviation integral into the motor power increase. At the same time, the user's bone density constraint is combined to limit the damping adjustment amplitude.

[0127] The process of obtaining posture adjustment information can be understood as: mapping the motion deviation value to the joint posture plane. For example, the sagittal plane deviation of the knee joint is generated by a mapping table to generate a 5° flexion correction. Combined with the user's height and weight constraints, the adjustment range is limited to 3°, and finally the posture instruction with a safety margin is output.

[0128] In step S107, when the probability of motion risk is lower than the preset stage migration threshold, the next stage of training is triggered; otherwise, the current stage is maintained and parameter adjustment is performed. Furthermore, the preset stage migration threshold can be determined by integrating the group rehabilitation benchmark and the individual dynamic characteristics: an initial threshold is established based on the historical training data of patients with similar injuries (such as the peak value of the gait risk probability distribution of the fracture group of 0.25), and the user's real-time electromyographic signal attenuation rate, bone density growth curve and other physiological indicators are superimposed for offset correction (±0.1 interval), and the boundary is dynamically tightened by the standard deviation of the risk fluctuations of the last 7 training sessions. For example, after knee surgery, the user's initial threshold is set to 0.3. When the muscle strength increases by 20% and the risk probability is stable at 0.25 for three consecutive times, the threshold is lowered to 0.22 to trigger advancement, so as to avoid hindering the rehabilitation process due to a single abnormal fluctuation.

[0129] Optionally, the correction information is displayed visually through an interactive interface, and the equipment operating parameters are adjusted in real time by the control system.

[0130] This embodiment provides a real-time analysis method for lower limb motion data, enabling precise rehabilitation guidance by constructing a group-individual adversarial pre-training model. The method first invokes a personalized model based on user information. A lightweight LSTM network is used to encode historical motion feature vectors in real time and decode predicted future multi-frame motion trends (i.e., multi-frame future motion feature vectors). Motion deviation values ​​are calculated within a preset threshold range, and device parameter adjustment values ​​and posture correction information are simultaneously generated. While ensuring computational efficiency, this method achieves personalized adaptation of the rehabilitation training process by synergizing device power output with user motion posture. Firstly, the adversarial training framework balances group universality with individual variability, enhancing the robustness of motion risk prediction. Secondly, a reverse mapping mechanism is used to convert motion deviations into two-dimensional adjustments of device parameters and posture, mitigating the risk of secondary injury while accelerating functional recovery. A dynamic threshold mechanism combines historical risk fluctuations with the evolution of physiological indicators to ensure the rationality and safety of rehabilitation stage transitions, effectively avoiding misjudgments caused by local data anomalies, and forming a closed-loop optimized intelligent rehabilitation decision-making system.

[0131] See also Figure 2 In some embodiments, the motion risk ratio of each future motion feature vector is calculated based on the motion deviation value, and multiple motion risk ratios are weightedly fused to obtain the motion risk probability, including:

[0132] S201, obtaining a preset joint weight distribution table, wherein the joint weight distribution table stores weight coefficients of each lower limb joint at different rehabilitation stages, and the weight coefficients are positively correlated with the importance of the lower limb joints in rehabilitation training;

[0133] S202, extracting a corresponding weight coefficient set from a joint weight distribution table according to the user's current rehabilitation stage;

[0134] S203: For each future motion feature vector, perform a matrix multiplication operation on the motion deviation components of each lower limb joint contained therein and the corresponding weight coefficient to obtain a dimensionally compressed joint weighted deviation value;

[0135] S204: Input the joint weighted deviation value into the rehabilitation risk quantification model for nonlinear normalization processing, and output the motion risk ratio of each future motion feature vector. The rehabilitation risk quantification model is obtained by the following steps:

[0136] Obtain a preset age-parameter mapping table and call a basic parameter set of the sigmoid function that matches the user's age;

[0137] Query the stage-slope comparison table according to the current rehabilitation stage to obtain the corresponding slope adjustment coefficient;

[0138] The basic parameter group and the slope adjustment coefficient are fused through a one-dimensional convolutional layer to generate a user-adaptive normalization function, namely the rehabilitation risk quantification model;

[0139] S205 , performing temporal fusion on the motion risk proportions of multiple frames using a sliding time window method to obtain a motion risk probability.

[0140] In step S201, the formation of the joint weight distribution table is based on a comprehensive analysis of clinical rehabilitation medicine experience and historical training data. The difference in the importance of each lower limb joint at different rehabilitation stages is reflected in the gradient change of the weight coefficient: for example, in the early stage after fracture surgery, the hip joint is assigned a higher weight because it plays the main supporting function. As the rehabilitation process progresses, the weight of knee flexion and extension gradually increases. The joint weight distribution table is formed by statistically analyzing the degree to which joint motion deviations of patients with similar injuries at specific stages affect the rehabilitation effect, combined with expert experience and calibration, to ensure that the weight coefficient matches the priority of joint function recovery.

[0141] In step S202, the system determines the user's current rehabilitation stage by dynamically comparing the real-time motion risk probability with a preset stage transition threshold. If the motion risk probability remains consistently below the current stage threshold over multiple consecutive training sessions, the system determines the next stage. This step dynamically adapts joint importance assessments by querying the joint weight distribution table and extracting the set of weight coefficients that matches the current stage, providing a staged parameter benchmark for subsequent weighted calculations.

[0142] In step S203, each future motion feature vector contains independent motion deviation components of each lower limb joint, such as the hip, knee, and ankle. By multiplying with the corresponding elements of the weight coefficient set, the multidimensional deviation vector is compressed into a scalar joint weighted deviation value. In essence, it is a linear combination of the importance-weighted deviations of each joint, so that the joint deviations that have a greater impact on the rehabilitation effect occupy a higher proportion in subsequent calculations.

[0143] In step S204, the process of obtaining the motion risk ratio of each future motion feature vector can be understood as follows: the construction of the rehabilitation risk quantification model depends on the dual parameter adaptation of age and rehabilitation stage. Specifically, the age-parameter mapping table pre-stores the basic parameter group of the S-type function according to the physiological differences in the motor function recovery ability of users of different age groups, which is used to adjust the steepness of the normalized curve; the stage-slope comparison table adjusts the slope of the curve according to the rehabilitation process. For example, a flat curve is used in the early stage to avoid excessive sensitivity, and the slope is increased in the later stage to improve the risk response accuracy; the basic parameter group and the slope adjustment coefficient are fused through a one-dimensional convolution layer to generate a user-adaptive S-type function, and the joint weighted deviation value is nonlinearly mapped to the motion risk ratio in the range of 0-1, so that the risk quantification result reflects both individual physiological characteristics and rehabilitation stage characteristics.

