Physical training management system based on AI intelligence

Abstract

The invention discloses a body training management system based on AI intelligence, and relates to the technical field of artificial intelligence, and the system comprises a data acquisition module, an action evaluation module, a data processing module and a data processing module, the input module is used for inputting the original data sequence into a pre-trained AI model and outputting a corresponding action standard evaluation result, the feedback generation module is used for generating feedback information in real time based on the action standard evaluation result and transmitting the feedback information to terminal equipment of a user, and the posture adjustment module is used for adjusting the posture of the user. The user is guided to adjust action postures according to the feedback information to avoid sport injuries caused by improper postures; according to the AI intelligence-based body training management system, the actual activation condition of muscles can be recognized from the physiological level, the recognition deviation caused by only depending on image or motion track information is made up, and the recognition ability for tiny motion differences is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to an AI-based body training management system. Background Technology

[0002] With increasing health awareness, physical training and sports rehabilitation are gradually becoming important parts of daily life. Especially in the fields of fitness, rehabilitation medicine, and athletic performance improvement, scientific and personalized training guidance is becoming a mainstream demand. Traditional training guidance typically relies on the coach's experience and judgment or posture assessment methods based on visual observation.

[0003] However, these methods are highly subjective and lack real-time performance, making it difficult to achieve high-precision motion recognition and risk warnings. This is especially true for individual users without professional guidance, who are prone to muscle strains, joint injuries, and other sports-related risks due to improper posture. Currently, some training systems have incorporated motion capture devices (such as inertial sensors or video image recognition technology) for posture analysis, but these systems often fail to deeply identify muscle activity states and lack effective perception of hidden dangers such as muscle fatigue and abnormal load. Furthermore, traditional motion recognition systems struggle to achieve high-frequency, high-sensitivity real-time feedback and lack effective personalized motion evaluation mechanisms. Electromyography (EMG), as a bioelectrical signal reflecting muscle electrical activity, can directly reflect the degree of muscle contraction and neural control characteristics. Therefore, an increasing number of studies are attempting to combine EMG signals with artificial intelligence algorithms for motion recognition and physiological state assessment. Although preliminary applications have been made, most remain in the experimental verification stage, lacking a systematic and practical overall solution. In particular, existing technologies still have significant shortcomings in real-time assessment of motion standardization, establishing individualized motion models, and providing dynamic feedback during training to prevent sports injuries. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based intelligent body training management system that performs real-time assessment of movement standardization based on muscle electrical activity signals to solve the problem of sports injuries caused by improper posture.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-based intelligent body training management system, the system comprising: The data acquisition module is used to collect the user's muscle electrical activity signals based on sensors and to preprocess them to generate raw data sequences; The action evaluation module, connected to the data acquisition module, is used to input the raw data sequence into the pre-trained AI model and output the corresponding action standardization evaluation results, including the matching score between each action and the standard template output by the AI ​​model, determining the action violation probability, setting the action violation probability threshold, and recording and correcting actions whose action violation probability exceeds the action violation probability threshold. The feedback generation module, connected to the motion assessment module, is used to generate feedback information in real time based on the motion standardization assessment results and transmit it to the user's terminal device. The posture adjustment module, connected to the feedback generation module, guides users to adjust their postures based on feedback information to avoid sports injuries caused by improper postures. This includes obtaining the lateral displacement value of the skeletal points of the user's actual movement, obtaining the distance between the hand and the camera and the standard projection depth, calculating the projection offset value of the user's current posture on the standard plane, setting the maximum projection offset value of the user's current posture on the standard plane, and prompting the user to correct the movement if the projection offset value exceeds the maximum projection offset value, and displaying a correction direction arrow.

[0006] Preferably, the data acquisition module collects user electromyographic activity signals based on sensors and performs preprocessing to generate raw data sequences. This includes dividing the electromyographic activity signals into multiple fixed-length time periods, using a template matching method to determine whether each period is a known waveform repeated, accumulating the number of detected repetitive electromyographic patterns, using the ratio of the accumulated number of detected repetitive electromyographic patterns to the total number of sampling segments in the acquisition window as a signal repetition score, setting an abnormal repetition threshold, and indicating the presence of muscle tremors or signal grounding problems if the signal repetition score is greater than the abnormal repetition threshold.

[0007] Preferably, the feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device. This includes determining the position coordinates of the target muscle group that needs to trigger feedback, obtaining the current feedback point position of the user's wearable device, calculating the distance between the feedback point position and the target muscle group position coordinates, calculating the user's perceived stimulus intensity, setting a user's perceived stimulus intensity threshold, and adjusting the feedback device or changing the feedback method when the user's perceived stimulus intensity is less than the user's perceived stimulus intensity threshold.

[0008] Preferably, the action evaluation module inputs the original data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation result. The specific steps for determining the action violation probability include subtracting one hundred from the matching score between each action output by the AI ​​model and the standard template to obtain the action violation probability.

[0009] Preferably, the feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device to calculate the user's perceived stimulus intensity. The specific steps include calculating the three-dimensional distance between the position coordinates of the target muscle group that needs to trigger feedback and the current feedback point position of the user's wearable device, using an exponential decay operation, inputting the calculated three-dimensional distance as the negative part of the exponential operation, and using the operation result as the current user's perceived stimulus intensity.

[0010] Preferably, the specific steps in the posture adjustment module for guiding the user to adjust their posture based on feedback information to avoid sports injuries caused by improper posture include calculating the projection offset value of the user's current posture on the standard plane. These steps include multiplying the lateral displacement value of the skeletal point of the user's actual movement by the distance between the hand and the lens and the standard projection depth, and dividing the multiplication result by the actual straight-line distance between the skeletal point and the current distance from the camera lens to obtain the projection offset value of the user's current posture on the standard plane.

