Detection system and method for discriminating traditional Chinese medicine pulse condition based on time-frequency domain feature fusion

Abstract

The invention discloses a detection system and method for judging the traditional Chinese medicine pulse condition based on time-frequency domain feature fusion. The method comprises the steps that pulse wave fluctuation pressure in the pulse taking process of a doctor and time sequence signals of pulse taking finger pressure of the doctor are collected; the collected time sequence signals are preprocessed, and representative feature information of the time sequence signals is extracted for fusion; and forming a training set by the traditional Chinese medicine calibration data and the fused feature data, and constructing and training a pulse condition classification model based on an artificial intelligence classification recognition algorithm so as to perform pulse condition discrimination by using the pulse condition classification model. Data are synchronously collected in the pulse taking process of a doctor, the consistency of data calibration is guaranteed, meanwhile, compared with a traditional passive sampling scheme with fixed pressure and fixed position, by means of active intervention of the doctor, the positions of cun, guan and chi and the finger pressure of floating, middle and sinking can be flexibly adjusted according to the age, height, weight, gender and the like of different crowds, and the accuracy of pulse taking is improved. And the reliability and effectiveness of data acquisition are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of traditional Chinese medicine pulse detection, in particular to a detection system and method for distinguishing traditional Chinese medicine pulse based on time-frequency domain feature fusion. Background Art

[0002] In the process of traditional Chinese medicine diagnosis, pulse diagnosis is one of the important diagnostic methods. The pulse of a healthy person is called a normal pulse. Generally, it is neither floating nor sinking, neither large nor small, neither strong nor weak, neither fast nor slow, but even and gentle, with a regular rhythm, also known as a flat pulse or a slow pulse.

[0003] The pulse pattern undergoes physiological or temporary changes under the influence of internal and external factors. For example, the younger the age, the faster the pulse. The pulse of an infant is rapid and rapid, 120 to 140 times per minute; the pulse of a five- or six-year-old child is usually six beats per breath, 90 to 110 beats per minute; the pulse of a young and middle-aged person is strong, and the pulse is usually strong; the pulse of an elderly person is weak, and the pulse of an adult female is thinner and slightly faster than that of an adult male; the pulse of a thin person is more floating, and the pulse of a fat person is more sinking; the pulse of a person who does heavy physical labor, does strenuous exercise, walks long distances, drinks alcohol, eats a lot, or is emotionally excited is usually fast and strong, and the pulse is weak when hungry.

[0004] When taking the pulse, you also need to use three different finger strengths to press the pulse. Gently pressing on the skin is "floating"; moderately pressing on the muscles is "medium"; and heavy pressing on the tendons and bones is "deep". The three parts of Cun, Guan, and Chi have three signs of floating, medium, and sinking, collectively known as "three parts and nine signs". The pulses taken with different techniques have different clinical significance. Usually, if the pulse is floating on the outside, the disease is shallow, and if it is sinking on the inside, the disease is deep.

[0005] However, traditional pulse diagnosis relies on the doctor's finger touch and subjective experience, lacks objective and quantitative standards, and therefore has large individual differences and instability. With the development of sensor technology, signal processing technology, and artificial intelligence recognition technology, it has become possible to develop equipment and methods that can objectively and accurately detect and analyze pulse conditions. Summary of the invention

[0006] The purpose of the present invention is to provide a detection system and method for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion, which is used to solve the problems of low extraction rate of effective pulse signals, poor stability of pulse signals, low pulse recognition accuracy, etc. in the prior art.

[0007] The present invention solves the above problems through the following technical solutions:

[0008] The detection method for distinguishing TCM pulse based on the fusion of time-frequency domain features includes:

[0009] Collect the pulse wave fluctuation pressure and the timing signal of the doctor's pulse pressure during the doctor's pulse diagnosis process;

[0010] Preprocess the collected time series signals, extract the representative feature information of the time series signals for fusion; combine the TCM calibration data and the fused feature data into a training set, build and train the pulse classification model based on the artificial intelligence classification and recognition algorithm, and use the pulse classification model to distinguish the pulse.