[0144] In step S205, the time window length is matched to the device sampling frequency, and the weight coefficients of each frame within the window are adjusted according to the time decay factor, with near-end frames given higher weights to emphasize the influence of recent motion trends. A temporal attention mechanism is introduced into the fusion process to automatically identify abnormally fluctuating frames and reduce their contribution. Finally, a weighted summation is performed to output a smoothed motion risk probability. This step suppresses transient noise interference, enhances the temporal coherence of risk assessment results, and improves decision reliability, providing a stable basis for stage transition judgment.

[0145] This embodiment realizes dynamic calibration of joint importance adaptively during the rehabilitation stage through the joint weight allocation table, and completes the weighted fusion of multi-dimensional motion deviation components in combination with matrix multiplication operations, effectively strengthening the rehabilitation impact weights of key joints. The rehabilitation risk quantification model is based on the two-dimensional parameter fusion of the age-parameter mapping table and the stage-slope comparison table, constructs a user-adaptive S-type function, and nonlinearly maps the joint weighted deviation value to the motion risk ratio. The sliding time window mechanism is used in conjunction with the temporal attention algorithm to capture the motion risk trend while suppressing instantaneous noise, and realize the smooth output of risk probability. This method improves the physiological fit of risk assessment through the synergy of staged weight adaptation and personalized normalization processing; optimizes decision stability with the help of temporal fusion, ensures the scientificity and safety of rehabilitation stage migration judgment, and forms an accurate and reliable dynamic risk quantification system.

[0146] See also Figure 3 In some embodiments, the motion deviation value is reversely mapped to the device parameter matrix to obtain multiple device parameter adjustment values ​​including:

[0147] S301. Build a device parameter mapping relationship library, including:

[0148] Obtain the preset rehabilitation stage-power response curve and establish a nonlinear mapping relationship between the motion deviation value and the device power adjustment amount at different rehabilitation stages;

[0149] Configure the motion mode-speed comparison matrix to define the reference speed and allowable adjustment range corresponding to each phase of the gait cycle;

[0150] Loading a safety constraint parameter table to store the damping adjustment threshold boundary based on the user's physiological characteristics;

[0151] S302: Decomposing the motion deviation value according to the device parameter mapping relationship library, including:

[0152] According to the current rehabilitation stage index to the rehabilitation stage-power response curve, the time domain integral of the motion deviation value is input into the power conversion module, and the power adjustment amount adapted to the stage is output;

[0153] Analyze the frequency domain characteristic components of the motion deviation value and match the motion mode-speed comparison matrix to obtain the reference speed correction coefficient;

[0154] The spatial distribution characteristics of the motion deviation value and the user's physiological parameters are input into the damping regulator, and the compliant damping adjustment amount is generated in combination with the safety constraint parameter table;

[0155] S303: Outputting device parameter adjustment values, where the device parameter adjustment values ​​include a power adjustment amount, a reference speed correction coefficient, and a compliance damping adjustment amount.

[0156] In step S301, the rehabilitation stage-power response curve is established based on a coupled analysis of historical rehabilitation data and clinical expert experience. By statistically analyzing the correspondence between user motion deviation values ​​and device power adjustment amounts at different rehabilitation stages, and combining the sensitivity of injury types to power response requirements, a nonlinear mapping curve with stage characteristics is fitted. For example, a low-slope curve is used in the early stages of rehabilitation to achieve gentle power compensation, while a high-slope curve is used in the later stages to enhance response strength.

[0157] The baseline speed for the movement pattern-speed comparison matrix is ​​defined by analysis of a standard gait cycle. The gait cycle can be understood as the complete motion process of a user's unilateral limb from heel strike to ipsilateral heel strike again, encompassing stance and swing phases. The baseline speed is set based on statistically analyzed gait phase durations in healthy individuals, while the adjustable range is dynamically expanded based on the user's range of motion test results to ensure adjustments are within physiological tolerance limits.

[0158] The safety constraint parameter table integrates the user's age, bone density, and muscle strength level physiological indicators, and combines biomechanical simulation to calculate the safety threshold of the damping force of each joint to form the dynamic boundary condition of the damping adjustment amount, which is the damping adjustment amount threshold boundary.

[0159] In step S302, the power conversion module converts the nonlinear mapping relationship between the rehabilitation stage and the power response curve into a quantitative output of the power adjustment amount through the time accumulation effect of the accumulated motion deviation value in the time domain integral amount. For example, for a persistent foot drop deviation, an increase in the integral amount triggers a higher power compensation.

[0160] Preferably, the frequency domain feature components are used to extract the periodic fluctuation characteristics of the motion deviation through fast Fourier transform, match the gait phase-speed relationship preset in the motion pattern-speed comparison matrix, identify abnormal frequency components and generate a reference speed correction coefficient. If an abnormal shortening of the swing phase period is detected, the reference speed is reduced according to the reference speed correction coefficient.

[0161] The spatial distribution characteristics reflect the vector distribution of motion deviation values ​​in the hip, knee, and ankle joints. Combined with the muscle balance and joint stability indicators in the user's physiological parameters, the damping force difference required for each joint is calculated through the damping regulator, and then limited according to the safety constraint parameter table to generate a compliant damping adjustment amount that meets the user's current physiological state.

[0162] In step S303, the power adjustment factor compensates for the time-domain cumulative effects of motion deviations by driving the device's output torque; the reference speed correction factor dynamically adapts to the user's actual gait rhythm; and the compliance damping adjustment factor balances joint force distribution in the spatial dimension. These three factors act together on the device's actuators, achieving precise parameter adjustment across multiple domains, including time, frequency, and space, ensuring optimal motion trajectory tracking during rehabilitation training within physiological safety boundaries.

[0163] This embodiment constructs a device parameter mapping relationship library that includes a rehabilitation stage-power response curve, a motion mode-speed comparison matrix, and a safety constraint parameter table to achieve multi-dimensional decoupling analysis of the time domain integral, frequency domain characteristic components, and spatial distribution characteristics of the motion deviation value. Combined with the phase compensation mechanism of the power conversion module, the reference speed dynamic correction algorithm, and the biomechanical constraint strategy of the damping regulator, a coordinated output of the power adjustment amount, the reference speed correction coefficient, and the compliant damping adjustment amount is generated, effectively balancing the device response sensitivity and physiological tolerance, improving the motion trajectory tracking accuracy while ensuring the safety boundary of the joint movement, and forming an adaptive parameter control system that conforms to the characteristics of the rehabilitation stage and the individual differences of users.