[0011] Preferably, the data acquisition module, based on the sensor, acquires the user's electromyographic activity signal and performs preprocessing to generate the raw data sequence, further includes: The sampling frequency from which electromyographic signals are extracted from the sensor; After sampling is completed, the absolute values ​​of all sampling points are counted, and the absolute values ​​of the sampling points are summed. The average value is obtained by dividing the total number of sampling points by the sum of the absolute values ​​of the sampling points. The original signal is input into a bandpass filter module, which retains components between 20 Hz and 500 Hz. Iterate through each sampling point and determine whether the absolute amplitude of the sampling point is higher than the average level of the overall signal.

[0012] Preferably, the specific steps of traversing each sampling point and determining whether the absolute amplitude of the sampling point is higher than the average level of the overall signal include: if the absolute amplitude of the sampling point exceeds the average amplitude multiplied by a preset multiple threshold, wherein the multiple threshold is between 2 and 3, the sampling point is considered a real active signal; if the absolute amplitude of the sampling point does not exceed the average amplitude multiplied by a preset multiple threshold, the sampling point is considered an artifact or invalid noise and is automatically discarded or ignored.

[0013] Preferably, the action evaluation module further includes inputting the raw data sequence into a pre-trained AI model and outputting the corresponding action standardization evaluation result, including: Feature extraction is performed on the preprocessed electromyographic (EMG) signals, which is divided into two dimensions: the first dimension is time features, which reflect the characteristics of the EMG signals in the time series, including mean, variance, waveform length, and zero crossover rate; the second dimension is frequency features, which involves analyzing the spectral composition of the EMG signals through Fourier transform or wavelet transform to extract peak frequency, average frequency, and frequency band energy parameters.

[0014] Preferably, the action evaluation module further includes inputting the raw data sequence into a pre-trained AI model and outputting the corresponding action standardization evaluation result, including: The time features and frequency features are weighted and fused according to their respective weights. The time features are multiplied by a time weight coefficient, and the frequency features are multiplied by a frequency weight coefficient. Then, the result of multiplying the time features by a time weight coefficient and the result of multiplying the frequency features by a frequency weight coefficient are added together to obtain the comprehensive feature value. The comprehensive feature values ​​are input into the AI ​​model, which outputs a standard action score. The standard action score is set between zero and one. If the standard action score is less than the preset standard threshold, the action is judged as a non-standard action, and the user is prompted to adjust their posture.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This AI-powered body training management system uses a data acquisition module to collect muscle electrical activity signals from sensors and preprocess them to generate raw data sequences. A motion assessment module inputs these raw data sequences into a pre-trained AI model and outputs corresponding motion standardization assessment results. A feedback generation module generates feedback information in real time based on the motion standardization assessment results and transmits it to the user's terminal device. A posture adjustment module guides the user to adjust their posture based on the feedback information to avoid sports injuries caused by improper posture. It can physiologically identify the actual activation of muscles, compensating for the recognition bias caused by relying solely on image or motion trajectory information, effectively improving the ability to recognize subtle differences in movement, and can monitor muscle contraction in real time. It mitigates hidden risks such as skill level, fatigue level, and abnormal load, enhancing the comprehensive control over the user's movement status. It dynamically adjusts the weights of time and frequency domain information based on different user characteristics, and combines standard threshold evaluation results to output personalized action scores and intelligent judgment results. This enables intelligent feedback strategy adjustments based on offset distance, including vibration point control and voice prompt switching, thereby quickly guiding users to correct their posture and avoid sports injuries caused by improper movements. It is widely applicable to various scenarios such as home fitness, post-operative rehabilitation, and professional training, possessing advantages such as strong wearability, flexible deployment, and automated data processing. It overcomes the limitations of existing systems, which are highly experimental and lack practicality, helping to reduce the incidence of sports injuries and improve training safety and effectiveness stability. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the system modules of the present invention. Detailed Implementation

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

[0018] like Figure 1 As shown, the present invention provides a technical solution: an AI-based intelligent body training management system, the system comprising: The data acquisition module is used to collect the user's muscle electrical activity signals based on sensors and to preprocess them to generate raw data sequences; The action evaluation module, connected to the data acquisition module, is used to input the raw data sequence into the pre-trained AI model and output the corresponding action standardization evaluation results, including the matching score between each action and the standard template output by the AI ​​model, determining the action violation probability, setting the action violation probability threshold, and recording and correcting actions whose action violation probability exceeds the action violation probability threshold. The feedback generation module, connected to the motion assessment module, is used to generate feedback information in real time based on the motion standardization assessment results and transmit it to the user's terminal device. The posture adjustment module, connected to the feedback generation module, guides users to adjust their postures based on feedback information to avoid sports injuries caused by improper postures. This includes obtaining the lateral displacement value of the skeletal points of the user's actual movement, obtaining the distance between the hand and the camera and the standard projection depth, calculating the projection offset value of the user's current posture on the standard plane, setting the maximum projection offset value of the user's current posture on the standard plane, and prompting the user to correct the movement if the projection offset value exceeds the maximum projection offset value, and displaying a correction direction arrow.

[0019] In the AI-based intelligent body training management system, the data acquisition module is used to collect the user's muscle electrical activity signals based on sensors and perform preprocessing to generate raw data sequences. The sensors may include surface electromyography (EMG) sensors, accelerometers, or inertial measurement units, which can monitor the electrical activity changes of specific muscle groups in the user's body in real time. The preprocessing process includes signal denoising, amplitude normalization, and feature extraction to generate high-quality raw data sequences, providing a stable data foundation for subsequent AI model analysis.

[0020] The action evaluation module, connected to the data acquisition module, inputs the raw data sequence into a pre-trained AI model. This AI model, employing a deep neural network structure (such as a Convolutional Neural Network (CNN) or a Long Short-Term Memory (LSTM) network), has been trained through supervised learning using a large number of standard and abnormal action samples to acquire the ability to recognize actions and detect anomalies. Upon receiving the raw data sequence, the model outputs the corresponding action standardization evaluation result. This evaluation result includes a matching score between each action and a standard template, reflecting the similarity between the current action and the standard reference.