[0011] As a further improvement, the pulse classification model is constructed; the specific method includes:

[0012] Step a, calibrating the collected time series signal data and the pulse type of traditional Chinese medicine diagnosis by calibration values ​​one by one to form complete training data;

[0013] Step b, extracting and sorting the collected pulse wave fluctuation pressure signals of Cun, Guan and Chi and the doctor's pulse feeling finger pressure signals;

[0014] Step c, respectively aligning the time sequence of the doctor's pulse pressure signals of Cun, Guan and Chi with the pulse wave fluctuation signals, and at the same time taking the derivative of the pulse wave fluctuation signals, separating the floating, middle and sinking pulse wave fluctuation signals in combination with the gradient change of the pulse wave fluctuation signals, and finally generating nine independent signals;

[0015] Step d, denoising, baseline calibration and stable periodic signal extraction for each signal;

[0016] Step e, the stable periodic signal extracted in step d is analyzed in time domain and frequency domain to extract characteristic data that can reflect pulse information;

[0017] Step f, separately splice and fuse the characteristic data of each signal in step e to form fused characteristic data that can reflect the current patient's pulse type, so that the calibration value of the TCM calibration data and the fused characteristic data together form a complete training set data, and the complete training set is formed by the diagnostic sampling of several patients.

[0018] As a further improvement, in step b, the three pulse wave fluctuation pressure signals are sorted out, and the specific method is: effectively analyzing each signal collected, and taking the mean value of the effective signal as the pulse wave fluctuation pressure signal at the current position;

[0019] The validity analysis refers to calculating the energy spectrum of the pulse wave signal. When the energy value reaches a set threshold and the time domain characteristics show quasi-periodicity, it is determined to be a valid signal.

[0020] The quasi-periodicity refers to the fact that the pulse wave period values ​​present a certain similarity or alternating regularity, that is, it is determined to be a quasi-periodic signal.

[0021] As a further improvement, in step c, the timing alignment is specifically performed by aligning the finger pressure signal and the pulse wave signal based on the horizontal coordinate of the sampling time, and determining the waveform intervals of the floating, middle and sinking signals with reference to the finger pressure signal waveform trend;

[0022] The pulse wave fluctuation signal is separated by taking the first-order derivative of the pulse wave based on the determined waveform interval, and the point where the waveform slope fluctuates abnormally is used as the cutoff point for separation.

[0023] As a further improvement, in step d, each signal is subjected to denoising, baseline calibration and stable periodic signal extraction; the specific method is as follows:

[0024] The signal denoising uses a bandpass filter to filter out environmental noise greater than 30 Hz and noise introduced by human movement or breathing less than 0.5 Hz;

[0025] The baseline calibration is to perform baseline calibration on periodic signals with different pulse wave fluctuation pressures by extreme point detection combined with partial least squares fitting;

[0026] The stable periodic signal extraction will introduce a large number of interference signals due to the interference of the doctor's flexible adjustment of the finger position and the unknown factors of the patient and the surrounding environment during the sampling process. It is necessary to analyze the distribution of the collected pulse signals based on the waveform periodic data and eliminate the signals that seriously deviate from the periodic distribution law.

[0027] As a further improvement, in step e, the extracted time domain feature data include the main wave amplitude h1, the dicrotic wave amplitude h3, the pulse output SV and the average pulse period t;

[0028] Among them, pulse output SV = coef × t / (K × K); coefficient coef = 0.283; difference K = Ps-Pd, Ps and Pd are the highest peak value and baseline value of the pulse wave respectively.

[0029] As a further improvement, in step e, the extracted frequency domain feature data includes reflection coefficient Rf, propagation velocity PWV and wavelet coefficient Wf;

[0030] Wherein, the reflection coefficient Rf = (P1-Pmin) / P1; P1 is the spectrum amplitude corresponding to the first harmonic in the spectrum diagram after the signal is Fourier transformed, and Pmin is the spectrum amplitude corresponding to the first minimum value of the harmonic;

[0031] The propagation velocity PWV = 4 × fmin × L; fmin is the frequency corresponding to the first minimum value of the harmonic; L is the distance that the reflected wave is transmitted from the terminal to the radial artery; the wavelet coefficient Wf is the high and low frequency coefficients obtained after the signal is transformed by wavelet.

[0032] As a further improvement, the pulse classification model is constructed; the specific method also includes:

[0033] g. Based on the machine learning qualitative analysis modeling algorithm, the training set is used to find the best model through cross-validation, and the optimal model is used as the final pulse classification model;

[0034] h. Package the pulse classification model in the specified format and upload it to the platform for backup.

