Patent Application
20240001196
This application claims priority to, and the benefit of, China Patent Application No. 202210781291.7 filed on Jul. 4, 2022, the disclosure of which is hereby incorporated by reference in its entirety.
The present invention relates to an artificial intelligence (AI) assisted personal training system, a personal training device and a control device.
The benefits of exercise are considered to “solve one of the biggest public health problems of the 21st century” and are receiving increasing public attention worldwide. Exercise is Medicine® is a global health initiative administered by the American College of Sports Medicine (ACSM) that recognizes the importance of physical activity for optimal health and encourages physicians and other healthcare providers to provide patients with physical activities when designing the patients' treatment plans. In fact, more and more people are trying to incorporate exercise into their daily lives and develop the habit of regularly doing exercise. According to certain reports, between 2009 and 2017, the number of gymnasiums in Hong Kong increased from 548 to 743 (a growth rate of about 35.6%). Resistance training (also known as strength training or weight training) has rapidly become an emerging fitness trend in recent years. Resistance training is a form of physical activity designed to increase strength, anaerobic endurance and skeletal muscle size through muscle contraction against resistance.
However, sports may lead to injuries. The main reasons may include: insufficient training level, poor physical fitness, incorrect movement, lack of self-protection ability, no preparatory activity or insufficient preparatory activity before doing exercise, lack of training to adapt to the environment, and improper organization of teaching and competition works. In 2018, sports and equipment-related injuries ranked the second among all injury types in the U.S. for 25- to 44-year-olds. Proper posture and use of muscles during athletic activities not only maximizes training efficiency, but also minimizes additional stress on ligaments and joints, thereby reducing the risk of injury.
There are already smart clothing or wearables on the market that can monitor heart rate, posture, body temperature, stress or stress levels when the user is doing exercise. Table 1 and
The prior-art products described above are shown in
The Athos training system from Mad Appareal in the United States includes smart clothing that can be adapted to any type of training, such as T-shirts, shorts, and tights with embedded sEMG. This training system can provide feedbacks to users through a mobile app that is only compatible with iOS, and provide reports after training sessions as well as weekly reports. However, it requires the coach/athlete to read the report or weekly report, to manually assess the wearer's or user's performance and progress or to provide real-time performance feedback.
Hexoskin from Cane Technologies of Montreal, Canada is a smart garment that can be used to continuously monitor heart, lung, activity and sleep data. It includes smart sports vests, jerseys and body condition sensors. It is equipped with sensors to monitor heartbeat, breathing and body condition, and is suitable for data tracking during exercise and during sleep, helping users understand their breathing rates, heart rates, sweating behavior, etc. It can also calculate the number of footsteps taken every day, the amount of calories burned, etc. The device has a companion mobile app (APP), but is also compatible with most popular third-party health apps. It still requires professionals (such as fitness trainers) to interpret the collected data to provide feedbacks on the user's exercise performance, exercise intensity, and exercise posture.
The Nadi X yoga pants from Wearable X in Sydney, Australia are smart clothing that can monitor movement posture and can be connected to a mobile phone to provide voice guidance. Nadi X yoga pants have sensors on the lower back, knees, feet, etc., and are powered by a battery above the left knee to accurately collect user data. Nadi X yoga pants can be connected to the exclusive App via Bluetooth, and after the connection, the user's training progress can be tracked. Users can choose the yoga difficulty that suits them to start training. Nadi X yoga pants collect the user's posture through the above sensors. When the user's yoga posture is not standard, the sensor will vibrate to different degrees, and the hips, knees and ankles will vibrate softly. There are ways to improve posture and remind users to adjust to the correct movement, thus helping users to adopt more correct yoga movements and effectively achieve movements that may not have been possible in the past.
Despite availability of the above-mentioned prior-art products, current smart wearable technologies cannot satisfy more advanced physical exercises, such as insufficient performance and efficiency evaluation of advanced physical exercises, and cannot provide real-time feedbacks and suggestions through AI algorithms. The latest sEMG and sports wearable sensor technologies allow for simultaneous and accurate measurement of muscle activity, body posture and body movement. The most common smart wearable devices on the market, such as compression T-shirts, leggings, etc., are capable of measuring the user's body temperature, heart rate, heart rate variability, respiratory rate, calorie consumption, steps, and more. However, the biosignals provided by current wearable devices are insufficient to assess the performance and efficiency of advanced physical exercises, such as resistance exercise tasks. To the Inventors' knowledge, the Athos training system described in Table 1 is the only wearable technology on the market that can measure the activity of major muscle groups. However, its functionality is still premised on the biometric tracking phase.