[0021] The motion assessment module further calculates the probability of motion violations based on the matching scores. The system establishes a classification model through the score distribution in historical samples and determines a "motion violation probability threshold" within the statistical model. This threshold is determined based on the intersection range of the matching scores between standard and non-standard motions, and the optimal critical point is often determined through ROC curve analysis to ensure a balance between sensitivity and specificity. When the probability of a motion violation exceeds the threshold, the motion is judged to have a posture problem, the system records it in the log, and generates a correction prompt.

[0022] A feedback generation module connected to the motion assessment module generates feedback information in real time based on the motion standardization assessment results and transmits it to the user's terminal device. This feedback information includes text, images, voice, or vibration prompts. Its generation logic automatically adjusts the feedback content and method based on the severity of the motion violation, the type of motion, and user preferences. The terminal device can be a smartphone, a smart wearable device, or a fitness interactive screen, realizing a real-time feedback loop during the user's training process.

[0023] The posture adjustment module, connected to the feedback generation module, guides the user to adjust their posture based on feedback information to avoid sports injuries caused by improper posture. Specifically, it acquires the lateral displacement values ​​of skeletal points during the user's actual movement; this skeletal point data is obtained through image recognition or depth camera equipment. Simultaneously, it collects the distance between the hand and the lens and the standard projection depth to reconstruct the projection relationship of the user's three-dimensional posture on a standard reference plane. The system calculates the projection offset value of the user's current posture on the standard plane based on these parameters.

[0024] The system further sets the maximum projection offset value of the user's current posture on the standard plane. This value is a threshold determined by statistically analyzing the posture projection range of a large number of standard movements in each frame, typically using the maximum projection deviation within the 95% confidence interval of the standard movement samples as this threshold. When the system detects that the projection offset value of the user's current posture on the standard plane exceeds the maximum projection offset value, it determines that the current posture deviates from the standard trajectory and triggers a posture correction prompt. The correction prompt is presented graphically on the user's terminal device, including an arrow indicating the correction direction, guiding the user to adjust their limb position in the expected direction, completing precise correction, and forming a dynamic posture training closed loop.

[0025] By integrating AI model evaluation, real-time sensor monitoring, and dynamic feedback mechanisms, this system effectively monitors the user's movement standardization, promptly identifies and alerts users to incorrect movements, helping to reduce the risk of sports injuries and improve training quality and efficiency. The system supports the import of personalized standard templates, adapting to different sports and individual user differences, enabling intelligent and customized movement correction. Simultaneously, it employs a combination of visual projection and skeletal point analysis to accurately quantify and provide feedback on user posture errors, enhancing the targetedness and operability of posture correction.

[0026] The data acquisition module collects user's electromyographic activity signals based on sensors and performs preprocessing to generate raw data sequences. This includes dividing the electromyographic activity signals into multiple fixed-length time periods, using template matching for each segment to determine whether it is a known waveform repeated, accumulating the number of detected repetitive electromyographic patterns, and using the ratio of the accumulated number of detected repetitive electromyographic patterns to the total number of sampling segments in the acquisition window as the signal repeatability score. An abnormal repeatability threshold is set. If the signal repeatability score is greater than the abnormal repeatability threshold, it indicates the presence of muscle tremor or signal grounding problems.

[0027] In the AI-based intelligent body training management system, the data acquisition module collects user electromyographic (EMG) signals from sensors and preprocesses them to generate raw data sequences, including dividing the EMG signals into multiple fixed-length time segments. Signal segments within each time segment are compared using a preset template matching method to determine if there are repeated occurrences of known waveforms. The template matching method can employ a sliding window and correlation coefficient calculation approach, comparing each signal segment one by one with a set of defined standard EMG waveform templates. If the similarity exceeds a predetermined threshold, the segment is determined to be a repeated waveform. The system processes all time segments within the sampling window, accumulating the number of detected repetitive EMG patterns as an indicator parameter for repetitive behavior. The ratio of the accumulated number of detected repetitive EMG patterns to the total number of sampling segments within the acquisition window is calculated to obtain a signal repeatability score. This score reflects the degree of signal repetition in the current acquisition window and has the ability to quantitatively describe periodic tremors or abnormal grounding interference. The system sets an abnormal repetition threshold, which is used to determine whether the signal repeatability score is abnormally high. The abnormal repetition threshold can be obtained based on the statistical distribution of repetition rates from a large amount of user electromyography (EMG) data under normal training conditions. Typically, statistical methods are used to extract the upper limit of the repetition score distribution, for example, setting the threshold at the mean plus two standard deviations (mean + 2σ) of the normal distribution. Alternatively, a 95% confidence interval boundary can be set based on a 5% anomaly rate. When the signal repetition score exceeds the abnormal repetition threshold, the system determines that there may be abnormal repetition in the current EMG signal, indicating a possible muscle tremor or signal grounding problem. Such prompts can trigger further signal quality checks or prompt the user to adjust electrode attachment and body position to eliminate potential interference and improve data quality and assessment reliability.

[0028] This invention introduces template matching and signal repeatability scoring mechanisms into muscle electrical activity signals to effectively identify and warn of abnormal repetitive patterns in the signals (such as muscle tremors or grounding problems), significantly improving signal quality control capabilities during the data acquisition phase and enhancing system stability and the accuracy of motion assessment.

[0029] The feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device. This includes determining the position coordinates of the target muscle group that needs to trigger feedback, obtaining the current feedback point position of the user's wearable device, calculating the distance between the feedback point position and the target muscle group position coordinates, calculating the intensity of the perceived stimulus by the user, setting a threshold for the intensity of the perceived stimulus by the user, and adjusting the feedback device or changing the feedback method when the intensity of the perceived stimulus by the user is less than the threshold.