[0035] At the same time, the present invention also solves the above problems through the following technical solutions:

[0036] The detection system for distinguishing TCM pulse based on time-frequency domain feature fusion is used to implement the detection method for distinguishing TCM pulse based on time-frequency domain feature fusion as described above, comprising:

[0037] A data acquisition module is used to collect the timing signal of the pulse wave fluctuation pressure and the pressure of the doctor's pulse finger pressure;

[0038] A signal processing module, used for preprocessing the collected time series signal;

[0039] Feature extraction module, used to extract representative feature information of time series signals and fuse them;

[0040] Intelligent discrimination module, which is used to combine TCM calibration data and feature fusion data into a training set, and to build and train the pulse classification model based on artificial intelligence classification and recognition algorithm;

[0041] The data storage module is implemented based on the cloud platform and is used to store the original collected pulse wave fluctuation pressure and the pressure data of the doctor's pulse feeling and finger pressure, the pulse classification model and the classification recognition results;

[0042] The control display module is used to control the parameter configuration, real-time waveform display and pulse condition identification result display of the data acquisition module.

[0043] As a further improvement, the data acquisition module has a structural form including a finger cuff and a wristband, the signal processing unit is integrated in the wristband, the flexible pressure sensor is integrated at the fingertip of the finger cuff, and the transmission line between the signal processing unit and the flexible pressure sensor is distributed on the back of the hand in a wiring manner with adjustable length.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] (1) The detection system proposed in the present invention is worn by doctors and is used to identify TCM pulse based on the fusion of time-frequency domain features. The system ensures compatibility with different hand sizes and finger thicknesses by means of Velcro wrapping and retractable wiring. Data is collected synchronously during the doctor's pulse diagnosis process, thus ensuring the consistency of data calibration. Compared with the traditional passive sampling scheme with fixed pressure and fixed position, with the active intervention of doctors, the positions of Cun, Guan and Chi as well as the finger pressures of floating, middle and sinking can be flexibly adjusted according to the age, height, weight, gender, etc. of different people, thereby ensuring the reliability and effectiveness of data collection.

[0046] (2) The present invention can comprehensively capture the subtle changes in the pulse signal through time-frequency fusion feature analysis, effectively shielding the original signal from being easily interfered by the environment, equipment, doctors and patients themselves, thereby improving the stability of the pulse data; the machine learning qualitative analysis modeling algorithm is used for pulse discrimination, which can extract some potential non-quantitative characteristics of the pulse data and improve the accuracy of pulse category recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a framework diagram of a detection system for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion in embodiment 1 of the present invention;

[0048] Figure 2 It is a flow chart of a detection method for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion according to embodiment 2 of the present invention;

[0049] Figure 3 The present invention is to provide a flow chart of the method for building a pulse classification model in Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described in more detail below in conjunction with the drawings in the preferred embodiments of the present application. In the drawings, the same or similar reference numerals throughout represent the same or similar parts or parts with the same or similar functions. The described embodiments are part of the embodiments of the present application, not all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limitations on the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0051] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0052] In the description of the present application, it should be noted that, unless expressly provided and limited otherwise, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or display comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or displays.

[0053] The following will be combined Figure 1 , a detection system and method for distinguishing TCM pulse condition based on time-frequency domain feature fusion involved in the embodiment of the present application are described in detail. It is worth noting that the following embodiments are only used to explain the present application and do not constitute a limitation to the present application.

[0054] Embodiment 1:

[0055] Refer to the attached Figure 1 ,This embodiment provides a detection system for distinguishing TCM pulse based on time-frequency domain feature fusion, including: a data acquisition module, a signal processing module, a feature extraction module, an intelligent discrimination module, a data storage module and a control display module;

[0056] The data acquisition module is used to collect the pulse wave fluctuation pressure and the timing signal of the doctor's finger pressure during the doctor's pulse diagnosis process;

[0057] The data acquisition module has a specific structural form of a finger sleeve + wristband type, that is, the signal processing unit is integrated in the wristband, the flexible pressure sensor is integrated in the finger tip of the finger sleeve, and the transmission line between the signal processing unit and the sensor is distributed on the back of the hand in the form of a cable. The length of the cable can be flexibly adjusted according to the size of the doctor's hand, so as to improve the adaptability of the product. Preferably, the finger sleeve adopts an independent Velcro wrapping wearing solution for the index finger, middle finger, and ring finger to adapt to the thickness of different doctors' fingers, ensure that the flexible pressure sensor can be worn in the upper center of the fingertip each time, and improve the effectiveness of data sampling.

