A breathing training system using mantras

Abstract

The present application relates to a kind of breathing training systems using six-word formula, including user end hardware and cloud server, using the VR head-mounted integrated machine of user end hardware and elastic fabric waistband, guide and detect the breathing training condition of user, realize the accurate classification of abdominal breathing, chest breathing, mixed breathing by means of EMG, displacement, triple fusion calculation mode of flow, utilize mouth shape identification auxiliary phoneme confidence correction, accurately identify different exhale mode of six-word formula, introduce mouth shape identification, phoneme model, physiological sensing fusion multi-modal scoring mechanism, compared with traditional single chest and abdominal displacement monitoring scheme or single audio identification scheme, accurately evaluate the training quality of user from many aspects and comprehensive score, it is convenient for user to understand the improvement direction and progress space of self breathing training.

Description

Technical Field

[0001] The present invention belongs to the technical field of respiratory training systems, and particularly relates to a respiratory training system using the six-character formula. Background Art

[0002] Ultimately, the six-character formula is a health preservation method developed by ancient people based on the principle of sound systems. As is well known, sound originates from vibration. That is to say, the health preservation principle of the six-character formula can be attributed to the angle of human vibration. The emission of specific sounds in the human body requires specific exhalation methods and resonance parts, that is, specific parts of the human body need to move and do work, specific parts of the chest and abdomen need to be pressurized, and relevant internal organs need to resonate. Therefore, different exhalation and pronunciation methods will train different parts and internal organs, which may be the mechanism by which the six-character formula plays its role.

[0003] Patent No. "CN202110993145.6" discloses a respiratory assessment and training control system that detects the abdominal movement fluctuation data of the user through a chest and abdomen belt, measures the respiratory flow data at the mouth, and analyzes to obtain a training plan. However, the existing respiratory training systems lack the detection of the user's pronunciation, mouth shape, and electromyogram. On the one hand, only the exhalation intensity of the user can be detected. When it is necessary to perform respiratory training with the six-character formula, different exhalation methods of the user cannot be recognized. On the other hand, there is a lack of accurate guidance for the user after the training, and the data generated during the process is not quantified according to the detection items, so the user cannot clearly know how much room for improvement there is in their training from the training report.

[0004] Therefore, designing a respiratory training system using the six-character formula that can detect different exhalation methods of users and provide accurate evaluation after training has become an urgent technical problem to be solved. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a respiratory training system using the six-character formula.

[0006] The technical solution of the present invention is a respiratory training system using the six-character formula, including a user-side hardware and a cloud server.

[0007] The user-side hardware includes a VR headset integrated machine and an elastic fabric belt.

[0008] The VR headset integrated machine is equipped with a display for displaying a 3D pronunciation mouth shape template, a breathing rhythm ring, and an air track, a speaker for playing voice prompts, a built-in microphone array for detecting respiratory audio characteristics, an infrared mouth shape camera and an IMU module for user mouth shape recognition, and a main control SoC that is electrically connected to the cloud server, the display, the speaker, the infrared mouth shape camera, and the IMU module for control.

[0009] The elastic fabric waistband is worn next to the sternum, under the costal arch, and in the rectus abdominis area. It is equipped with a chest and abdominal displacement sensor to detect the magnitude of chest and abdominal displacement, a miniature respiratory flow meter to detect peak and duration of respiratory flow, an EMG sensor to detect the energy of the rectus abdominis and intercostal muscles, a skin conductance sensor (GSR) to monitor emotional state, and a thermistor to detect changes in abdominal skin temperature. The electrical signals are connected to the chest and abdominal displacement sensor, the miniature respiratory flow meter, the EMG sensor, the GSR, and the thermistor in a BLE low-power SoC. The BLE low-power SoC wirelessly connects to a cloud server.

[0010] The cloud server acquires data detected by the user-end hardware through signals, performs linear normalization of features, and obtains sub-scores on a percentage basis for pronunciation accuracy, breathing depth, rhythm consistency, stability, and breathing pattern accuracy. The sub-scores are then weighted and summed to obtain a comprehensive score S. The sub-scores and comprehensive score are used to generate a training end report, which is then fed back to the user-end hardware.