Second, people who are physically active often consult personal fitness trainers to pursue more advanced training goals, such as resistance training. Professional fitness trainers have an extensive knowledge of muscle anatomy and a good understanding of the exact muscle groups that need to be activated and deactivated during different movement tasks. Most importantly, a qualified fitness trainer can provide immediate feedback on whether certain exercises/tasks have been performed correctly and provide guidance to correct mistakes, which is essential for building muscle. However, existing wearables are unable to provide precise real-time feedbacks and recommendations through AI algorithms.
The Inventors have found that with the help of the latest technologies in wearable sensors and AI methods, with proper data input and training, the knowledge and skills of fitness training trainers are likely to be replaced by AI. The present invention has been developed based on this observation.
A first aspect of the present invention is to provide a personal training device.
The training device comprises an upper garment part, a trousers part and plural sensor accommodating units. The upper garment part and trousers part are both used for wearing by a user. The sensor accommodating units are distributed on the upper garment part and trousers part. Plural sensors are installed in the sensor accommodating units such that an individual sensor accommodating unit is equipped with one or more of the sensors. Particularly, positions of the sensor accommodating units on the upper garment part and trousers part respectively correspond to locations of muscles of major muscle groups of the user such that the sensors, or electrodes thereof, of the sensor accommodating units are positioned on the corresponding muscle locations of the user when the user wears the personal training device, thereby allowing a posture of the user and a muscle activity of the main muscle groups to be monitored during the user doing an exercise.
In certain embodiments, the upper garment part and trousers part are separate garment articles, are collectively formed as a one-piece garment, or are formed from plural straps. The upper garment part and trousers part are tight-fitting or skin-tight.
In certain embodiments, the upper garment part, trousers part and sensor accommodating units are made of one or more fabrics. A main fabric selected among the one or more fabrics and used for forming the upper garment part and trousers part is tricot knitted. A main fabric selected among the one or more fabrics and used for forming the sensor accommodating units is warp-knitted stretch mesh.
In certain embodiments, the sensors include surface electromyography sensors and inertial measurement unit sensors. The muscles of the major muscle groups include upper trapezius, triceps, erector spinae, biceps femoris, pectoralis major, biceps, rectus abdominis, and rectus femoris.
In certain embodiments, the sensor accommodating units include 14 units for accommodating 16 sensors.
In certain embodiments, the individual sensor accommodating unit is realized as a pocket sewn, snap-attached, or affixed, to the upper garment part or the trousers part.
In certain embodiments, the pocket is a Type-1 pocket for accommodating a single sensor. The Type-1 pocket has a length and a width given by L1=(a+c)×p and W1=(b+2c)×p′ where: L1 is the length of the Type-1 pocket; W1 is the width of the Type-1 pocket; a, b and c are length, width and depth of the accommodated sensor, respectively; p is between 80% and 90% inclusively; and p′ is between 60% and 65% inclusively.
The Type-1 pocket may include an inner layer opening. The inner layer opening is rectangular, oval or square in shape, or conforms to a shape of the accommodated sensor. The inner layer opening has a length and a width given by L3=d+s and W3=e×3 where: L3 is the length of the inner layer opening; W3 is the width of the inner layer opening; d is a sum of lengths of all electrodes in the accommodated sensor; s is between 2 mm and 4 mm inclusively; and e is a common width of the electrodes.
In certain embodiments, the pocket is a Type-2 pocket for accommodating two sensors. The Type-2 pocket has a length and a width given by L2=(a+c)×p and W2=[(2b+2c)+q]×p′ where: L2 is the length of the Type-2 pocket; W2 is the width of the Type-2 pocket; a, b and c are length, width and depth of an individual accommodated sensor, respectively; p is between 80% and 90% inclusively; p′ is between 60% and 65% inclusively; and q is between 5 mm to 10 mm inclusively.