[0164] See also Figure 4 In some embodiments, the motion deviation value is reversely mapped to the user posture matrix to obtain the posture adjustment information of the current user, including:

[0165] S401: Constructing a user pose matrix, including:

[0166] Obtain a preset joint-posture mapping table and establish the corresponding relationship between the motion deviation components of each lower limb joint and the ideal posture parameters;

[0167] Configure the rehabilitation stage-correction intensity comparison table to define the maximum posture adjustment range allowed at different rehabilitation stages;

[0168] Load the user's physiological-posture constraint table and store the personalized posture safety range based on the user's height and weight;

[0169] S402, decomposing the motion deviation value according to the user posture matrix, including:

[0170] Decompose the motion deviation value into each joint motion plane, and obtain the initial posture correction value through the joint-posture mapping table;

[0171] According to the current rehabilitation stage, query the rehabilitation stage-correction intensity comparison table, perform intensity standardization on the initial posture correction amount, and obtain the correction intensity coefficient;

[0172] The correction intensity coefficient, initial posture correction amount and user physiological parameters are input into the posture optimizer, and the safe posture adjustment instruction and the final posture correction amount are generated in combination with the user physiological-posture constraint table;

[0173] S403: Outputting posture adjustment information, where the posture adjustment information includes a correction strength coefficient, an initial posture correction amount, a final posture correction amount, and a safe posture adjustment instruction.

[0174] In step S401, the construction of the joint-posture mapping table is based on the coupling analysis of the lower limb kinematic model and clinical gait data. By collecting the motion angle parameters of the hip, knee, and ankle joints in the sagittal and coronal planes during the gait cycle of healthy people and combining them with the joint coordination law of the ideal gait trajectory, a quantitative mapping relationship between the joint motion deviation component (such as the knee flexion angle deviation) and the ideal posture parameter (such as the ideal flexion angle) is established.

[0175] Preferably, the definition of the maximum posture adjustment amplitude follows the staging principle of rehabilitation medicine: low amplitude limit is adopted in the early rehabilitation stage to avoid excessive stretching of damaged tissues, and the adjustment amplitude is gradually increased in the later stage to enhance the training intensity. The specific threshold is calibrated comprehensively through expert experience and biomechanical tolerance testing.

[0176] The user's physiological-posture constraint table calculates the personalized posture safety range by establishing a statistical regression model of height, weight, and lower limb joint mobility. For example, it dynamically adjusts the mechanical limit threshold of the hip joint abduction angle based on the length of the user's lower limb to ensure that the posture adjustment conforms to the human anatomical characteristics.

[0177] In step S402, the motion deviation value decomposition is achieved through multi-plane motion analysis, and the overall motion deviation vector is projected onto the sagittal plane, coronal plane and other motion planes of each joint. For example, the foot inversion deviation is decomposed into the ankle joint coronal plane inversion angle deviation component, and then the initial posture correction amount (such as 5° ankle abduction) is obtained based on the joint-posture mapping table matching.

[0178] Intensity normalization is performed by scaling the initial posture correction with the maximum allowable amplitude for the corresponding stage in the rehabilitation stage-correction intensity comparison table, generating a correction intensity coefficient in the range of 0-1. For example, if the initial correction is 8° and the maximum amplitude for the stage is 10°, the coefficient is 0.8. The posture optimizer uses a constrained gradient descent algorithm, using the correction intensity coefficient as a weighting factor to scale the initial posture correction. Simultaneously, it performs a limited iterative calculation based on the safety range boundaries in the user's physiological-posture constraint table (e.g., knee flexion does not exceed 120°). The final posture correction that meets physiological constraints is output along with the corresponding safety adjustment instructions (e.g., "adjust knee flexion correction from 10° to 8°").

[0179] In step S403, the correction intensity coefficient quantifies the adjustment force adapted during the rehabilitation stage, the initial posture correction value reflects the original correction requirement of the motion deviation value, the final posture correction value represents the actual execution value after safety constraint optimization, and the safe posture adjustment instruction transmits the action execution strategy to the device in text or coded form. The four work together to ensure that the posture adjustment gradually approaches the ideal motion trajectory within the personalized physiological safety range.

[0180] This embodiment constructs a user posture matrix including a joint-posture mapping table, a rehabilitation stage-correction intensity comparison table, and a user physiological-posture constraint table. Based on multi-plane motion analysis, the joint-level decomposition of the motion deviation value and the matching of the initial posture correction amount are achieved. The posture optimizer is combined with the user physiological-posture constraint table to generate the final posture correction amount and safety adjustment instructions. Under the premise of ensuring the safe range of personalized joint movement, the anatomically precise correction of motion deviation and progressive posture adjustment are achieved, effectively balancing the intensity of rehabilitation training and physiological tolerance, ensuring the dynamic adaptability and execution safety of the posture adjustment instructions, and promoting the user's motion trajectory to reliably approach the ideal parameters.

[0181] In some embodiments, the preset threshold range is configured to be initialized by the following steps when the user first trains:

[0182] Obtaining user information, and matching a first preset number of group samples in a database according to the user information, where matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity;

[0183] Obtain a basic adversarial model and iteratively train it using group samples, including:

[0184] The projected gradient descent method is used to generate disturbance samples that meet the preset disturbance range;

[0185] Repair the disturbed sample according to the preset constraint conditions, which are biomechanical constraint conditions;

[0186] The population samples and the perturbation samples are mixed and input into the basic adversarial model and iteratively trained until a first preset training accuracy threshold is met, indicating that the basic adversarial model training is completed;

[0187] Input the group samples into the trained basic adversarial model and output the sample threshold range for each group sample;

[0188] Arrange multiple sample threshold ranges in order of the degree of matching between group samples and user information, and select the sample threshold range with the highest matching degree as the preset threshold range.

[0189] In this embodiment, the determination of the first preset number requires a balance between database coverage and computational efficiency. This is achieved by statistically analyzing the size of the injury type group to which the user belongs and dynamically adjusting the number based on training resource constraints. For example, a representative sample size is selected based on the clustering results of injury type similarity. In the step of matching group samples based on user information, injury type similarity is preferably compared using ICF classification codes, physiological characteristic similarity is calculated using the Euclidean distance after normalization of parameters such as height, weight, and bone density, and movement abnormality similarity is evaluated based on the cosine similarity of joint angle deviation vectors within the gait cycle. The weighted scores of the three are then combined to select the nearest neighboring samples in the database.

[0190] During iterative training of the basic adversarial model, the preset perturbation range is based on clinical research data on biomechanical parameter variability. For example, the natural fluctuation range of hip flexion angle is statistically determined by multiplying the standard deviation of gait parameters in healthy individuals. When generating perturbation samples using the projected gradient descent method, a small perturbation is first applied along the gradient direction within the motion parameter space of the original population sample. The perturbed parameter vector is then projected into a credible region defined by the preset perturbation range to ensure that the perturbed sample conforms to anatomical kinematic constraints.