[0011] As a further improvement of the present invention, the cloud server is equipped with a scoring model based on 1D-CNN and Bi-LSTM. The scoring model extracts audio features and physiological features from the data detected by the user terminal hardware, and outputs a 6-dimensional softmax, breathing pattern classification and breathing-pronunciation synchronization score corresponding to the exhalation method of the Six Healing Sounds. After linear normalization of the output results, a 1-dimensional sub-score of the exhalation method is obtained.

[0012] As a further improvement of the present invention, the audio features include 13-dimensional MFCC, energy features, and zero-crossing rate with a frame length of 25ms and a frame shift of 10ms. The physiological features include chest and abdominal displacement, respiratory rate, EMG activity, GSR changes, temperature changes, and mouth alignment. The mouth alignment includes 5-dimensional tongue position and 4-dimensional lip shape features output by the infrared mouth camera.

[0013] As a further improvement of this invention, the pronunciation accuracy sub-score is calculated based on phoneme confidence and the Euclidean distance between the mouth shape and the template. The maximum score of the one-dimensional softmax function is obtained through linear normalization and used as the phoneme confidence. The matching degree is obtained by comparing the degree of tongue descent and mouth opening with the corresponding standard template for exhalation. The phoneme confidence and matching degree are weighted and calculated using the following formula: ,in, For pronunciation accuracy, For phoneme confidence, For matching accuracy; respiratory depth is calculated by constructing an expiratory flow rate curve using peak respiratory flow and duration, obtaining the average expiratory flow rate, and then calculating the ratio with a preset adult expiratory template baseline. The calculation formula is as follows: ,in Q represents the respiratory depth, and Q represents the measured average expiratory flow rate. The template is used as the baseline; stability is obtained based on the variance of the normalized expiratory flow rate curve, and the calculation formula is as follows: ,in, For stability, The variance of the expiratory flow rate curve is given; rhythm consistency is calculated based on the normalized RMS time error between the displayed beat loop and the detected user breathing, using the following formula: ,in, For rhythm consistency, The time error is considered; the accuracy of the breathing pattern is obtained by calculating the proportion of the EMG energy of the rectus abdominis muscle in the total EMG energy of the intercostal muscles using the EMG energy of the rectus abdominis muscle, and the calculation formula is as follows: ,in, For the accuracy of breathing patterns, The percentage of total EMG energy in the rectus abdominis muscle is given; the comprehensive score is weighted at w1-5, with values ​​of 0.3, 0.25, 0.2, 0.15, and 0.1 respectively. The formula for calculating the comprehensive score is as follows: .

[0014] As a further improvement to the present invention, the training completion report includes improvement suggestions, which modify the self-assessment scores displayed in the training completion report by manually setting improvement scores; the improvement scores include a decrease in breathing depth of 12 points and a decrease in stability of 5 points, when... When the score is <0.04, the stability score is first adjusted to 95, then reduced further. The score for rhythm consistency is... When the time was <50ms, the score was changed to 95, and the breathing pattern accuracy was within [percentage missing]. When the value is ≥0.75, the Boolean flag is set to 1 from within the model, and the score is displayed as 100.

[0015] As a further improvement of the present invention, the built-in microphone array includes four MEMS microphones, positioned corresponding to the sides of the nose and cheekbones of the human body. The outer side of the MEMS microphones is provided with a wind noise prevention structure composed of foam and a waterproof and breathable membrane. The chest and abdominal displacement sensor includes an electrically connected strain gauge and a 9-axis IMU fusion sensor, with a range of 0–50 mm, an accuracy of ±1 mm, and a measurement frequency of 50 Hz. The miniature respiratory flow meter is a thermal mode flow meter with a range of 0–180 L / min, an accuracy of ±3%, and a measurement frequency of 100 Hz. The EMG sensor has a sampling rate of 500 Hz. The skin conductivity sensor has a range of 0–20 μS, and the thermistor has an accuracy of ±0.1 °C.