[0030] In the AI-based intelligent body training management system, the feedback generation module generates feedback information in real time based on the results of the movement standardization assessment and transmits it to the user's terminal device. During the feedback information generation process, the feedback generation module first determines the position coordinates of the target muscle group that needs to trigger feedback. These coordinates can be located using muscle movement tags and skeletal point data associated with the AI ​​model in the system. Subsequently, the feedback generation module obtains the current feedback point position of the user's wearable device, i.e., the actual attachment coordinates of the current wearable feedback device (such as a vibration motor or electrical pulse patch), obtained through sensors or user-defined parameters. The system calculates the distance between the feedback point position and the target muscle group's position coordinates to determine whether the stimulation signal accurately acts on the desired area. Combining this offset distance with the feedback device's output parameters, the system calculates the user-perceived stimulation intensity. This intensity comprehensively considers factors such as feedback type (e.g., mechanical stimulation, electrical stimulation), signal amplitude, and skin sensitivity in the feedback area to construct a standardized perceived stimulation scoring model. The system sets a user-perceived stimulation intensity threshold to determine whether the current feedback is sufficient for the user to perceive. This threshold is obtained statistically based on skin sensitivity data of different users and the minimum identifiable stimulus value recorded in multiple rounds of trials. For example, the average threshold perceived by 80% of users in the experimental group can be used as the default setting, or it can be fine-tuned in combination with individual settings. When the intensity of the perceived stimulus is detected to be less than the threshold, the system automatically adjusts the parameters of the feedback device (such as vibration intensity and electrical stimulation frequency) or switches to other types of feedback (such as switching from mechanical vibration to voice prompts) to ensure that the feedback information is effectively conveyed to the user, thereby improving the responsiveness of the interaction and the user's training participation.

[0031] By introducing a target muscle group localization and feedback point location matching mechanism, this invention can accurately identify the specific areas requiring feedback during user training, achieving spatially directional transmission of feedback information. Combined with real-time monitoring of the feedback point location and stimulus intensity on the user's wearable device, the system can dynamically adjust feedback parameters to ensure the feedback signal is fully perceptible to the user. Setting a user-perceived stimulus intensity threshold enables the system to adaptively adjust the feedback method, automatically triggering device adjustments or feedback format switching when the user fails to effectively perceive feedback, effectively improving the accuracy of feedback response and the practicality of training guidance. This mechanism enhances the system's adaptability to various training scenarios and individual differences, strengthens human-computer interaction, and further improves the intelligence level and user experience of the body training management system.

[0032] The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results. The specific steps for determining the action violation probability include subtracting one hundred from the matching score between each action output by the AI ​​model and the standard template to obtain the action violation probability.

[0033] In the AI-based body training management system, the action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results. The specific steps for determining the probability of action violation include subtracting one hundred from the matching score between each action and the standard template output by the AI ​​model to obtain the action violation probability. Specifically, the raw data sequence is acquired and preprocessed by the data acquisition module and then input into the pre-trained AI model. This AI model analyzes the time series features and electromyographic pattern features between the action data sequence and the preset standard action template, outputting a matching score. This matching score is a value between 0 and 100, representing the similarity of each action to the standard template; a higher score indicates a more standard action. To quantify the degree of inconsistency between the action and the standard action, the action evaluation module subtracts one hundred from the matching score, i.e., using the formula: Action Violation Probability = 100 - Matching Score, thus converting the matching score inversely into a probability index representing the possibility of a violation. The higher the action violation probability value, the greater the deviation of the action from the standard action. The process of obtaining the probability of action violations is clear and controllable, and is applicable to subsequent action recording, abnormal prompts and feedback generation and processing modules, forming an important basis for action recognition and quality control in the whole system.

[0034] By subtracting one hundred from the matching score between each action output by the AI ​​model and the standard template in the action evaluation module to obtain the action violation probability, the system directly quantifies the degree of deviation of user actions. This concise calculation logic not only enhances the system's computational efficiency and real-time processing capabilities but also avoids accuracy shifts or computational delays caused by complex function transformations. After quantifying the violation probability, the system can perform more granular training quality control, such as hierarchical feedback and individualized teaching suggestions, significantly improving the interpretability and operability of the AI ​​model evaluation results. This mechanism helps trainees quickly understand their own action deviations, develop an immediate corrective mindset, and thus promote the optimization of training effects. Furthermore, as a standardized indicator, the action violation probability facilitates horizontal comparisons between different users, different time periods, or different action types, improving the consistency of data analysis and training evaluation, as well as system adaptability, enhancing the practical value and versatility of this invention.

[0035] The feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device. The specific steps for calculating the user's perceived stimulus intensity include calculating the three-dimensional distance between the position coordinates of the target muscle group that needs to trigger feedback and the current feedback point position of the user's wearable device, using an exponential decay operation, and inputting the calculated three-dimensional distance as the negative part of the exponential operation. The calculation result is used as the current user's perceived stimulus intensity.

[0036] In the AI-based intelligent body training management system, the feedback generation module generates feedback information in real time based on the motion standardization assessment results and transmits it to the user's terminal device. The specific steps for calculating the user's perceived stimulus intensity include: calculating the three-dimensional distance between the position coordinates of the target muscle group requiring feedback and the current feedback point position on the user's wearable device; using an exponential decay operation, the calculated three-dimensional distance is input as the negative part of the exponential operation, and the result is used as the current user's perceived stimulus intensity. Specifically, the system first identifies the position coordinates of the target muscle group requiring feedback intervention in the current movement through the motion standardization assessment results, and obtains the corresponding position coordinates by combining the user's skeletal point data identified by the posture adjustment module or image recognition system. Simultaneously, by detecting the feedback point position on the user's wearable device, its current three-dimensional spatial coordinates attached to the body surface are obtained. The feedback generation module calculates the Euclidean distance between the two three-dimensional position coordinates to obtain the three-dimensional distance between the feedback point and the target muscle group. This distance value is input into an exponential decay function, used as the negative part of the exponent. The specific calculation expression is: User-perceived stimulus intensity = A × exp(-k × d), where d is the three-dimensional distance, A is the maximum stimulus value, and k is the decay coefficient. This function expresses the natural decay characteristics of the feedback signal in space; that is, the farther away from the target muscle group, the weaker the user's actual perception. This calculation result is used by the system as an estimate of the current user-perceived stimulus intensity, participating in the determination of whether the "user-perceived stimulus intensity threshold" has been reached, and is used for subsequent feedback method adjustments and reach strategy optimization.