The Type-2 pocket may include an inner layer opening. The inner layer opening is rectangular, oval or square in shape, or conforms to a shape of the individual accommodated sensor. The inner layer opening has a length and a width given by L4=d+s and W4=e×3+f×2 where: L4 is the length of the inner layer opening; W4 is the width of the inner layer opening; d is a sum of lengths of all electrodes in the individual accommodated sensor; s is between 2 mm and 4 mm inclusively; and e is a common width of the electrodes.
In certain embodiments, the personal training device further comprises a wire opening located on an outside part of the pocket and spaced from a sensor insertion opening of the pocket by 1 cm to 2 cm. The wire opening is parallel to a side edge of the pocket or angled to a side of the pocket to facilitate placement of sensor electrodes through the wire opening.
The personal training device based on the AI-assisted wearable sensor developed by the present invention can simultaneously monitor the posture, muscle activity and muscle fatigue in real time during the user doing an exercise. Furthermore, according to the collected data, the personal training device uses an AI algorithm developed by the present invention to evaluate sporting performance, thereby providing real-time feedbacks to improve training efficiency and reduce injury risk.
A second aspect of the present invention is to provide a control device for processing data generated by a personal training device. The control device comprises an input unit, a database, a data preprocessing unit; a training analysis unit and an output unit. The input unit is used for receiving detected-signal data from sensors on the personal training device and inputting a user profile of a user of the personal training device. The database is used for storing the received detected-signal data and the user profile. The data preprocessing unit is used for cleaning and preprocessing the detected-signal data. The training analysis unit is used for executing an artificial intelligence algorithm to perform fatigue analysis on the user's movement and providing training load recommendations based on the user profile and the detected-signal data. An output unit is configured to output an estimated remaining number of repetitions of an exercise to be performed by the user, a training load suggestion given by the training analysis unit, or a number of repetitions of the exercise completed by the user.
In certain embodiments, the control device further comprises a coaching module having a graphical user interface (GUI). The GUI is used for displaying training information and real-time visual feedback of the user's movement, tutorial videos for different types of exercise as stored in the device's exercise library, as well as used for providing audio guidance and feedback on the exercise in real time.
In certain embodiments, the training analysis unit comprises a first deep neural network unit and a second deep neural network unit. The first neural network unit is configured to estimate the user movement according to a current muscle activation signal. The second deep neural network unit is configured to suggest an optimal training load to the user according to the current muscle activation signal and the estimated remaining number of repetitions.
In certain embodiments, the control device further comprises a posture detection model unit and a machine learning posture classifier. The posture detection model unit is configured to receive a signal of the user's human body detected from a camera input unit, and to generate a human body marker based on the detected signal of the user's human body. The machine learning gesture classifier is configured to calculate a vector representing the user's ongoing exercise program and a motion state from the human body marker generated by the posture detection model unit, whereby the control device is realized as an artificial intelligence fitness training system.
In certain embodiments, the data preprocessing unit includes a signal preprocessing unit for performing one or more of the following preprocessing functions on the detected-signal data: bandpass filtering; highpass filtering; lowpass filtering; root mean square calculation; moving average calculation; mean absolute value calculation; median frequency calculation; data segmentation; data normalization; and data anomaly identification.
A third aspect of the present invention is to provide a personal training system.
The personal training system comprises any of the embodiments of the personal training device and any of the embodiments of the control device.
In certain embodiments, the personal training system further comprises a machine learning server, an application server and a database server. The machine learning server is used for executing artificial intelligence algorithms for machine learning. The application server is used for storing one or more application programs of the control device, allowing the user to download the one or more application programs. The database server is used for storing collected data, providing data to the artificial intelligence algorithm and supporting data analysis.
Other aspects of the present disclosure are disclosed as illustrated by the embodiments hereinafter.
Skilled artisans will appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been depicted to scale.
As used herein, “upper garment part” means a clothing item, or a portion thereof, that is intended to be worn on a torso of a person. Examples of an upper garment part include a shirt, a sweater, a coat and a jacket, as well as an upper portion of a one-piece clinging garment that covers the person from his/her shoulder to ankles.
As used herein, “trousers part” means a clothing item, or a portion thereof, that is shaped like a pair of trousers and is intended to be worn below a waist of a person for covering at least a hip part of the person. Examples of a trousers part include a pair of trousers and a pair of shorts, as well as a lower portion of a one-piece clinging garment that covers the person from his/her shoulder to ankles.