[0016] After adopting the above method, the user wears a VR headset and an elastic fabric waistband. An infrared lip-sync camera and IMU module are aligned with the mouth to detect the user's mouth movements, obtaining the user's mouth movement data. The Euclidean distance is calculated with the movement parameters of the 3D pronunciation lip-sync template, and auxiliary correction is performed when calculating the phoneme confidence. The MEMS microphones that constitute the built-in microphone array are wrapped with foam and a waterproof and breathable membrane to reduce the impact of wind noise on phoneme acquisition and improve the recognition accuracy of different exhalation methods of the Six Healing Sounds. By combining multiple physiological sensor detections on the elastic fabric waistband, the user's exhalation data is statistically analyzed to construct an exhalation flow curve. The stability of the user's exhalation is identified by the curve variance. The user's exhalation and the rhythm The timing error of the ring tapping is used to identify rhythm consistency, and the ratio of EMG energy is used to identify the force exerted by the user's chest and abdomen during exhalation. The calculation structure of these three factors is integrated to achieve accurate classification of abdominal breathing, chest breathing, and mixed breathing, forming a multimodal scoring mechanism. The average scoring error after training any of the six-sound breathing methods is less than 3%, which is more than 22% higher than the traditional single chest and abdominal displacement monitoring scheme or single audio recognition scheme. The confidence level of phoneme recognition during standard training is stable above 0.90, and it can accurately identify whether the user is using the deep abdominal breathing required by the six-sound breathing method, effectively avoiding errors such as "shallow breathing" or "holding breath", and promptly capturing and recording errors in pronunciation during training.

[0017] By linking the beat ring with data on chest and abdominal movement and respiratory flow detected by the breath trajectory and elastic fabric waistband, and providing visual guidance, the system achieves a quantitative leap from "rapid and uneven" breathing rhythm to "stable and controllable" breathing rhythm. The rhythm consistency score improves by an average of 18.6% after three training sessions, and the variance of the expiratory flow curve corresponding to expiratory stability decreases by 42%. Using a GSR and thermistor to monitor the user's skin conductance and temperature in real time, the system identifies whether the user is in a tense or relaxed state. When the detected data fluctuates, it feeds back to the main control SoC and plays relaxing voice prompts, reducing the training failure rate caused by abnormal user emotions. By constructing a scoring model containing the user's personal breathing curve when the user first uses this breathing training model, subsequent training can use the initial model as a basis to obtain training scores that better match the user's individual situation, improving overall training completion and user compliance.

[0018] By incorporating improved scoring into the end-of-training reports in the early and mid-term stages, the displayed self-scores are lower than the actual scores. For scoring items such as breathing depth, stability, and rhythm consistency, which are easy to approach full marks, the user's lost scores are artificially amplified. This leaves room for score adjustments when making suggestions for subsequent training, forming positive guidance. This satisfies the incentive requirements for users under continuous training, while retaining the comprehensive score calculated from the real self-scores, which reflects the user's true training level and ensures the seriousness of the end-of-training reports. Detailed Implementation

[0019] A breathing training system using the Six Healing Sounds includes user-end hardware and a cloud server.

[0020] The user-end hardware includes a VR headset and an elastic fabric belt.

[0021] The VR headset is equipped with a display for showing 3D lip-sync templates, breathing rhythm rings and breath trajectories, a speaker for playing voice prompts, a built-in microphone array for detecting breathing audio characteristics, an infrared lip-sync camera and IMU module for user lip-sync recognition, and a main control SoC that is electrically connected to a cloud server, display, speaker, infrared lip-sync camera and IMU module for control.

[0022] The elastic fabric waistband is worn next to the sternum, under the costal arch, and in the rectus abdominis area. It is equipped with a chest and abdominal displacement sensor to detect the magnitude of chest and abdominal displacement, a miniature respiratory flow meter to detect peak and duration of respiratory flow, an EMG sensor to detect the energy of the rectus abdominis and intercostal muscles, a skin conductance sensor (GSR) to monitor emotional state, and a thermistor to detect changes in abdominal skin temperature. The electrical signals are connected to the chest and abdominal displacement sensor, the miniature respiratory flow meter, the EMG sensor, the GSR, and the thermistor in a BLE low-power SoC. The BLE low-power SoC wirelessly connects to a cloud server.