[0191] The repair of perturbation samples based on preset constraints is achieved through a biomechanical rule engine. For example, when the perturbation causes the knee flexion angle to exceed the physiological limit, a quadratic programming algorithm is used to adjust the angles of adjacent joints to maintain the closure of the lower limb kinematic chain, or a feasible posture is recalculated based on inverse kinematics.

[0192] After the population samples and the restored perturbation samples are mixed and input into the basic adversarial model, the model learns the discriminative features of the real samples and the perturbation samples through an adversarial training mechanism. During the iterative training process, the classifier's misclassification rate drives the optimization of the perturbation strategy until the model achieves stable discrimination ability for the mixed sample set. Preferably, the first preset training accuracy threshold is set based on clinical validation results, generally requiring the model's classification accuracy for normal gait and abnormal perturbations on an independent test set to exceed a preset clinical decision confidence level.

[0193] The trained basic adversarial model extracts the hidden feature distribution of group samples through forward propagation, and combines the kernel density estimation method to output the dynamic threshold range of the motion parameters corresponding to each sample, which is the sample threshold range.

[0194] When finally selecting the sample threshold range with the highest matching degree, a weighted fusion strategy is used to multiply the injury type weight, physiological feature weight, and movement abnormality weight, and the threshold interval corresponding to the first place in the comprehensive matching degree ranking is selected as the initialization result, which is the preset threshold range.

[0195] This embodiment matches highly correlated group samples based on injury type, physiological characteristics, and similarity of motion abnormalities, combines projected gradient descent to generate perturbation samples under biomechanical constraints, uses adversarial training to optimize the model's ability to distinguish between real and perturbation samples, outputs a dynamic threshold range, and then uses weighted fusion to screen the optimal matching results. This allows personalized initialization of the preset threshold range during the user's first training, effectively improving the adaptation accuracy of the threshold range to injury characteristics and motion abnormalities, ensuring biomechanical compliance, and enhancing the basic adversarial model's robust recognition of abnormal fluctuations in motion parameters.

[0196] In some embodiments, the personalized model further includes a threshold adjustment module, and the method further includes:

[0197] When the user is not training for the first time, the preset threshold range is configured to be updated through the threshold adjustment module, including:

[0198] After the user finishes exercising, obtain the user's current batch of real-time exercise data, record it as the latest exercise data, and obtain the historical exercise data of each batch within a preset historical period;

[0199] Generate personal adversarial samples based on the latest and historical motion data;

[0200] Acquire the latest user information of the user, and match a second preset number of latest group samples in the database according to the latest user information, where the matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity;

[0201] The latest group sample and the individual adversarial sample are mixed and input into the threshold adjustment module and iteratively trained until the second preset training accuracy threshold is met, indicating that the threshold adjustment module is updated;

[0202] The latest motion data is input into the updated threshold adjustment module to obtain the updated preset threshold range.

[0203] In this embodiment, the determination of the preset historical period needs to comprehensively consider the progress speed of the rehabilitation stage and the stability requirements of the motion data. For example, for users in the early stages of recovery, a shorter period is used to capture rapidly changing characteristics, while for users in the later stages, the period is extended to ensure data continuity.

[0204] Preferably, generating individual adversarial samples based on the latest and historical motion data can be understood as: extracting clinically significant feature frames from the latest motion data (e.g., frames with the highest degree of movement completion, frames with the largest deviation, and frames with typical errors), combining them with diffusion model decoupling to obtain mechanical parameters such as joint angles and muscle activation patterns, then generating variant samples based on the physical constraints in the historical motion data and the individual's movement pattern characteristics. Finally, these variant samples are inserted into the historical data in their original time sequence to form adversarial samples. Alternatively, a desensitization correction strategy that matches similar anomalies in a federated shared latent space can be used to reconstruct resonant adversarial samples that conform to the individual's biomechanical characteristics. Alternatively, standard motion features with interpretable labels can be generated based on biomechanical constraints, and effective adversarial samples can be screened in combination with risk assessment indicators to achieve a diverse construction of personalized adversarial samples. Details are described below.

[0205] Preferably, the second preset number is determined following the same coverage-efficiency balance principle as the first preset number, but additional consideration must be given to the impact of the user's latest physiological state changes on sample matching, such as dynamically adjusting the matching sample size based on rehabilitation stage transitions. When matching the latest group samples based on the latest user information, the injury type similarity uses a dynamically updated ICF coding comparison, the physiological feature similarity recalculates the normalized distance based on the latest physical measurement parameters, and the movement abnormality similarity optimizes the deviation vector weight using the latest gait analysis results, forming a multi-dimensional similarity evaluation system to screen the most relevant samples in the database.

[0206] After the latest group samples and individual adversarial samples are mixed and input into the threshold adjustment module, the model learns the discrimination boundary between real group data and personalized adversarial perturbations through an adversarial training mechanism. Its classifier continuously adjusts the threshold decision boundary in the process of distinguishing normal movement patterns from adversarial samples until the model achieves stable discrimination performance for the mixed dataset. Preferably, the determination of the second preset training accuracy threshold requires reference to historical training convergence curves and clinical expert evaluation results. It is generally required that the model's recognition accuracy rate for threshold-exceeding events on the validation set reaches a clinically acceptable level.

[0207] The updated threshold adjustment module analyzes the distribution offset of the latest motion data in the hidden feature space, combines the correlation between the group sample threshold range and the perturbation intensity of the adversarial sample, and dynamically adjusts the upper and lower limits of the thresholds of each motion parameter. It finally outputs an updated preset threshold range that conforms to the group rules and adapts to the individual's rehabilitation progress.

[0208] This embodiment sets a preset historical period by dynamically adapting the rehabilitation stage, generates personal adversarial samples based on the latest motion data and historical motion data, combines the latest group samples with multi-dimensional similarity matching, and uses adversarial training to optimize the threshold decision boundary, thereby achieving a dynamic balance between the threshold range of group rules and individual rehabilitation progress, and improving the personalized adaptability of motion abnormality recognition.

[0209] In some embodiments, generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes:

[0210] Extracting significant feature information from the latest motion data, the significant feature information includes the first motion frame with the highest motion completion, the second motion frame with the largest motion deviation, and the third motion frame with clinical typicality;

[0211] In the dormant state, semantic feature reconstruction is performed on the salient feature information, including:

[0212] The pre-set diffusion model is used to decouple the significant feature information and obtain the kinematic parameters, which include joint angles, joint torque distribution, motion velocity curves, and muscle force patterns.