Disclosed herein are a personal training device, a control device for processing data generated by the personal training device, and an AI-assisted personal training system that includes the personal training device and the control device.
The personal training device comprises an upper garment part, a trousers part, and plural sensor accommodating units. Both of the upper garment part and trousers part are used for wearing by a user. The sensor accommodating units are distributed on the upper garment part and trousers part. Plural sensors are installed in the sensor accommodating units such that an individual sensor accommodating unit is equipped with one or more of the sensors. Positions of the sensor accommodating units on the upper garment part and trousers part respectively correspond to locations of muscles of major muscle groups of the user such that the sensors, or electrodes thereof, of the sensor accommodating units are positioned on the corresponding muscle locations of the user when the user wears the personal training device. Thereby, it allows a posture of the user and a muscle activity of the main muscle groups to be monitored during the user doing an exercise.
The control device comprises an input unit, a database, a data preprocessing unit, a training analysis unit, and an output unit. The input unit is used for receiving detected-signal data from sensors on the personal training device and inputting a user profile of a user of the personal training device. The database is used for storing the received detected-signal data and the user profile. The data preprocessing unit is used for cleaning and preprocessing the detected-signal data. The training analysis unit is used for executing an artificial intelligence algorithm to perform fatigue analysis on the user's movement and providing training load recommendations based on the user profile and the detected-signal data. The output unit is configured to output an estimated remaining number of repetitions of an exercise to be performed by the user, a training load suggestion given by the training analysis unit, or a number of repetitions of the exercise completed by the user.
Specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
The AI-assisted personal training system of the present invention may include the wearable personal training device, the control device, and a server.
The wearable personal training device may have the shape of a garment to facilitate donning and doffing by a user. For example, the wearable personal training device includes a top (or T-shirt) and shorts; alternatively, a plurality of straps, such as waist belts, shoulder straps, girdle straps, leggings, shin guards, etc., which can be interconnected for ease of use. The wearable personal training device includes a plurality of sensors, and the plurality of sensors can be respectively mounted on the wearable personal training device through, for example, pockets or Velcro, so that after the user wears the wearable personal training device, the, the plurality of sensors or their electrodes are respectively located at positions corresponding to different muscle parts of the user.
The control device may be implemented as a stand-alone hardware device, or as application software to be installed in a mobile phone or other portable communication device. Alternatively, the control device may be a computer program product installed in a computer with a Windows or iOS operating system. The control device may be wired to the sensor via a wire, or wirelessly connected to the sensor via WiFi or Bluetooth, or the like. When the user performs exercise or exercises after putting on the wearable personal training device, the control device can simultaneously monitor the user's posture and muscle activity during the exercise through the plurality of sensors. Preferably, the monitoring may be performed in real time. The AI algorithm module built into the control device evaluates the user's training performance based on the sensor data, and can provide real-time feedback and suggestions to the user through a GUI and/or audio feedback.
The servers may include machine learning servers, application servers, and database servers. The Inventors have developed a centralized database with multiple uses that can: store collected data; provide data to AI algorithms; and support data analysis. All data collected, used and processed are stored in a database. The data can include, for example, user credentials, training records, sEMG signal readings, and some personalization parameters optimized by AI algorithms.
The individual components of the AI-assisted personal training system of the present invention will be described in detail below with reference to the accompanying drawings.
The wearable personal training device of the present invention is based on AI-assisted sensors. One embodiment of the wearable personal training device may include a top 1 and a pair of shorts 2 (or trousers), as shown in
The top 1 and the shorts 2 each have a plurality of sensor accommodating units, such as pockets 3, to accommodate up to 16 sensors, and the installation positions of these sensors can respectively correspond to the skin of up to 16 different body parts. The tops and shorts may also be constructed as a one-piece garment comprising an upper garment part and a trousers part, or in other styles, such as multiple straps 4 for easy attachment and donning, with pockets, locks or Velcro Removably mounts the sensor and is easy to wear on the user. Alternatively, the sensor may be fixed to the personal training device rather than being detachable.