[0023] The cloud server acquires data detected by the user-end hardware through signals, performs linear normalization of features, and obtains sub-scores on a percentage basis for pronunciation accuracy, breathing depth, rhythm consistency, stability, and breathing pattern accuracy. The sub-scores are then weighted and summed to obtain a comprehensive score S. The sub-scores and comprehensive score are used to generate a training end report, which is then fed back to the user-end hardware.

[0024] The cloud server is equipped with a scoring model based on 1D-CNN and Bi-LSTM. The scoring model extracts audio and physiological features from the data detected by the user-end hardware and outputs a 6-dimensional softmax, breathing pattern classification and breathing-pronunciation synchronization score corresponding to the exhalation method of the Six Healing Sounds. After linear normalization of the output results, a 1-dimensional sub-score of the exhalation method is obtained.

[0025] Audio features include 13-dimensional MFCC, energy features, and zero-crossing rate with a frame length of 25ms and a frame shift of 10ms. Physiological features include chest and abdominal displacement, respiratory rate, EMG activity, GSR changes, temperature changes, and lip alignment. Lip alignment includes 5-dimensional tongue position and 4-dimensional lip shape features output by the infrared lip camera.

[0026] The pronunciation accuracy sub-score is calculated based on phoneme confidence and the Euclidean distance between the mouth shape and the template. The maximum score obtained through linear normalization and one-dimensional softmax is used as the phoneme confidence. The matching degree is obtained by comparing the degree of tongue descent and mouth opening with the corresponding standard template for exhalation. The phoneme confidence and matching degree are then weighted and calculated using the following formula: ,in, For pronunciation accuracy, For phoneme confidence, For matching accuracy; respiratory depth is calculated by constructing an expiratory flow rate curve using peak respiratory flow and duration, obtaining the average expiratory flow rate, and then calculating the ratio with a preset adult expiratory template baseline. The calculation formula is as follows: ,in Q represents the respiratory depth, and Q represents the measured average expiratory flow rate. The template is used as the baseline; stability is obtained based on the variance of the normalized expiratory flow rate curve, and the calculation formula is as follows: ,in, For stability, The variance of the expiratory flow rate curve is given; rhythm consistency is calculated based on the normalized RMS time error between the displayed beat loop and the detected user breathing, using the following formula: ,in, For rhythm consistency, The time error is considered; the accuracy of the breathing pattern is obtained by calculating the proportion of the EMG energy of the rectus abdominis muscle in the total EMG energy of the intercostal muscles using the EMG energy of the rectus abdominis muscle, and the calculation formula is as follows: ,in, For the accuracy of breathing patterns, The percentage of total EMG energy in the rectus abdominis muscle is given; the comprehensive score is weighted at w1-5, with weights of 0.3, 0.25, 0.2, 0.15, and 0.1 respectively. The formula for calculating the comprehensive score is as follows: .

[0027] The training end report includes improvement suggestions, which are adjusted by manually setting improvement scores to modify the self-assessment scores displayed in the training end report. These improvement scores include a 12-point reduction in breathing depth and a 5-point reduction in stability. When the score is <0.04, the stability score is first adjusted to 95, then reduced further. The score for rhythm consistency is... When the time was <50ms, the score was changed to 95, and the breathing pattern accuracy was within [percentage missing]. When the value is ≥0.75, the Boolean flag is set to 1 from within the model, and the score is displayed as 100.

[0028] The built-in microphone array includes four MEMS microphones positioned to correspond to the sides of the nose and cheekbones. The MEMS microphones are surrounded by a wind-noise-resistant structure consisting of foam and a waterproof and breathable membrane. The chest and abdominal displacement sensor includes an electrically connected strain gauge and a 9-axis IMU fusion sensor, with a range of 0–50 mm, an accuracy of ±1 mm, and a measurement frequency of 50 Hz. The miniature respiratory flow meter is a thermal flow meter with a range of 0–180 L / min, an accuracy of ±3%, and a measurement frequency of 100 Hz. The EMG sensor has a sampling rate of 500 Hz. The skin conductivity sensor has a range of 0–20 μS, and the thermistor has an accuracy of ±0.1℃.

[0029] Example 1: Training with the "He" sound.

[0030] User behavior:

[0031] ① The user wears a VR headset and belt to enter the "He He Technique Training Mode".