[0037] A represents the strongest feedback stimulus value that the user can perceive when the feedback point location completely coincides with the target muscle group location, i.e., the three-dimensional distance is 0. This value is the upper limit of the perceived stimulus intensity, which usually depends on the maximum output capability of the feedback device (such as the maximum vibration amplitude, the maximum electrical pulse intensity, etc.) and the maximum acceptable value of the human body for this type of stimulus.

[0038] k is a coefficient that controls the rate of stimulus intensity decay, determining how quickly the perceived stimulus intensity decreases as the distance between the feedback point and the target muscle group increases. The larger k is, the more rapidly the stimulus intensity weakens with increasing distance; the smaller k is, the more gradually the stimulus decay process occurs.

[0039] Methods for determining A: Experimental method: In human trials, the output intensity of the feedback device is gradually increased, and the maximum stimulus value that most users can clearly perceive without discomfort when close to the target muscle group is recorded as the upper limit of A. Device specification setting: The value of A is set according to the factory parameters or design limit output capacity of the feedback device to ensure the maximum perceived intensity within a safe range. Methods for determining k: Model fitting: By collecting user perception rating data of feedback stimuli at different three-dimensional distances, the optimal decay curve is fitted using an exponential function model, and the optimal k value is determined by regression. Statistical strategy: Select a k value that allows more than 90% of users to still perceive the feedback when the distance between the feedback point and the target muscle group is less than a set threshold (e.g., 5cm), ensuring the effectiveness of feedback coverage in actual training. Task sensitivity adjustment: The k value is dynamically adjusted according to the accuracy requirements of the training task. For example, a larger k is set in high-precision tasks to enhance the positional accuracy requirements.

[0040] By introducing an exponential attenuation calculation method based on the three-dimensional distance between the target muscle group location and the feedback point location, this invention achieves quantitative estimation of the user-perceived stimulus intensity, effectively reflecting the spatial correlation between the feedback signal and the body part. This method can intuitively and accurately describe the stimulus attenuation effect caused by the feedback point deviating from the target muscle group, providing a calculable basis for the system to adjust the feedback intensity and form.

[0041] By employing an exponential decay model, the system can naturally simulate the nonlinear decrease in perceived stimulus with increasing distance, improving the physiological rationality of feedback control and the accuracy of human-computer interaction. This mechanism helps ensure that feedback signals are correctly perceived, thereby enhancing training guidance, reducing false prompts and user misjudgments, and improving the system's intelligent response capabilities and user experience consistency.

[0042] The posture adjustment module guides users to adjust their posture based on feedback information to avoid sports injuries caused by improper posture. The specific steps for calculating the projection offset value of the user's current posture on the standard plane include multiplying the lateral displacement value of the skeletal point of the user's actual movement with the distance of the hand from the lens and the standard projection depth, and dividing the result by the actual straight-line distance of the skeletal point and the current distance from the camera lens to obtain the projection offset value of the user's current posture on the standard plane.

[0043] In the AI-based intelligent body training management system, the posture adjustment module guides users to adjust their postures based on feedback information to avoid sports injuries caused by improper posture. The specific steps for calculating the projection offset value of the user's current posture on the standard plane include multiplying the lateral displacement value of the user's actual skeletal points by the distance between the hand and the camera lens and the standard projection depth. The result is then divided by the actual straight-line distance between the skeletal points and the camera lens to obtain the projection offset value of the user's current posture on the standard plane. Specifically, the system first extracts the lateral displacement value of the user's actual skeletal points from the image capture device (such as an RGB-D camera), representing the degree of lateral deviation of key skeletal points relative to the standard centerline during the user's movement. Simultaneously, the system obtains the distance between the hand and the camera lens, representing the current distance between the hand's spatial position and the camera, as well as the system-preset standard projection depth, which is determined based on the spatial reference plane of the standard movement in the training environment. The system then multiplies the lateral displacement value of the user's skeletal points by the distance between the hand and the camera lens and the standard projection depth. The result reflects the projection characteristics of the movement in real space after being affected by depth. The product is then divided by the skeletal point and the actual straight-line distance from the camera lens to eliminate distortion errors caused by changes in spatial depth, thereby accurately calculating the projection offset value of the user's current posture on the standard plane. This projection offset value serves as an important parameter for posture assessment, used to determine whether there is a deviation in the current posture and whether a posture adjustment prompt should be triggered.

[0044] This invention proposes a method for accurately quantifying user posture projection offset by multiplying the lateral displacement value of skeletal points, the distance of the hand from the camera, and the standard projection depth, combined with standardized division of the actual straight-line distance between the skeletal points and the camera. This calculation logic effectively eliminates image distortion interference introduced by different depth levels, improving the stability and accuracy of posture recognition in three-dimensional space. The system can evaluate the deviation between the user's current posture and the standard posture in spatial projection in real time, enabling the posture adjustment module to have stronger targeting and data support when correcting actions, effectively reducing misjudgments and feedback delays, and further improving user training safety and correction efficiency. This solution provides key basic data for AI-assisted posture correction, enhancing the intelligent recognition and dynamic guidance capabilities of the entire system.