According to the invention, the personal training device is designed for the combination of sensors 5 to detect the muscle activity and body movement of the user by means of said sensors. The sensor may also be called an external sensing device, and is used to send or transmit the detected signal to an external device such as a control device. The personal training device can be designed to include men's and women's styles. See
Preferably, the personal training device is designed to be tight-fitting or skin-tight to ensure that the sensor mounted in the sensor accommodating unit 3 can be in close contact with the user's skin and that the sensor is on the skin when the user is exercising has the smallest displacement. Preferably, the personal training device can be designed as a short-sleeved t-shirt and shorts to provide maximum flexibility for the user to exercise and easily insert the sensor. There may be as many as 16 or more pockets on the personal training device, each for insertion of sensors.
The top 1 and shorts 2 of the personal training device may be made of, for example, nylon spandex fabric. For example, warp-knitted tricot fabrics have excellent stretch, providing a good fit and free body movement for fitness training, ensuring accurate and stable muscle activity and body movement detection. The fabric typically has good breathability and moisture absorption to maintain good thermophysiological and ergonomic wearing comfort when the sensor is mounted on the personal training device and the user is training.
The pockets used to accommodate the sensors are usually made of fabric that is elastic for easy insertion of the sensor and has see-through properties for monitoring the condition of the sensor. Warp-knitted powernet fabrics are an example of this requirement. Table 2 lists an example of fabric properties for suggested fabrics.
The sEMG and inertial measurement unit (IMU) sensors on the personal training device of the present invention can be any sEMG sensor that meets the following specification requirements: (1) has a communication interface available; (2) is capable of acquiring sEMG signals at a sampling rate of >1000 Hz without notch filtering (3) capable of collecting 9-axis IMU data at a sampling rate of >200 Hz, including data from 3 accelerometers, 3 gyroscopes, and 3 magnetometers; (4) all sensor data are time-synchronized; (5)) The sensor is connected to a pair of disposable electrodes by wires, which can be between 3 and 5 inches in length. The personal training device of the present invention can use state of the art sEMG sensors, such as the Noraxon Ultium EMG sensor system, which includes a receiver and 16 wireless surface EMG sensors with an internal IMU.
The personal training device shown in
Regarding the position of the sensor accommodating unit relative to the body of the user who wears the personal training device, that is, the position of the above-mentioned sensor installed on the user's body, the Inventors have conducted long-term in-depth research and a large number of experiments, Finally, the sensor accommodating unit (that is, the sensor) is selected to be placed at the position corresponding to the above-mentioned eight pairs of muscles. The reasons mainly include: (1) The eight pairs of muscles are superficial layers that can be detected by sEMG; muscles, and (2) the eight pairs of muscles are the main muscle groups that are often trained in resistance training. Of the eight pairs of muscles, the biceps and triceps are responsible for flexion and extension of the elbow. The rectus femoris and biceps femoris are responsible for flexion and extension of the knee. The pectoralis major is responsible for flexion, adduction, and internal rotation of the arm. The rectus abdominis and erector spinae are very important postural muscles. The rectus abdominis controls trunk flexion, and the erector spinae extends and rotates the spine. The upper traps support the weight of the arms and stretch the neck. Another reason to choose a position that corresponds to the upper trapezius is that most people use it unintentionally when it should be relaxed. Therefore, according to the present invention, by placing the sensors at positions corresponding to these muscles, it is possible to more accurately detect the motion of the user and provide feedback to the user.
In contrast, the Hexoskin and Nadi X yoga pants listed in Table 1 do not measure muscle activity at all. Although the Athos training system can measure more muscles than the present invention, its performance analysis is completely different from that of the present invention. The Athos training system compares measured sEMG data, such as muscle activation levels, to reference values based on the target muscle being trained. For example, for a squat, the rectus femoris should reach 40% activation. The present invention further considers muscle fatigue by using variables in the frequency domain on top of the muscle activation level determined from the magnitude of the sEMG signal. Therefore, the detection results of the present invention are more reliable because the magnitude of the sEMG signal can be severely affected by the skin condition, especially in the presence of sweat, which is often the case when a user is exercising. Furthermore, the present invention uses a camera or camera to monitor the user's gestures and movements, improving the accuracy of motion detection compared to using wearable sensors alone.
According to the present invention, the personal training device may comprise two types of pockets.
Type-1 design—for single sensor insertion.
Type-2 design—for dual sensor insertion.