[0032] ② The system prompts: "Please prepare, take a deep breath, and make an 'ah' sound for 3 seconds."

[0033] ③ The user slowly inhales for 2 seconds, then opens their mouth and makes a "ha" sound for about 3 seconds. The sound is steady and there is no obvious tremor.

[0034] Module parameter Audio features The MFCC's main frequency is concentrated at 1.2kHz, with stable energy and a stable zero-crossing rate. The phoneme recognition confidence level is 0.94. respiratory flow Peak flow rate was 90 L / min, with a smooth expiratory curve and a duration of 3.1 seconds. chest and abdominal displacement Abdominal contraction of 18mm and chest change of <2mm indicate abdominal breathing. EMG signal The rectus abdominis muscle was involved in 78% of the breathing, and the intercostal muscles were involved in 22%, which is consistent with the abdominal breathing pattern. GSR & Temperature GSR decreased slightly by 0.3 μS, skin temperature increased by 0.1℃, and mood remained stable. Lip recognition Tongue position decreased, mouth opening was 65%, and the match rate with the standard "heh" template was 0.91.

[0035] Training Completion Report:

[0036] Pronunciation accuracy: 94 points

[0037] Breathing depth: 88 points

[0038] Rhythm consistency: 92 points

[0039] Stability: 90 points

[0040] Breathing pattern accuracy 100%

[0041] Phoneme confidence =0.94, Euclidean distance between lip shape and template =0.09,

[0042] =100×(0.7×0.94+0.3(1-0.09)),

[0043] Peak respiratory flow rate 90 L / min, optimal template =90 L / min,

[0044] =100×min(1,90 / 90)=100, the score is reduced to 88.

[0045] Expiratory flow rate curve variance =0.05,

[0046] =100×(1-0.05)=95, the score is reduced to 90.

[0047] Time error RMS =0.08,

[0048] =100×(1-0.08)=92, >50ms, no improvement or tuning will be made.

[0049] EMG energy =0.78,

[0050] =100×0.78=78, X4≥0.75, the accuracy rate of the displayed breathing pattern is 100%.

[0051] S = 0.3 × 95.2 + 0.25 × 100 + 0.2 × 95 + 0.15 × 92 + 0.1 × 78 = 93.5, rounded to one decimal place.

[0052] Overall rating: 93.5 / 100

[0053] VR Feedback: Green Halo + Voice "Pronunciation is standard, breathing is steady, keep it up!"

[0054] Breath trajectory: The light spot perfectly matches the track without deviation.

[0055] Example 2: Training with the "Call" technique.

[0056] User behavior:

[0057] ① The user enters the "Calling Technique Training Mode"

[0058] ② The system prompts: "Please pronounce the sound 'hu' and remember to exhale using your abdomen."

[0059] ③ The user inhales quickly for 1 second, then makes a short "whoosh" sound, which is high-pitched and lasts for only 1.8 seconds, with obvious chest rise and fall.

[0060] Module parameter Audio features Large energy fluctuations, abnormal zero-crossing rate, and phoneme recognition confidence level of 0.68. respiratory flow The peak flow rate was 135 L / min, the curve was steep, and the duration was 1.8 seconds, which was determined to be rapid exhalation. chest and abdominal displacement Abdominal contraction was only 5mm, and chest contraction was 12mm, which was determined to be thoracic breathing. EMG signal The rectus abdominis muscle is involved in 25% of the breathing, the intercostal muscles in 75%, and thoracic breathing is dominant. GSR & Temperature GSR increases by 1.2 μS, temperature decreases by 0.05°C, indicating emotional stress. Lip recognition Mouth opening 45%, tongue position slightly high, matching degree with "exhale" template = 0.63

[0061] Training Completion Report:

[0062] Pronunciation accuracy: 68 points; Breathing depth: 55 points; Rhythmic consistency: 60 points; Stability: 52 points; Breathing pattern accuracy: 0%

[0063] S = 0.3 × 68 + 0.25 × 55 + 0.2 × 60 + 0.15 × 52 + 0.1 × 52 = 59.2, rounded to one decimal place.

[0064] Overall rating: 59.2 / 100VR Feedback: Red warning + voice message "Breathing too shallow, please slow down and use abdominal breathing." Breath trajectory: The light spot deviated from its trajectory 3 times and eventually fell.