[0058] Furthermore, for flexible pressure sensors, multiple sensors are selected on a single fingertip for array arrangement. The sensor array arrangement can be selected in a V-shaped, parallel, or surrounding shape to increase the contact area of ​​the sensor and improve the stability of the data. In the data acquisition module, the sampling method is that the Chinese medicine practitioner wears the data acquisition device composed of the above-mentioned finger sleeve + wristband to perform pulse-taking data sampling and calibration on the patient. The finger sleeve and the sensor are made of a one-time film-forming extremely thin process material to ensure that the Chinese medicine practitioner collects data and takes the pulse simultaneously without changing the traditional Chinese medicine consultation process and habits.

[0059] Optionally, the data acquisition module and the circuit separate the pressure data collected by the sensor into pulse wave fluctuation data and the doctor's pulse feeling and finger pressure data independently through a DC component separation mechanism, thereby ensuring the purity of the pulse wave fluctuation signal.

[0060] In this embodiment, taking the doctor wearing the data acquisition device on the right hand as an example, the doctor wears the device on the index finger, middle finger and ring finger to perform pulse diagnosis on the patient's Cun, Guan and Chi points, and simultaneously completes the collection of pulse wave fluctuations and the doctor's pulse pressure value timing signal data and uploads it to the cloud. Each finger cuff of the device consists of three flexible pressure sensors arranged in a herringbone shape for multi-channel data acquisition.

[0061] The doctor's pulse pressure signals, including real-time recording of floating, middle and sinking pressure values, are used to guide doctors to collect data in a standardized manner. The collected data specifically includes a total of nine types of pulse signals corresponding to the floating, middle and sinking parts of Cun, Guan and Chi, as well as the corresponding doctor's pressure signals.

[0062] A signal processing module, used for preprocessing the collected time series signal;

[0063] The preprocessing operation mainly includes signal denoising, floating, middle and sinking truncation extraction, baseline calibration and stable periodic signal extraction. Among them, signal denoising uses a bandpass filter to filter out environmental noise greater than 30Hz and noise introduced by human movement or breathing less than 0.5Hz. Floating, middle and sinking truncation extraction aligns the doctor's finger pressure signal with the pulse wave signal in time sequence, and combines the gradient change of the pulse wave signal to effectively separate the floating, middle and sinking pulse wave signals. Baseline calibration, through extreme point detection combined with partial least squares fitting, baseline calibration is performed on periodic signals with different pulse wave fluctuation pressures, so as to accurately extract the effective characteristic values ​​within and between the pulse wave signal cycles. Stable periodic signal extraction operation, because during the sampling process, the doctor will flexibly adjust the finger position and the interference of unknown factors of the patient and the surrounding environment will introduce a large number of interference signals. It is necessary to analyze the distribution of the collected pulse signal based on the waveform periodic data, and eliminate the signals that seriously deviate from the periodic distribution law to improve the effectiveness of the signal. The distribution analysis of periodic data is specifically the distance distribution and mode extraction between cycles.

[0064] In this embodiment, the signal processing module performs validity analysis on the three channel data corresponding to the collected Cun, Guan and Chi signals, retains the valid signals for averaging processing, intercepts the floating, middle and sinking data of the processed signals, and filters the intercepted data and extracts stable periodic signals.

[0065] Feature extraction module, used to extract representative feature information of time series signals and fuse them;

[0066] In this embodiment, the extracted features mainly include time domain features and frequency domain features. Among them, the time domain features are mainly the amplitude and period parameters related to the main wave, pre-wave, dicrotic wave, and descending gorge that can reflect the difference in pulse characteristics in the single cycle signal. The frequency domain features are mainly to extract features such as pulse rate and reflection coefficient by frequency conversion of the full cycle signal.

[0067] In the present embodiment, the nine groups of data after signal preprocessing are respectively analyzed in time domain and frequency domain. The time-frequency domain characteristic data of each group are spliced ​​to form the pulse data corresponding to the current sampled patient. If the pulse data is used as model training data, it is necessary to calibrate it in conjunction with the doctor's diagnosis result. If it is used as model test data, it is not necessary to calibrate it.

[0068] Intelligent discrimination module, which is used to combine TCM calibration data and feature fusion data into a training set, and to build and train the pulse classification model based on artificial intelligence classification and recognition algorithm;

[0069] The model building is specifically based on an artificial intelligence classification and recognition algorithm to perform qualitative analysis and modeling on the input training set data. The artificial intelligence classification and recognition algorithm can be selected from one of the LDA algorithm, logistic regression algorithm, support vector machine algorithm, decision tree algorithm, random forest algorithm, Xgboost algorithm, LightGBM algorithm, ANN algorithm, CNN algorithm, DNN algorithm or neural network algorithm.