Type-3 design—for single sensor insertion. It is actually a variant of the Type-1 design, which differs from the Type-1 design in that the mounting position and/or orientation of the electrodes 8 and the orientation of the linear openings 7 have been modified to suit different bodies or muscle areas. An example of a pocket design for a sensor for the pectoralis major is shown in
As shown in
It is understood that the inner layer opening 6 and wire opening 7 are optional; different sensors 5 may require inner layer opening 6 and wire opening 7 of different shapes or sizes, or may not require inner layer opening 6 and wire opening 7. In addition, both the inner layer opening 6 and the wire opening 7 may be located in the inner layer of the pocket 3; alternatively, the inner layer opening 6 may be used as a wire opening, so that a separate wire opening 7 may not be required.
In order to fix the position of the sensor during the movement of the user, the pocket 3 can generate sufficient tension or restraint force on the sensor 3 after the sensor 5 is inserted into the pocket 3. Thus, the size of the pocket may depend on the size of the sensor used.
The opening 6 in the pocket 3 can be in common shape such as rectangle, circle, ellipse, square, etc., or the same shape as the sensor 5, so as to facilitate the manufacture and installation of the sensor 5. As shown in
One aspect of the important contribution of the personal training device of the present invention is the optimization of the positions of the various sensors on the personal training device, which are particularly advantageous for detection of specific movements or forms of exercise performed by the user.
For example, in order to detect the EMG signal of the target muscle, a pair of EMG electrodes must be mounted to the muscle belly position of the target muscle, and then connected to the EMG sensor through a pair of short wires. In order to collect more accurate and stable body part IMU data, the EMG sensor (where the IMU is located) is preferably placed parallel and/or perpendicular to the body part.
Preferably, the personal training device of the present invention allows for the simultaneous monitoring of 8 pairs of muscles in the user's body, which are the major muscle groups of the user's body.
The personal training device of the present invention can perform motion recognition and abnormality detection for the user when the user is exercising after wearing. In order to train major muscle groups, the present invention designs a whole-body resistance training program including 10 exercises.
Different exercise modes generate different muscle activation patterns for the user, and these exercise modes can be distinguished in the AI algorithm of the control device of the personal training system of the present invention, i.e, motion recognition. One of the goals of the AI algorithm is to identify the type of movement in real time. The AI algorithm can receive as input readings from 16 sEMG sensors, with or without camera feeds, and then use K-means (k-means) clustering to identify exercise patterns or exercise types. In addition to motion recognition, the algorithm is also able to deduce the current position of the user performing the motion and anomalies, such as incorrect or overuse of certain muscle groups and underuse of stabilizers (if any). A real-time warning is issued to the user if the abnormal situation may increase the risk of injury to the user. For exercises other than those shown in
The control device of
(1) Signal Input Unit 1101
The signal input unit 1101 may receive signals input from, for example, 16 sEMG sensors, and compose time-series data from the signals reported by the sEMG. The input signals may be, for example, in the order of microvolts (μV). The sEMG sensor may be the aforementioned sensor 5 located on the personal training device worn by the user. The signal input unit 1101 may be coupled to the sEMG sensor wirelessly or wired. For example, the signal input unit 1101 may receive sEMG and IMU data collected by each sensor (e.g., the sensor 5 as described above) as shown in
In addition, the signal input unit 1101 may also receive information about the user or user data, such as the user's gender, weight, height, BMI (Body Mass Index), body fat ratio, muscle mass, and exercise selected by the user to be performed, exercise type, etc. Such information or data can be inputted in a prescribed or predetermined format, and can even be directly inputted into the first deep neural network unit 1103 for further processing.
(2) Signal Preprocessing Unit 1102
The signal input unit 1101 transmits the input data to the signal preprocessing unit 1102 for preprocessing to form data in a predetermined format, which is then supplied to other units (for example, the first deep neural network unit 1103 or the second deep neural network unit 1105) for further use. For example, the raw sEMG and IMU data imported from the signal input unit 1101 are cleaned and preprocessed by the signal preprocessing unit 1102. The preprocessing may include (but is not limited to): (1) bandpass, highpass and/or lowpass filtering; (2) rms/moving average/mean absolute/median frequency calculations; (3) data segmentation—which can be done by repetition; (4) data normalization; and (5) data anomaly identification.