[0213] Construct physical constraint information and personal constraint information based on motion mechanics parameters;

[0214] Generate variant samples similar to salient feature information based on historical movement data, physical constraint information, and personal constraint information;

[0215] The mutated samples are inserted into the historical motion data according to the execution sequence of the first action frame, the second action frame, and the third action frame in the latest action data to obtain personal adversarial samples.

[0216] In this embodiment, the generation of personal adversarial samples is achieved by analyzing the significant feature information in the user's latest motion data and fusing historical motion data for mutation reconstruction.

[0217] The first action frame in the significant feature information reflects the user's current optimal motion performance and is used to capture the biomechanical characteristics under the correct motion mode; the second action frame reveals the abnormal mechanical state corresponding to the maximum motion deviation, which is used to locate the weak links in rehabilitation training; the third action frame is associated with clinical typical movement disorder patterns (such as foot drop and knee hyperextension), which is used to match common abnormal features to enhance the medical relevance of the sample.

[0218] Semantic feature reconstruction is performed in a dormant state, meaning background processing tasks are initiated during the system's non-real-time operation period. Feature decoupling of multimodal motion data is performed using a pre-trained diffusion model based on a variational autoencoder architecture. Noise perturbations are gradually injected through a forward diffusion process, and kinematic parameters such as joint angles and joint torque distributions are separated through an inverse denoising network, achieving the conversion of raw sensor data into interpretable biomechanical indicators.

[0219] Constructing physical constraints based on kinematic parameters can be understood as extracting human kinematic limits (such as maximum joint range of motion) from decoupled joint angles and torque distributions to form rigid boundary constraints. Personal constraints, on the other hand, generate soft adaptation rules (such as gait phase offset tolerance) based on user-specific velocity curve patterns and muscle force timing from historical motion data. These two constraints constrain the generation logic of variant samples from the perspectives of biomechanical universality and individual motion characteristics, respectively, ensuring that they conform to anatomical laws while maintaining the user's personalized motion pattern.

[0220] When generating variant samples, physical constraint information is used as hard boundary conditions and personal constraint information is used as soft optimization targets. A generative adversarial network is used to search for feasible solutions in the latent space of historical motion data that are similar to the information mechanics of significant features but with parameter perturbations, thus forming new samples that retain user personalized features while introducing controllable variations.

[0221] Preferably, the process of inserting the variant samples into the historical data in the original time sequence is achieved through a motion phase alignment algorithm. For example, the first motion frame variant sample is inserted at the beginning of the gait support phase to strengthen the memory of the correct pattern, and the second motion frame variant sample is inserted at the end of the swing phase to simulate the extreme deviation scenario, thereby constructing an adversarial sample set covering multiple scenarios.

[0222] This embodiment extracts motion mechanics parameters that are deeply coupled with clinical characteristics through analysis of key action frames (i.e., significant feature information), and combines the physical-individual dual constraint mechanism to generate safe and compliant variant samples, providing the threshold adjustment module with adversarial training data that has both individual adaptability and biomechanical rationality, thereby enhancing the model's generalization ability for complex motion patterns.

[0223] In some embodiments, generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes:

[0224] Extracting first abnormal feature information from the latest motion data, the first abnormal feature information including joint motion trajectory deviation features, muscle activation abnormality features, and motion timing disorder features;

[0225] Mapping the first anomaly feature information to the federated shared latent space for similarity matching includes:

[0226] Calculate the latent space projection similarity with the second abnormal feature information of other users in the database;

[0227] Filter matching cases whose latent space projection similarity exceeds the preset similarity threshold, and obtain the desensitization correction strategy associated with the matching cases;

[0228] The first abnormal feature information is fused and reconstructed according to the desensitization correction strategy to generate a resonance adversarial sample after the first abnormal feature information is corrected;

[0229] The generated resonance adversarial sample is mixed with the historical motion data and the latest motion data to obtain the personal adversarial sample.

[0230] In this embodiment, the first abnormal feature information is used to accurately characterize the user's current motor dysfunction pattern. Mapping the first abnormal feature to a federated shared latent space for similarity matching can optimize the generation of local adversarial examples while desensitizing the data and leveraging the correction experience of similar abnormal cases in the group knowledge base. The federated shared latent space refers to a cross-institutional feature encoding space obtained through pre-training within a federated learning framework, enabling feature-level information exchange without sharing the original data.

[0231] Preferably, when calculating the latent space projection similarity, the cosine similarity measurement method of the feature vector is used to evaluate the directional alignment of the latent vector after encoding the first abnormal feature with the second abnormal feature latent vector of other users in the database, and screen out matching cases with highly similar motion abnormality patterns.

[0232] The preset similarity threshold is determined based on the cluster analysis results of historical rehabilitation cases. By statistically analyzing the distribution density of abnormal features of users with similar injuries in the latent space and combining expert experience, the minimum similarity threshold (such as 0.75) is set to ensure that the matching cases have clinical reference value.

[0233] The desensitization correction strategy refers to the anonymization of sensitive parameters involving personal identity information in matching cases (such as age and height), while retaining its abnormal correction logic (such as the adjustment coefficient of the knee hyperextension correction angle).

[0234] A feature interpolation algorithm weights the first anomaly feature and the correction parameters of the matching case to reconstruct a resonant adversarial sample that retains the user's individual anomaly characteristics while incorporating collective correction experience. For example, to address a user's inversion trajectory deviation, the matching case provides a strategy for enhancing the valgus torque. After desensitization, the similarity-weighted fusion generates a correction sample tailored to the user's current muscle strength level.

[0235] Finally, the resonant adversarial samples are mixed with historical and latest motion data to form a dataset covering abnormal patterns in both individual and group dimensions, providing the threshold adjustment module with adversarial training samples that are both biomechanically reasonable and privacy-compliant, thereby enhancing the model's generalized recognition capabilities for complex movement disorders.

[0236] This embodiment achieves privacy-preserving matching of cross-user abnormal features and adversarial sample generation through a federated shared latent space. After extracting features of joint trajectory deviation, muscle activation anomalies, and timing imbalances, the first abnormal feature is mapped to the shared latent space constructed by federated learning for cosine similarity matching. A desensitization correction strategy is then applied to select highly similar cases. Resonant adversarial samples are generated through feature interpolation and fusion, and finally, historical and recent motion data are mixed to form a training set. While protecting user privacy, this method generates adversarial samples that are both biomechanically sound and individually adaptable through the dynamic fusion of group correction experience and individual abnormal features. This effectively improves the threshold adjustment module's generalized recognition accuracy for complex movement disorder patterns and the reliability of rehabilitation decisions.