[0070] In this embodiment, the data to be tested is uploaded to the platform, and prediction calculations are performed with the trained model to classify and identify the pulse type of the patient, which can be a single pulse or a combination of two or three pulses.

[0071] The data storage module is implemented based on the cloud platform and is used to store the original collected pulse wave fluctuation pressure and the pressure data of the doctor's pulse feeling and finger pressure, the pulse classification model and classification recognition results, etc.;

[0072] In this embodiment, the data acquisition module uploads the collected data to the cloud platform through the network for storage. At the same time, the pulse classification model and the discrimination results are also stored in the cloud platform for flexible calling and information tracking.

[0073] The control display module is used to control the parameter configuration of the data acquisition module, real-time waveform display, and pulse condition identification result display, etc.

[0074] In this embodiment, the control display module is connected to the data acquisition module through the mobile terminal via Bluetooth / network to control the number of data sampling, naming, uploading, waveform display and other operations of data samples.

[0075] Embodiment 2:

[0076] See attached Figure 2-3 , the method of distinguishing TCM pulse based on the fusion of time-frequency domain features includes:

[0077] Collect the pulse wave fluctuation pressure and the timing signal of the doctor's pulse pressure during the doctor's pulse diagnosis process;

[0078] Preprocess the collected time series signals, extract the representative feature information of the time series signals for fusion; combine the TCM calibration data and the fused feature data into a training set, build and train the pulse classification model based on the artificial intelligence classification and recognition algorithm, and use the pulse classification model to distinguish the pulse.

[0079] In a specific embodiment, this embodiment is based on the detection method embodiment of the time-frequency domain feature fusion discrimination of the Chinese medicine pulse provided in Example 1. Figure 2 , including the following steps:

[0080] S1. Set parameters for the sampling terminal, i.e., the data acquisition device, and complete the network connection;

[0081] In this step, a mobile terminal is selected to connect to the sampling terminal via Bluetooth, and the sampling parameter configuration is completed through the mobile terminal display interface, where the number of data sampling points is set to 5000 and the wearing position is selected as the right hand.

[0082] S2, the traditional Chinese medicine wearable data acquisition device collects pulse data, collects the pulse wave fluctuation pressure and the timing signal of the doctor's pulse pressure during the doctor's pulse diagnosis;

[0083] In this step, the Chinese medicine practitioner wears a bracelet and finger sleeves to take the patient's pulse. During the data collection process, the waveform sampled by each sensor can be displayed in real time on the screen of the mobile terminal at the same time.

[0084] S3, pulse data upload to the cloud;

[0085] In this step, after completing data sampling, click the "Data Upload" button on the mobile terminal app to complete one-click upload of data to the cloud.

[0086] S4, pre-processing the collected time series signals, extracting representative feature information of the time series signals for fusion, and forming a training set with the TCM calibration data and the fused feature data; and using the pulse classification model built and trained based on the artificial intelligence classification and recognition algorithm to perform pulse discrimination;

[0087] In this step, the cloud platform takes the pulse data to be detected, i.e., the training set, as input, and uses a pre-established pulse classification model based on an artificial intelligence classification and recognition algorithm to perform pulse category detection and store the results.

[0088] S5. Feedback the result to the display terminal for display;

[0089] In this step, the pulse condition results determined by the cloud platform are fed back to the mobile terminal through the network for result display.

[0090] Next, in step S4, based on the pre-processing of the collected time series signal, the representative feature information of the time series signal is extracted and fused, and the TCM calibration data and the fused feature data are combined into a training set; and the pulse classification model built and trained based on the artificial intelligence classification recognition algorithm is used to distinguish the pulse condition; the whole process is described in detail, and the flowchart is shown in Figure 3 :

[0091] a. TCM pulse data calibration: calibrate the collected time series signal data with the pulse type diagnosed by TCM one by one to form complete training data;

[0092] In this step, a total of 30,000 pulse data from different departments and different populations for half a year are collected, and based on the pulse diagnosis of famous Chinese medicine practitioners, each data is calibrated with calibration values ​​to form complete training data.

[0093] b. Pulse data extraction: extract and sort out the collected pulse wave pressure signals of Cun, Guan and Chi and the pulse pressure signals of doctors;

[0094] Furthermore, the three-way signal sorting mainly involves effective analysis of the signals collected by each array-arranged sensor, and taking the mean of the effective signals as the pulse data signal at the current position.