(3) The First Deep Neural Network Unit 1103
The signal preprocessing unit 1102 first transmits the preprocessed data to the first deep neural network unit 1103. The first deep neural network unit 1103 is used for executing the AI algorithm of the present invention to perform fatigue analysis on the user's movement, and to estimate the remaining number of repetitions that the user can do according to the current muscle activation signal.
One of the goals of the AI algorithm of the control device is to identify the type of movement in real time. The AI algorithm can receive as input the readings of the 16 sEMG sensors from the signal input unit 1101 (which may or may not be fed with the camera input unit 110), and then use K-means (k-means) clustering to identify the exercise style or type of exercise. In addition to the recognition of the type of exercise, the AI algorithm can also use the first deep neural network unit 1103 to deduce the current position and abnormal situation of the user's exercise, such as the incorrect or overuse of certain muscle groups and the stability of the Inadequate muscle use (if any). A real-time warning is issued to the user if the abnormal situation may increase the risk of injury to the user.
(4) The First Output Unit 1104
The first output unit 1104 represents the estimated number of repetitions remaining, which is an estimate of the number of repetitions the user can complete after the current repetition. It is a positive integer, such as 6, which means the user can perform 6 more repetitions.
(5) The Second Deep Neural Network Unit 1105
The second deep neural network unit 1105 obtains the processed signal data from “signal preprocessing”, and obtains the output from the first output unit 1104. The second deep neural network unit is responsible for executing the AI algorithm of the present invention, suggesting to the user the optimal training load (e.g., in kilograms) based on the current muscle activation signal and the estimated number of remaining repetitions. For example, regardless of the weight of the current training load, its output could be, for example, “20 kg”.
The suggested training load should ideally allow the user to complete the training set with the number of repetitions defined by the training plan of the present invention, and not perform fewer or more repetitions. The recommendations are updated after each training set, so the user is expected to change the training load between training sets if the optimal load is different.
(6) Second Output Unit 1106
The second output unit 1106 represents the user's optimal training load. The recommended training load is expressed in kilograms (kg), and the output for the example is 20 kg. The second output unit 1106 is only updated between training sets.
(7) Camera Input Unit 1107
The camera input unit 1107 represents a video input from the camera. The camera is used to capture the actions of the user, and the user should keep his body in the picture of the camera when performing training.
(8) Posture Detection Model Unit 1108
The output of the camera input unit 1107 is passed to the gesture detection model unit 1108, which receives the user body signal detected from the camera input unit 1107 and generates a body marker based on the user body signal. The gesture detection model unit 1108 may, for example, take a 3-dimensional vector (x, y, z) to output a total of 33 body markers.
(9) Machine Learning Pose Classifier 1109
The machine learning gesture classifier 1109 calculates an 18-dimensional vector (feature vector) from the human body markers output by the gesture detection model unit 1108 to represent the exercise item and the movement state that the user is performing. The vectors can be respectively shown in Table 3 below.
Note that the number of vectors can be changed to suit the training style or sport the user is using.
The machine learning gesture classifier 1109 classifies the user's gestures by using the K-nearest neighbor algorithm based on the feature vectors calculated from the human body markers output from the gesture detection model unit 1108. The machine learning gesture classifier 1109 also calculates the number of repetitions that the user needs to continue to perform training by observing the probability of the user training. For example, if you need to count the number of repetitions a user does a bicep curl, the algorithm looks at the probability of a bicep curl (up) and a bicep curl (down). If the probability of biceps curling (downward) exceeds 80%, the algorithm marks the biceps curling as starting. When the probability of a biceps curl (up) exceeds 80% and the biceps curl is marked as started, the algorithm should mark it as done and increment the repetition counter.
(10) The Third Output Unit 1110
The third output unit 1110 displays the number of repetitions of the exercise program completed by the user. The number of repetitions of the exercise program performed by the user is a positive integer, such as 8 times.
Similarly, the second deep neural network unit 1105 can also be formed by, for example, a deep convolutional neural network, respectively, using the AI algorithm of the present invention to suggest an optimal training load to the user according to the current muscle activation signal and the estimated number of remaining repetitions, and the user's optimal training load is given through the second output unit 1106.
In addition, the machine learning posture classifier 1109 uses the AI algorithm of the present invention to calculate an 18-dimensional vector (feature vector) from the human body markers output from the posture detection model unit 1108 to represent the user's ongoing exercise program and exercise state.