[0045] The data acquisition module, which collects user muscle electromyography signals based on sensors and performs preprocessing to generate raw data sequences, also includes: The sampling frequency from which electromyographic signals are extracted from the sensor; After sampling is completed, the absolute values ​​of all sampling points are counted, and the absolute values ​​of the sampling points are summed. The average value is obtained by dividing the total number of sampling points by the sum of the absolute values ​​of the sampling points. The original signal is input into a bandpass filter module, which retains components between 20 Hz and 500 Hz. Iterate through each sampling point and determine whether the absolute amplitude of the sampling point is higher than the average level of the overall signal.

[0046] In the AI-based intelligent body training management system, the data acquisition module, based on sensors, collects user muscle electromyography (EMG) activity signals and performs preprocessing to generate raw data sequences. This process includes: First, extracting the sampling frequency of the EMG signals from the sensors to ensure subsequent signal analysis is performed on a unified time reference. Then, after sampling, the absolute values ​​of all sampling points are counted and summed as a basis for measuring overall signal strength. Next, the total number of sampling points is divided by the sum of their absolute values ​​to obtain the average value, which serves as a reference for subsequent signal characteristic assessment. Then, the raw signal is input into a bandpass filter module. This module retains components between 20Hz and 500Hz, filtering out low-frequency noise (such as motion artifacts) and high-frequency interference (such as power supply frequency and high-frequency electrical noise), extracting effective frequency band signals related to muscle activation. Finally, the system iterates through each sampling point to determine whether the absolute amplitude of the sampling point is higher than the average level of the overall signal. This step is used to initially identify possible muscle activation peaks, which helps in the accurate execution of subsequent EMG pattern extraction, motion assessment, and other processes.

[0047] This invention establishes a standardized and efficient electromyography (EMG) signal preprocessing workflow by combining sampling frequency, absolute signal value processing, average value calculation, and filtering mechanisms, thereby improving the quality of raw data and the stability of analysis. By extracting the sampling frequency of the EMG signal and performing statistical analysis on the absolute values ​​of the sampling points, the system achieves basic quantification of signal intensity, providing a reference for subsequent signal discrimination. Combined with a bandpass filtering module that retains components between 20Hz and 500Hz, the system significantly reduces the interference of low-frequency drift and high-frequency noise on the EMG signal, improving the signal's specificity and usability. By traversing the sampling points and determining whether their absolute amplitude is higher than the overall signal average, potential high-activity regions can be quickly screened, providing accurate input for EMG pattern recognition and motion segmentation. This scheme enhances the system's ability to extract EMG signal features, improves the robustness and accuracy of motion recognition, and ultimately enhances the intelligence and evaluation reliability of the entire body training management system.

[0048] The process involves iterating through each sampling point and determining whether the absolute amplitude of the sampling point is higher than the average level of the overall signal. Specifically, if the absolute amplitude of the sampling point exceeds the average amplitude multiplied by a pre-set multiple threshold, where the multiple threshold is between 2 and 3, the sampling point is considered a real active signal. If the absolute amplitude of the sampling point does not exceed the average amplitude multiplied by a pre-set multiple threshold, the sampling point is considered an artifact or invalid noise and is automatically discarded or ignored.

[0049] In the AI-based intelligent body training management system, the specific steps of traversing each sampling point and determining whether the absolute amplitude of the sampling point is higher than the average level of the overall signal include: if the absolute amplitude of the sampling point exceeds the average amplitude multiplied by a pre-set multiple threshold, where the multiple threshold is between 2 and 3, the sampling point is considered a true activation signal; if the absolute amplitude of the sampling point does not exceed the average amplitude multiplied by a pre-set multiple threshold, the sampling point is considered an artifact or invalid noise and is automatically eliminated or ignored. Specifically, the system first obtains the absolute amplitude of all sampling points according to the aforementioned steps and calculates the average level of the overall signal (i.e., the average amplitude). When traversing each sampling point, its absolute amplitude is compared with the average amplitude multiplied by a pre-set multiple threshold. If the absolute amplitude of the sampling point is greater than the product value, the system considers it to deviate significantly from the average amplitude, which meets the characteristics of a muscle activation event, and therefore marks it as a true activation signal. Otherwise, the system determines that the signal amplitude is within the normal fluctuation range or is caused by interference, and marks it as an artifact or invalid noise, performing automatic elimination or ignoring to avoid noise interfering with subsequent electromyography pattern analysis and AI evaluation.

[0050] The multiple threshold is a proportional criterion used to distinguish between real activation signals and background noise. Its value is set between 2 and 3, an empirical range selected based on statistical and physiological experimental analysis. Specific methods for determination include: Statistical outlier analysis: In a large amount of normal training electromyography (EMG) data, the amplitude distribution is statistically analyzed, using the standard deviation as the unit, and an amplitude 2-3 times the average level is set as the active signal threshold. Experimental verification: By comparing manually labeled or expert-scored real activation signal samples with background artifact data, ROC curve analysis is used to determine the multiple range corresponding to the optimal balance between sensitivity and specificity.

[0051] By introducing a multiplier threshold judgment mechanism, this invention effectively distinguishes true activation signals from noise and artifacts in electromyography (EMG) signals, improving the quality control capability of sampled data. The absolute amplitude of each sample point is compared with the overall average level multiplied by a pre-set multiplier threshold to form a clear discrimination rule. Noise removal can be completed without relying on complex models, enhancing the system's computational efficiency and real-time performance. The multiplier threshold range is set between 2 and 3, balancing the sensitivity and robustness of signal recognition, effectively reducing the false positive rate and ensuring that key muscle activation events can be accurately extracted for subsequent evaluation. Simultaneously, this mechanism provides an adjustable parameter basis for the system under different user physiological differences and different exercise scenarios, improving the system's adaptability and intelligence, and further enhancing the accuracy of motion recognition and the reliability of training management.