[0237] In some embodiments, generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes:

[0238] Extracting first motion feature information from the latest motion data, the first motion feature information including the user's habitual motion pattern features, typical incorrect motion features, and biomechanically reasonable motion features;

[0239] generating second motion characteristic information based on the biomechanical constraint condition, wherein the second motion characteristic information is configured to meet clinical standards but has a preset difference from the first motion characteristic information;

[0240] Add an explainability label to each second action feature information, where the explainability label contains specific quantitative indicators required for action improvement;

[0241] Inputting the first motion feature information and the second motion feature information into the personalized model respectively, obtaining a first risk assessment result corresponding to the first motion feature information and a second risk assessment result corresponding to the second motion feature information;

[0242] generating an acceptance index based on the second risk assessment result, and generating a stubbornness index based on the first risk assessment result;

[0243] The second action feature information is filtered according to the acceptance index and stubbornness index, and recorded as a valid adversarial sample. It is then mixed with the historical motion data and the latest motion data to obtain a personal adversarial sample.

[0244] In this embodiment, the first motion feature information extracted from the latest motion data includes the user's habitual motion pattern characteristics (such as a small knee flexion angle in the gait cycle), typical incorrect motion characteristics (such as excessive ankle inversion when the heel touches the ground) and biomechanically reasonable motion characteristics (such as a hip joint extension angle that meets standard gait parameters).

[0245] The process of generating the second motion feature information based on biomechanical constraints can be understood as constructing ideal motion parameters based on clinical kinematic specifications (such as the safe range of joint mobility), introducing controllable variations through preset differences (such as ±5° offset of joint angle, ±10% delay of timing phase), and forming a standard motion feature set with adjustable deviations from the user's current motion pattern.

[0246] The explainable labels added to each second action feature information contain specific quantitative indicators of action improvement requirements (such as "hip abduction angle needs to be increased by 3°" or "plantar pressure center moves forward 2cm"), which are used to clarify the optimization direction and execution standards of adversarial samples.

[0247] After inputting the first motion feature information and the second motion feature information into the personalized model, the first risk assessment result obtained reflects the motion risk probability of the user's current motion mode (such as habitual insufficient flexion leading to a risk value of 0.4 for knee arthritis), and the second risk assessment result evaluates the potential risk after the standard motion is adjusted (such as a risk value of 0.2 after increasing the hip abduction angle).

[0248] The acceptance index is generated by comparing the second risk assessment result with the user's physiological tolerance threshold, representing the feasibility of the standard action adjustment plan (for example, if the risk reduction reaches 30%, the acceptance is high); the stubbornness index is calculated based on the degree of deviation of the first risk assessment result relative to the historical training data, reflecting the potential harm of the user maintaining the current action pattern (for example, if the risk value is continuously 20% higher than the average of similar users, the stubbornness is high).

[0249] When screening for effective adversarial samples, priority is given to second motion feature information whose acceptance index is higher than the preset acceptance threshold and whose stubbornness index reaches the alarm level (e.g., acceptance > 0.7 and stubbornness > 0.6), ensuring that the adversarial sample can both guide the user to transition to biomechanically sound movements and break through the training inertia of their inherent error patterns. Preferably, the preset acceptance threshold is set by statistically analyzing the median distribution of the risk reduction rate after adjusting for standard movements in the historical training of users with similar injuries, for example, taking the lowest acceptance percentile of successful cases; the alarm level is determined based on the statistical fluctuation range of the user's historical risk assessment results, and is triggered when the stubbornness index exceeds two standard deviations of the individual's historical mean. The specific threshold is calibrated by clinical experts based on the injury type and rehabilitation stage to ensure that the screening criteria are both universal and individual.

[0250] Finally, historical motion data, the latest motion data and the screened effective adversarial samples are mixed to form a multi-dimensional training set covering habitual patterns, standard movements and gradual adjustment instructions. By repeatedly exposing users to optimized motion scenarios, the model's ability to guide behavioral pattern migration is enhanced, achieving targeted correction of incorrect movements in rehabilitation training and gradual reshaping of exercise habits.

[0251] It's important to note that all three aforementioned embodiments generate individual adversarial examples based on the latest motion data and historical motion data from a preset historical period. The differences are as follows: The first of these three embodiments focuses on semantic reconstruction based on the user's own motion characteristics. By combining physical constraints with historical data, it generates variant examples that conform to individual biomechanical characteristics, primarily enhancing the model's ability to identify deviations from specific motion patterns. The second of these three embodiments utilizes a correction strategy that matches external similar cases in a federated shared latent space, extending the model's generalization to rare abnormal patterns through resonant adversarial examples. The third of these three embodiments integrates clinical criteria to generate comparative examples with interpretable labels, combining risk assessment to screen effective adversarial examples and guide the model's identification of appropriate improvement directions. These three embodiments generate adversarial examples from three dimensions: individual feature enhancement, group knowledge transfer, and clinical rule guidance. These three methods complement each other: the first ensures personalized adaptation, the second supplements data diversity, and the third incorporates clinical prior knowledge. Together, these enhance the threshold adjustment module's dynamic adaptability to complex motion anomalies.

[0252] See also Figure 5 In the second aspect, this embodiment further provides a lower limb motion data real-time analysis system 1 for rehabilitation guidance, which is applicable to the real-time analysis method of the first aspect. The system includes a data acquisition module 11 and a logic processing module 12;

[0253] The data acquisition module 11 is used to obtain user information. The data acquisition module 11 is also used to obtain the real-time motion data of the current user according to a preset frequency;

[0254] The logic processing module 12 is used to call the personalized model corresponding to the current user according to the user information. The personalized model is configured as a group adversarial pre-training model and includes a real-time prediction module.

[0255] The logic processing module 12 is also used to input the real-time motion data into the personalized model to obtain the user's motion risk probability and correction information. The correction information includes device parameter adjustment values ​​and posture adjustment information, including:

[0256] The real-time motion data is converted into historical motion feature vectors of multiple frames arranged in chronological order and input into the real-time prediction module, which is configured to be built based on a lightweight LSTM network model;

[0257] The real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of multiple frames to obtain the future motion feature vectors of multiple frames arranged in chronological order;

[0258] Obtaining a preset threshold range, calculating the deviation value between each future motion feature vector and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order;

[0259] The motion risk ratio of each future motion feature vector is calculated based on the motion deviation value, and multiple motion risk ratios are weightedly fused to obtain the motion risk probability;

[0260] and, reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user;

[0261] Based on the risk probability of the movement, it is determined whether to execute the next stage of the training step, and the correction information is displayed and the corresponding equipment operating parameters are adjusted.