[0065] By having the user wear a VR headset and an elastic fabric waistband, an infrared lip-sync camera and IMU module are positioned at the mouth to detect the user's mouth movements. This data is used to calculate the Euclidean distance between the user and the lip-sync template's movement parameters, and is then used for auxiliary correction when calculating phoneme confidence. A foam and waterproof, breathable membrane are wrapped around the MEMS microphone array to reduce wind noise and improve the accuracy of recognizing different exhalation methods of the Six Healing Sounds. By combining multiple physiological sensors on the elastic fabric waistband, the user's exhalation data is statistically analyzed to construct an exhalation flow curve. The curve variance is used to identify the stability of the user's exhalation and the timing of the user's exhalation relative to the beat cycle. The system uses interval error to identify rhythm consistency, and the ratio of EMG energy to identify the force exerted by the user's chest and abdomen during exhalation. The calculation structure of these three factors is integrated to achieve accurate classification of abdominal breathing, chest breathing, and mixed breathing, forming a multimodal scoring mechanism. The average scoring error after training any of the six-character breathing methods is less than 3%, which is more than 22% higher than the traditional single chest and abdominal displacement monitoring scheme or single audio recognition scheme. The confidence level of phoneme recognition during standard training is stable above 0.90, and it can accurately identify whether the user is using the deep abdominal breathing required by the six-character breathing method, effectively avoiding errors such as "shallow breathing" or "holding breath", and promptly capturing and recording pronunciation errors during training.

[0066] By linking the beat ring with data on chest and abdominal movement and respiratory flow detected by the breath trajectory and elastic fabric waistband, and providing visual guidance, the system achieves a quantitative leap from "rapid and uneven" breathing rhythm to "stable and controllable" breathing rhythm. The rhythm consistency score improves by an average of 18.6% after three training sessions, and the variance of the expiratory flow curve corresponding to expiratory stability decreases by 42%. Using a GSR and thermistor to monitor the user's skin conductance and temperature in real time, the system identifies whether the user is in a tense or relaxed state. When the detected data fluctuates, it feeds back to the main control SoC and plays relaxing voice prompts, reducing the training failure rate caused by abnormal user emotions. By constructing a scoring model containing the user's personal breathing curve when the user first uses this breathing training model, subsequent training can use the initial model as a basis to obtain training scores that better match the user's individual situation, improving overall training completion and user compliance.

[0067] By incorporating improved scoring into the end-of-training reports in the early and mid-term stages, the displayed self-scores are lower than the actual scores. For scoring items such as breathing depth, stability, and rhythm consistency, which are easy to approach full marks, the user's lost scores are artificially amplified. This leaves room for score adjustments when making suggestions for subsequent training, forming positive guidance. This satisfies the incentive requirements for users under continuous training, while retaining the comprehensive score calculated from the real self-scores, which reflects the user's true training level and ensures the seriousness of the end-of-training reports.