[0052] The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results, which also include: Feature extraction is performed on the preprocessed electromyographic (EMG) signals, which is divided into two dimensions: the first dimension is time features, which reflect the characteristics of the EMG signals in the time series, including mean, variance, waveform length, and zero crossover rate; the second dimension is frequency features, which involves analyzing the spectral composition of the EMG signals through Fourier transform or wavelet transform to extract peak frequency, average frequency, and frequency band energy parameters.

[0053] In the AI-based intelligent body training management system, the motion evaluation module inputs the raw data sequence into a pre-trained AI model and outputs corresponding motion standardization evaluation results. This also includes feature extraction of the pre-processed electromyographic (EMG) signals, divided into two dimensions: The first dimension is time features, reflecting the characteristics of the EMG signals in the time series, including mean, variance, waveform length, and zero-crossing rate. The mean measures the overall average potential level of the signal; variance reflects the intensity of signal fluctuations; waveform length represents the total amplitude length of the signal change path; and the zero-crossing rate assesses the frequency of signal changes in the positive and negative intervals, reflecting the rate of muscle contraction. The second dimension is frequency features, which involves analyzing the spectral composition of the EMG signals using Fourier transform or wavelet transform. Frequency domain analysis can extract important parameters reflecting muscle frequency behavior, such as peak frequency (the frequency corresponding to the maximum power in the spectrum), average frequency (the power-weighted average frequency in the spectrum), and band energy parameters (the sum of signal energy within a specific frequency band). These features reflect differences in muscle activity, fatigue trends, and movement types. All the aforementioned time and frequency dimensions of features will be fed into the pre-trained AI model as input feature vectors to achieve a comprehensive evaluation of the standardization of user actions, generate action standardization evaluation results, and provide a basis for subsequent feedback generation and posture adjustment.

[0054] By introducing a feature extraction mechanism that combines temporal and frequency features, this invention achieves a more comprehensive and detailed representation of electromyographic signals in the action evaluation module, significantly improving the accuracy and stability of action recognition and evaluation. Temporal features directly reflect the changing patterns of muscle electrical activity in the time domain during user actions, while frequency features reveal the spectral structure hidden behind signal fluctuations, supplementing deeper information that is difficult to capture in the time domain. The feature extraction process combines Fourier transform or wavelet transform, taking into account both global and local frequency variation features, enabling the system to have stronger generalization ability and robustness when handling differences in action intensity, rhythm, and muscle group. By inductively inputting these high-value features into the AI ​​model, the action standardization evaluation results become more reliable and interpretable, further improving feedback response accuracy and user training quality. This scheme enhances the input foundation of the AI ​​model, improves the data processing depth and evaluation intelligence of the training system, and is conducive to building a high-performance, low-error human-computer interaction training platform.

[0055] The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results, which also include: The time features and frequency features are weighted and fused according to their respective weights. The time features are multiplied by a time weight coefficient, and the frequency features are multiplied by a frequency weight coefficient. Then, the result of multiplying the time features by a time weight coefficient and the result of multiplying the frequency features by a frequency weight coefficient are added together to obtain the comprehensive feature value. The comprehensive feature values ​​are input into the AI ​​model, which outputs a standard action score. The standard action score is set between zero and one. If the standard action score is less than the preset standard threshold, the action is judged as a non-standard action, and the user is prompted to adjust their posture.

[0056] In the AI-based intelligent body training management system, the action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation result. This includes: weighted fusion of time features and frequency features according to weights; multiplying the time feature by a time weight coefficient and the frequency feature by a frequency weight coefficient; and then adding the result of multiplying the time feature by a time weight coefficient to the result of multiplying the frequency feature by a frequency weight coefficient to obtain a comprehensive feature value. The time and frequency features originate from a two-dimensional feature extraction process, reflecting the temporal behavior and frequency structure of electromyographic signals, respectively. The time weight coefficient measures the importance of the time feature in the model input, and the frequency weight coefficient measures the importance of the frequency feature; both are positive real numbers and adjustable. Through the above weighted fusion, the system obtains a comprehensive feature value that includes both the time change trend and frequency pattern characteristics, improving the expressive power of the input features and the model's adaptability. The comprehensive feature value is input into the AI ​​model, outputting an action standardization score. The action standardization score is set between zero and one, representing the degree of similarity matching between the current action and the standard action. If the standardization score of the movement is less than the preset standard threshold, the system will judge the movement as a non-standard movement and prompt the user to adjust the posture to complete the standardization training.

[0057] **Time Weighting Coefficient:** Measures the relative weight of time features in action recognition. It can be determined through feature importance analysis during model training (e.g., Shapley value, information gain) or manually set for task adaptation. **Frequency Weighting Coefficient:** Measures the contribution of frequency features to model prediction. Its determination method is similar to the time weighting coefficient, and it can also be flexibly adjusted according to different sports. **Action Standardization Score:** A normalized score output by the AI ​​model, ranging from 0 to 1, used to express the degree of standardization matching of the action. A higher value indicates closer proximity to the standard action template. **Preset Standard Threshold:** A critical value used to distinguish between standard and non-standard actions. It is usually determined through historical data analysis, such as using the lower quartile of the training set scores as the threshold, or using an ROC curve to set the optimal distinction point.