[0262] In this embodiment, preferably, the data acquisition module 11 includes a wearable inertial measurement unit, a joint angle sensor and a physiological parameter acquisition device, which is used to obtain lower limb motion trajectory, joint angle changes, user heart rate, electromyography and other physiological data in real time, and integrates a user information entry interface to receive clinical information such as injury type and rehabilitation stage.

[0263] The real-time analysis system of this embodiment executes the real-time analysis method described in the first aspect. The specific content is referred to the aforementioned technical solution and will not be described here in detail.

[0264] By adopting the above technical solution, the present invention is different from the existing technology and has the following beneficial effects:

[0265] This invention achieves precise rehabilitation decision-making by constructing a group-individual adversarial pre-training model. Based on user information, it invokes a personalized model. A lightweight LSTM network is used to encode historical motion feature vectors in real time and decode them to predict future motion trends. A dynamic threshold mechanism is then used to calculate motion deviations, simultaneously generating device parameter adjustment values ​​and posture correction information. The adversarial training framework balances group universality with individual variability. Inverse mapping of the device parameter matrix enables multi-dimensional adjustment of power, speed, and damping. Safe correction instructions are generated based on the anatomical constraints of the user's posture matrix. A dynamic threshold initialization and update mechanism integrates group sample matching, biomechanical constraint repair, and a federated shared latent space matching strategy. It incorporates group correction experience to generate personalized adversarial examples while preserving privacy. Effective adversarial examples are screened based on biomechanically plausible motion characteristics and risk assessment indicators. Acceptance and stubbornness indicators are then combined to optimize motion transfer guidance. The system utilizes real-time motion risk probability assessment and device-posture coordinated adjustment to form a closed-loop optimized rehabilitation training process control. This system improves motion trajectory tracking accuracy while maintaining physiological safety boundaries, achieving accurate, real-time, and safe personalized rehabilitation guidance.

[0266] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of this application, this does not limit the scope of patent protection of this application. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concepts of this application using the contents recorded in the specification and drawings of this application, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the scope of patent protection of this application.

Claims

1. A real-time analysis method for lower limb motion data used in rehabilitation guidance, characterized in that: include: Obtaining user information, and calling a personalized model corresponding to the current user based on the user information, wherein the personalized model is configured as a group-individual adversarial pre-training model and includes a real-time prediction module; The real-time motion data of the current user is obtained at a preset frequency and input into the personalized model to obtain the user's motion risk probability and correction information, wherein the correction information includes device parameter adjustment values ​​and posture adjustment information, including: Converting the real-time motion data into a multi-frame historical motion feature vector arranged in chronological order, and inputting the historical motion feature vector into a real-time prediction module, wherein the real-time prediction module is configured to be constructed based on a lightweight LSTM network model; The real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of multiple frames to obtain future motion feature vectors of multiple frames arranged in chronological order; Obtaining a preset threshold range, calculating a deviation value between each of the future motion feature vectors and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order; Calculating the motion risk ratio of each of the future motion feature vectors according to the motion deviation value, and performing weighted fusion on multiple motion risk ratios to obtain the motion risk probability; and reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user; Evaluating whether to execute the next stage of training steps based on the motion risk probability, and displaying the correction information and adjusting corresponding equipment operating parameters; The motion risk ratio of each future motion feature vector is calculated according to the motion deviation value, and a plurality of motion risk ratios are weightedly integrated to obtain the motion risk probability, including: Obtaining a preset joint weight distribution table, wherein the joint weight distribution table stores weight coefficients of each lower limb joint at different rehabilitation stages, wherein the weight coefficients are positively correlated with the importance of the lower limb joints in rehabilitation training; Extracting a corresponding set of weight coefficients from the joint weight distribution table according to the user's current rehabilitation stage; For each of the future motion feature vectors, performing a matrix multiplication operation on the motion deviation components of each lower limb joint contained therein and the corresponding weight coefficient to obtain a dimensionally compressed joint weighted deviation value; The joint weighted deviation value is input into the rehabilitation risk quantification model for nonlinear normalization processing, and the motion risk ratio of each future motion feature vector is output. The rehabilitation risk quantification model is obtained by the following steps: Obtain a preset age-parameter mapping table and call a basic parameter set of the sigmoid function that matches the user's age; Query the stage-slope comparison table according to the current rehabilitation stage to obtain the corresponding slope adjustment coefficient; Performing feature fusion on the basic parameter group and the slope adjustment coefficient through a one-dimensional convolution layer to generate a user-adaptive normalization function, namely, the rehabilitation risk quantification model; Using a sliding time window method to perform time series fusion on the motion risk proportions of multiple frames to obtain the motion risk probability; The motion deviation value is reversely mapped to the device parameter matrix to obtain multiple device parameter adjustment values ​​including: Build a device parameter mapping relationship library, including: Obtain the preset rehabilitation stage-power response curve and establish a nonlinear mapping relationship between the motion deviation value and the device power adjustment amount at different rehabilitation stages; Configure the motion mode-speed comparison matrix to define the reference speed and allowable adjustment range corresponding to each phase of the gait cycle; Loading a safety constraint parameter table to store the damping adjustment threshold boundary based on the user's physiological characteristics; Decomposing the motion deviation value according to the device parameter mapping relationship library includes: According to the current rehabilitation stage index to the rehabilitation stage-power response curve, the time domain integral of the motion deviation value is input into the power conversion module, and the power adjustment amount adapted to the output stage is output; Analyzing the frequency domain characteristic components of the motion deviation value and matching the motion mode-speed comparison matrix to obtain a reference speed correction coefficient; Inputting the spatial distribution characteristics of the motion deviation value and the user's physiological parameters into the damping regulator, and combining the safety constraint parameter table to generate a compliant damping adjustment amount; Output device parameter adjustment values, where the device parameter adjustment values ​​include a power adjustment amount, a reference speed correction coefficient, and a compliance damping adjustment amount.

2. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 1, characterized in that: The motion deviation value is reversely mapped to the user posture matrix to obtain the posture adjustment information of the current user, including: Construct the user pose matrix, including: Obtain a preset joint-posture mapping table and establish the corresponding relationship between the motion deviation components of each lower limb joint and the ideal posture parameters; Configure the rehabilitation stage-correction intensity comparison table to define the maximum posture adjustment range allowed at different rehabilitation stages; Load the user's physiological-posture constraint table and store the personalized posture safety range based on the user's height and weight; Decomposing the motion deviation value according to the user posture matrix includes: Decomposing the motion deviation value into each joint motion plane, and obtaining the initial posture correction value through the joint-posture mapping table; According to the current rehabilitation stage, the rehabilitation stage-correction intensity comparison table is searched, and the initial posture correction amount is subjected to intensity normalization processing to obtain a correction intensity coefficient; Inputting the correction intensity coefficient, the initial posture correction amount and the user's physiological parameters into the posture optimizer, and generating a safe posture adjustment instruction and a final posture correction amount in combination with the user's physiological-posture constraint table; Output posture adjustment information, which includes a correction strength coefficient, an initial posture correction amount, a final posture correction amount, and a safe posture adjustment instruction.

3. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 1, characterized in that: The preset threshold range is configured to be initialized by the following steps when the user trains for the first time: Acquire user information, and match a first preset number of group samples in a database according to the user information, where matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity; Obtain a basic adversarial model, and iteratively train the basic adversarial model using population samples, including: The projected gradient descent method is used to generate disturbance samples that meet the preset disturbance range; Repairing the disturbed sample according to preset constraints, where the preset constraints are biomechanical constraints; Mixing the population sample and the perturbation sample and inputting them into the basic adversarial model for iterative training until a first preset training accuracy threshold is met, indicating that the basic adversarial model training is completed; Inputting the group samples into the trained basic adversarial model, and outputting the sample threshold range of each group sample; The plurality of sample threshold ranges are arranged in order of the degree of matching between the group sample and the user information, and the sample threshold range with the highest degree of matching is selected as the preset threshold range.

4. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 1, characterized in that: The personalized model further includes a threshold adjustment module, and the method further includes: When the user is not training for the first time, the preset threshold range is configured to be updated by the threshold adjustment module, including: After the user finishes exercising, obtain the user's current batch of real-time exercise data, record it as the latest exercise data, and obtain the historical exercise data of each batch within a preset historical period; generating a personal adversarial sample based on the latest motion data and the historical motion data; Acquiring the latest user information of the user, and matching a second preset number of latest group samples in the database according to the latest user information, where the matching constraints include injury type similarity, physiological feature similarity, and movement abnormality similarity; Mixing the latest group sample with the individual adversarial sample and inputting it into the threshold adjustment module for iterative training until a second preset training accuracy threshold is met, indicating that the threshold adjustment module has been updated; The latest motion data is input into the updated threshold adjustment module to obtain an updated preset threshold range.

5. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 4, characterized in that: Generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes: Extracting significant feature information from the latest motion data, the significant feature information including a first motion frame with the highest motion completion, a second motion frame with the largest motion deviation, and a third motion frame with clinical typicality; In the dormant state, semantic feature reconstruction is performed on the salient feature information, including: Decoupling the significant feature information through a preset diffusion model to obtain kinematic parameters, which include joint angles, joint torque distributions, motion velocity curves, and muscle force patterns; constructing physical constraint information and personal constraint information according to the kinematic parameters; generating a variant sample similar to the salient feature information based on the historical motion data, the physical constraint information, and the personal constraint information; The variant sample is inserted into the historical motion data according to the execution sequence of the first action frame, the second action frame, and the third action frame in the latest action data to obtain the personal adversarial sample.

6. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 4, characterized in that: Generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes: extracting first abnormal feature information from the latest motion data, wherein the first abnormal feature information includes a joint motion trajectory deviation feature, a muscle activation abnormality feature, and a motion timing disorder feature; Mapping the first abnormal feature information to the federated shared latent space for similarity matching includes: Calculate the latent space projection similarity with the second abnormal feature information of other users in the database; Filter matching cases whose latent space projection similarity exceeds the preset similarity threshold, and obtain the desensitization correction strategy associated with the matching cases; fusing and reconstructing the first abnormal feature information according to the desensitization correction strategy to generate a resonance adversarial sample after correction of the first abnormal feature information; The generated resonance adversarial sample is mixed with the historical motion data and the latest motion data to obtain the personal adversarial sample.

7. The real-time analysis method for lower limb motion data for rehabilitation guidance according to claim 4, characterized in that: Generating a personal adversarial sample based on the latest motion data and historical motion data within a preset historical period includes: Extracting first motion feature information from the latest motion data, the first motion feature information including a user's habitual motion pattern features, typical incorrect motion features, and biomechanically reasonable motion features; generating second motion characteristic information based on the biomechanical constraint condition, wherein the second motion characteristic information is configured to meet clinical standards but has a preset difference from the first motion characteristic information; Adding an explainability label to each second action feature information, wherein the explainability label includes a specific quantitative indicator of action improvement requirements; Inputting the first motion feature information and the second motion feature information into a personalized model respectively to obtain a first risk assessment result corresponding to the first motion feature information and a second risk assessment result corresponding to the second motion feature information; generating an acceptance index based on the second risk assessment result, and generating a stubbornness index based on the first risk assessment result; The second action feature information is filtered according to the acceptance index and the stubbornness index, and recorded as a valid adversarial sample, and mixed with the historical motion data and the latest motion data to obtain the personal adversarial sample.

8. A real-time analysis system for lower limb motion data used in rehabilitation guidance, characterized in that: The real-time analysis method according to any one of claims 1 to 7 is applicable, wherein the system comprises a data acquisition module and a logic processing module; The data acquisition module is used to obtain user information, and the data acquisition module is also used to obtain the real-time motion data of the current user according to a preset frequency; The logic processing module is used to call the personalized model corresponding to the current user according to the user information, the personalized model is configured as a group adversarial pre-training model, and the personalized model includes a real-time prediction module; The logic processing module is further configured to input real-time motion data into the personalized model to obtain the user's motion risk probability and correction information, wherein the correction information includes device parameter adjustment values ​​and posture adjustment information, including: Converting the real-time motion data into a multi-frame historical motion feature vector arranged in chronological order, and inputting the historical motion feature vector into a real-time prediction module, wherein the real-time prediction module is configured to be constructed based on a lightweight LSTM network model; The real-time prediction module performs encoding and decoding operations on the historical motion feature vectors of multiple frames to obtain future motion feature vectors of multiple frames arranged in chronological order; Obtaining a preset threshold range, calculating a deviation value between each of the future motion feature vectors and the preset threshold range, and obtaining motion deviation values ​​arranged in chronological order; Calculating the motion risk ratio of each of the future motion feature vectors according to the motion deviation value, and performing weighted fusion on multiple motion risk ratios to obtain the motion risk probability; and reversely mapping the motion deviation value to a device parameter matrix to obtain a plurality of device parameter adjustment values, and reversely mapping the motion deviation value to a user posture matrix to obtain posture adjustment information of the current user; Whether to execute the next stage training step is determined based on the motion risk probability assessment, and the correction information is displayed and the corresponding equipment operating parameters are adjusted.

Images