Claims

1. A breathing training system using mantras, characterized by: The user terminal hardware and the cloud server are included, The user terminal hardware includes a VR headset all-in-one machine and an elastic fabric waistband, The VR headset all-in-one machine is installed with a display for displaying a 3D pronunciation mouth shape template, a breathing rhythm ring and an air track, a loudspeaker for playing a voice prompt, a built-in microphone array for detecting breathing audio features, an infrared mouth shape camera and an IMU module for recognizing user mouth shapes, and a main control SoC for connecting the cloud server, the display, the loudspeaker, the infrared mouth shape camera and the IMU module through electrical signals to control them; The elastic fabric waistband is worn on the side of the sternum, under the costal arch and in the rectus abdominis area of the human body, and is installed with a chest and abdominal displacement sensor for detecting the displacement of the chest and abdomen, a miniature respiratory flow meter for detecting the peak value and duration of respiratory flow, an EMG sensor for detecting the energy of the rectus abdominis and intercostal muscles, a skin conductance sensor GSR for monitoring the emotional state, and a thermistor for detecting the temperature change of the abdominal skin, and a BLE low-power SoC for connecting the chest and abdominal displacement sensor, the miniature respiratory flow meter, the EMG sensor, the GSR and the thermistor through electrical signals, and the BLE low-power SoC is wirelessly connected to the cloud server; The cloud server obtains the data detected by the user terminal hardware through signals, performs linear normalization features, obtains pronunciation accuracy, breathing depth, rhythm consistency, stability, and respiratory mode accuracy sub-scores in percentage, and sums the sub-scores to obtain a comprehensive score S, generates a training end report using the sub-scores and the comprehensive score, and feeds back the training end report to the user terminal hardware; The cloud server is provided with a scoring model constructed based on 1D-CNN and Bi-LSTM, which extracts audio features and physiological features from the data detected by the user terminal hardware, outputs a 6-dimensional softmax corresponding to the six-word formula exhalation method, a breathing mode classification and a breathing-pronunciation synchronization score, and after linear normalization processing of the output results, a 1-dimensional sub-score of the exhalation method is obtained; The audio features include 13-dimensional MFCC, energy features, and zero-crossing rate with a frame length of 25ms and a frame shift of 10ms, and the physiological features include chest and abdominal displacement, breathing rate, EMG activity, GSR change, temperature change and mouth shape alignment, and the mouth shape alignment includes 5-dimensional tongue position and 4-dimensional lip shape features output by the infrared mouth shape camera; The pronunciation accuracy sub-score is calculated based on the phoneme confidence and the mouth shape-template Euclidean distance. The maximum value of one-dimensional softmax obtained by linear normalization is taken as the phoneme confidence. The matching degree is obtained by comparing the tongue position drop degree, mouth shape opening degree, and corresponding standard template of blowing way. The phoneme confidence and the matching degree are weighted and calculated. The calculation formula is wherein, is the pronunciation accuracy, is the phoneme confidence, is the matching degree; the breath depth is constructed by the breath flow peak value and the duration to obtain the average breath flow, and the breath depth is calculated by comparing the average breath flow with the preset adult breath template benchmark. The calculation formula is wherein is the breath depth, Q is the measured average breath flow, is the template benchmark; the stability is obtained based on the variance of the normalized breath flow curve. The calculation formula is wherein, is the stability, is the variance of the breath flow curve; the rhythm consistency is calculated based on the time error RMS between the display playing rhythm ring and the detected user breath after normalization. The calculation formula is wherein, is the rhythm consistency, x5 is the time error; the breath mode accuracy is obtained by using the EMG energy of the rectus abdominis muscle to calculate the proportion of the rectus abdominis muscle in the total EMG energy. The calculation formula is wherein, is the breath mode accuracy, is the proportion of the rectus abdominis muscle in the total EMG energy; the comprehensive score is weighted with w1-5, which are 0.3, 0.25, 0.2, 0.15, and 0.1 in sequence. The calculation formula of the comprehensive score is, .

2. The breath training system using the six-syllable mantra according to claim 1, wherein: The training end report includes improvement suggestions, the improvement suggestions set improvement scores by human, and modify the self-score values displayed in the training end report; the improvement scores include a breathing depth reduction score of 12 points, a stability reduction score of 5 points, when x3<0.04, the stability is first modified to 95 points, then the score is reduced, the rhythm consistency score is modified to 95 points when x5<50ms, and the respiratory mode accuracy is modified to 100 points when X4≥0.75, and the Boolean flag is set to 1 in the model.

3. The breath training system using the six-syllable mantra according to claim 1, wherein: The built-in microphone array includes 4 MEMS microphones, which are located on both sides of the nose bridge and the cheekbone of the human body. The outer side of the MEMS microphone is provided with a wind noise prevention structure composed of foam and a waterproof and breathable film. The chest and abdominal displacement sensor includes a strain gauge and a 9-axis IMU fusion sensor connected by electricity, with a range of 0-50 mm, an accuracy of ±1 mm, and a measurement frequency of 50 Hz. The miniature respiratory flow meter is a thermal mode flow meter, with a range of 0-180 L / min, an accuracy of ±3%, and a measurement frequency of 100 Hz. The sampling rate of the EMG sensor is 500 Hz. The range of the skin conductance sensor is 0-20 μS, and the accuracy of the thermistor is ±0.1℃.