[0058] By weightedly fusing temporal and frequency features to form a unified comprehensive feature value input to the AI ​​model, this invention achieves efficient integration of multi-dimensional information, improving the accuracy and generalization ability of action evaluation. The introduction of time and frequency weight coefficients allows the system to dynamically adjust the model input based on feature contributions in different training scenarios, effectively enhancing the model's adaptability to individual and action category differences. Employing a standardized action scoring mechanism, the output score is standardized to the 0-1 range, offering advantages such as strong intuitiveness and clear discrimination boundaries, facilitating user understanding and system decision-making. Combined with a preset standard threshold mechanism, the system can accurately identify non-standard actions and provide timely corrective feedback, improving the safety, standardization, and intelligence of training.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A body training management system based on AI intelligence, characterized in that, The system includes: The data acquisition module is used to collect the user's muscle electrical activity signals based on sensors and to preprocess them to generate raw data sequences; The action evaluation module, connected to the data acquisition module, is used to input the raw data sequence into the pre-trained AI model and output the corresponding action standardization evaluation results, including the matching score between each action and the standard template output by the AI ​​model, determining the action violation probability, setting the action violation probability threshold, and recording and correcting actions whose action violation probability exceeds the action violation probability threshold. The feedback generation module, connected to the motion assessment module, is used to generate feedback information in real time based on the motion standardization assessment results and transmit it to the user's terminal device. The posture adjustment module, connected to the feedback generation module, guides users to adjust their postures based on feedback information to avoid sports injuries caused by improper postures. This includes obtaining the lateral displacement value of the skeletal points of the user's actual movement, obtaining the distance between the hand and the camera and the standard projection depth, calculating the projection offset value of the user's current posture on the standard plane, setting the maximum projection offset value of the user's current posture on the standard plane, and prompting the user to correct the movement if the projection offset value exceeds the maximum projection offset value, and displaying a correction direction arrow.

2. The AI-based intelligent body training management system according to claim 1, characterized in that: The data acquisition module collects user electromyographic activity signals from sensors and performs preprocessing to generate raw data sequences. This includes dividing the electromyographic activity signals into multiple fixed-length time periods, using template matching for each period to determine whether a known waveform is repeated, accumulating the number of detected repetitive electromyographic patterns, and using the ratio of the accumulated number of detected repetitive electromyographic patterns to the total number of sampling segments in the acquisition window as the signal repeatability score. An abnormal repeatability threshold is set. If the signal repeatability score is greater than the abnormal repeatability threshold, it indicates the presence of muscle tremor or signal grounding problems.

3. The AI-based intelligent body training management system according to claim 1, characterized in that: The feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device. This includes determining the position coordinates of the target muscle group that needs to trigger feedback, obtaining the current feedback point position of the user's wearable device, calculating the distance between the feedback point position and the target muscle group position coordinates, calculating the user's perceived stimulus intensity, setting a user's perceived stimulus intensity threshold, and adjusting the feedback device or changing the feedback method when the user's perceived stimulus intensity is less than the user's perceived stimulus intensity threshold.

4. The AI-based intelligent body training management system according to claim 1, characterized in that: The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results. The specific steps for determining the action violation probability include subtracting one hundred from the matching score between each action output by the AI ​​model and the standard template to obtain the action violation probability.

5. The AI-based intelligent body training management system according to claim 1, characterized in that: The feedback generation module generates feedback information in real time based on the action standardization assessment results and transmits it to the user's terminal device. The specific steps for calculating the user's perceived stimulus intensity include calculating the three-dimensional distance between the position coordinates of the target muscle group that needs to trigger feedback and the current feedback point position of the user's wearable device, using an exponential decay operation, inputting the calculated three-dimensional distance as the negative part of the exponential operation, and using the calculation result as the current user's perceived stimulus intensity.

6. The AI-based intelligent body training management system according to claim 1, characterized in that: The posture adjustment module guides users to adjust their posture based on feedback information to avoid sports injuries caused by improper posture. The specific steps for calculating the projection offset value of the user's current posture on the standard plane include multiplying the lateral displacement value of the skeletal point of the user's actual movement with the distance of the hand from the lens and the standard projection depth, and dividing the multiplication result by the actual straight-line distance of the skeletal point and the current distance from the camera lens to obtain the projection offset value of the user's current posture on the standard plane.

7. The AI-based intelligent body training management system according to claim 1, characterized in that: The data acquisition module, which collects user electromuscular activity signals based on sensors and performs preprocessing to generate raw data sequences, also includes: The sampling frequency from which electromyographic signals are extracted from the sensor; After sampling is completed, the absolute values ​​of all sampling points are counted, and the absolute values ​​of the sampling points are summed. The average value is obtained by dividing the total number of sampling points by the sum of the absolute values ​​of the sampling points. The original signal is input into a bandpass filter module, which retains components between 20 Hz and 500 Hz; Iterate through each sampling point and determine whether the absolute amplitude of the sampling point is higher than the average level of the overall signal.

8. The AI-based intelligent body training management system according to claim 7, characterized in that: The specific steps of traversing each sampling point and determining whether the absolute amplitude of the sampling point is higher than the average level of the overall signal include: if the absolute amplitude of the sampling point exceeds the average amplitude multiplied by a preset multiple threshold, where the multiple threshold is between 2 and 3, then the sampling point is considered a real active signal; if the absolute amplitude of the sampling point does not exceed the average amplitude multiplied by a preset multiple threshold, then the sampling point is considered an artifact or invalid noise and is automatically discarded or ignored.

9. The AI-based intelligent body training management system according to claim 1, characterized in that: The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results, which also include: Feature extraction is performed on the preprocessed electromyographic (EMG) signals, which is divided into two dimensions: the first dimension is time features, which reflect the characteristics of the EMG signals in the time series, including mean, variance, waveform length, and zero crossover rate; the second dimension is frequency features, which involves analyzing the spectral composition of the EMG signals through Fourier transform or wavelet transform to extract peak frequency, average frequency, and frequency band energy parameters.

10. A body training management system based on AI intelligence according to claim 9, characterized in that: The action evaluation module inputs the raw data sequence into a pre-trained AI model and outputs the corresponding action standardization evaluation results, which also include: The time features and frequency features are weighted and fused according to their weights. The time features are multiplied by a time weight coefficient, and the frequency features are multiplied by a frequency weight coefficient. Then, the result of multiplying the time features by a time weight coefficient and the result of multiplying the frequency features by a frequency weight coefficient are added together to obtain the comprehensive feature value. The comprehensive feature values ​​are input into the AI ​​model, which outputs a standard action score. The standard action score is set between zero and one. If the standard action score is less than the preset standard threshold, the action is judged as a non-standard action, and the user is prompted to adjust their posture.

Images