One of the main functions of the machine learning based control device of the present invention is to identify muscle fatigue and thereby suggest optimal training loads for the user. The output of the fatigue analysis is real-time identification of muscle fatigue during resistance training, providing scientific evidence for recommending optimal training loads to the user in the next step.
The personal training system of the present invention may also include a database of user and sensor signals. The database can be contained in the control device or as a separate database, e.g., in a database server.
The Inventors have developed a centralized database with various uses, see
The training software 1304 may be an application program corresponding to the AI fitness training system 1111 of
In addition, the control device of the personal training system of the present invention can also have a coaching module, which can have a GUI and can provide audio feedback, the GUI of the coaching module can be seen in
The GUI allows the user to view all necessary information when performing a resistance training exercise using the personal training device of the present invention. For example, the information may include: (1) current date and time, total training time; (2) workout name included in the current exercise routine, including sets and repetitions, suggested training load; (3) real-time muscle activation level; (4) sensor status; (5) live camera viewer; (6) tutorial video.
The user's real-time muscle activation levels and sensor conditions during exercise can be displayed on a color-coded muscle map. If the control device detects any abnormal condition, a warning will be displayed through the GUI. If a camera is connected to the control, the user can also use the viewer window to check their posture and form.
The Inventors have spent a lot of time in developing the above-mentioned software for the sensor-equipped personal training device and the AI algorithm of the control device. Therefore, the contributing features of the personal training device of the present invention include: the structural design of the sensor pocket to ensure stable skin contact; the optimization of the position where the sensor is inserted into the pocket; the ability of optimizing separately; making samples of various styles, and conducting wearing tests on each subject to find out design deficiencies and wearing difficulties; grading patterns suitable for men's clothing and women's clothing.
The development of AI algorithm software includes: designing and developing AI algorithms to provide more accurate biofeedback and prediction, collecting data for each subject on a six-week fitness training program, training neural networks, calculating K-means The centroids for each exercise of the clustering algorithm, integrating the neural network into the training software system.
As mentioned above, the essential or core features of the present invention include: (1) AI algorithms, including motion recognition algorithms, motion anomaly detection algorithms, and training load recommendation algorithms; (2) personal training devices with wireless sensor interfaces, including sensor pockets and opening design and optimized sensor position. The personal training system of the present invention can solve the following problems: provide a scientific and objective measure of muscle fatigue, rather than a subjective score, during the user's exercise; evaluate exercise form and muscle utilization patterns; minimize risk; and measure and display muscle activation levels during exercise in a user-friendly way.
Therefore, the personal training system of the present invention is capable of measuring and evaluating the muscle activity of the user's major muscle groups, which is not available in other current wearable technologies. The personal training device of the present invention includes garments that allow for easy installation of sensors, which is critical for ordinary users who are less familiar with muscle anatomy. The present invention can be configured to evaluate some of the most common resistance training exercises requiring only dumbbells, without the need for large exercise equipment. The present invention provides users with reliable and real-time feedback on their exercise performance, especially correctness of form and movement, muscle utilization patterns, which are not available in any prior art systems on the market. For example, the corresponding relationship between the contributing technical features of the present invention and the corresponding technical advantages is shown in Table 4 below.
By collecting more experimental data to expand the training data set of the AI algorithm, the personal training device of the present invention can further improve the accuracy of the prediction result, verify the prediction result of the AI algorithm through experiments, and further develop the algorithm in the control device, includes motion recognition algorithms and muscle fatigue detection algorithms, visual and auditory feedback in the coaching module.
For example, an application scenario of the personal training system of the present invention may include an AI-assisted training device for resistance training. Moreover, the system architecture and characteristics of the clothing design of the personal training device of the present invention can be applied to other smart wearable technologies, smart wearable devices of advanced or professional training devices requiring real-time biofeedback, and other textile products.
While the present invention has been described in detail in connection with limited embodiments, it is to be understood that the invention is not limited to these disclosed embodiments. Those of ordinary skill in the art can devise other embodiments that are within the spirit and scope of the present invention, including variations, modifications, substitutions, or equivalent arrangements of parts, all of which fall within the scope of the present invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 202210781291.7 | Jul 2022 | CN | national |