[0095] Validity analysis refers to the calculation of the energy spectrum of the pulse wave signal. When the energy value reaches the set threshold and the time domain characteristics show quasi-periodicity, it is judged as a valid signal.

[0096] Quasi-periodicity means that the pulse wave period values ​​show a certain similarity or alternating regularity, which means it is judged as a quasi-periodic signal.

[0097] In this step, the collected pulse data signals and finger pressure signals of Cun, Guan and Chi are downloaded and extracted from the cloud platform, and the validity analysis is performed on the three groups of sensor data corresponding to each part, and the valid data is averaged to ensure the stability of the sampling signal.

[0098] c. Pulse data decomposition: align the time sequence of the doctor's finger pressure signals of Cun, Guan and Chi with the pulse wave signals respectively, and at the same time, take the derivative of the pulse wave signals, and separate the floating, middle and sinking pulse wave signals based on the gradient change of the pulse wave signals, and finally generate 9 independent signals.

[0099] Specifically, the timing alignment is to overlap and align the finger pressure signal and the pulse wave signal based on the horizontal coordinate of the sampling time, and determine the waveform intervals of the floating, middle and sinking signals with reference to the finger pressure signal waveform trend.

[0100] The pulse wave fluctuation signal separation is based on a determined waveform interval, first-order derivative is performed on the pulse wave, and the point where the waveform slope fluctuates abnormally is used as a cutoff point for separation.

[0101] In this step, the three signals of Cun, Guan and Chi are time-series aligned and derived based on the finger pressure signals, and the floating, middle and sinking signals are decomposed to finally generate 9 independent signals. The three decomposed pressure signals correspond to the floating, middle and sinking data of the intercepted pulse wave signal from small to large.

[0102] d. Pulse data preprocessing: denoising, baseline calibration and stable periodic signal extraction for each signal.

[0103] In this step, each signal is bandpass filtered to filter out system noise interference. Then, baseline calibration is introduced to pull the baseline of the periodic signal to near zero. Finally, the period of the calibrated data is calculated, and 3-5 signals with stable period distribution are extracted as stable periodic signal output.

[0104] e. Pulse data feature extraction: Perform time domain and frequency domain analysis on the stable and effective signal extracted in step d to extract feature data that can reflect the pulse information.

[0105] In this step, the output stable periodic signal is analyzed in time domain and frequency domain. The time domain feature extraction includes the main wave amplitude h1, the dicrotic wave amplitude h3, the pulse output SV and the average pulse period t, and the frequency domain feature extraction includes the reflection coefficient Rf, the propagation velocity PWV and the wavelet coefficient Wf.

[0106] Furthermore, the pulse output SV=coef×t / (K×K), wherein the coefficient coef=0.283, the difference K=Ps-Pd, and Ps and Pd are the highest peak value and baseline value of the pulse wave, respectively.

[0107] Reflection coefficient Rf = (P1-Pmin) / P1, where P1 is the spectrum amplitude corresponding to the first harmonic in the spectrum diagram of the signal after Fourier transform, and Pmin is the spectrum amplitude corresponding to the first minimum value of the harmonic.

[0108] The propagation velocity PWV = 4×fmin×L, where fmin is the frequency corresponding to the first minimum value of the harmonic, L is the distance the reflected wave is transmitted from the terminal to the radial artery, generally defined as 0.8m; the wavelet coefficient Wf is the high and low frequency coefficients obtained after the signal is transformed by wavelet.

[0109] The wavelet coefficients Wf are the high and low frequency coefficients obtained after the signal is transformed by wavelet.

[0110] f. Pulse data feature fusion: The feature data of each channel in step e are spliced ​​to form fusion feature data that can reflect the pulse type of the current patient, and together with the calibration value of the TCM calibration data, form a complete training set data. Diagnose and sample multiple patients to form a complete training set.

[0111] g. Artificial intelligence classification and recognition algorithm modeling: The training set is based on the machine learning qualitative analysis modeling algorithm, and cross-validation is used to traverse and optimize the best model, and the optimal model is used as the final pulse classification model.

[0112] The pulse classification model optimization is preferably implemented using an artificial intelligence classification and recognition algorithm selected from the group consisting of LDA algorithm, logistic regression algorithm, support vector machine algorithm, decision tree algorithm, random forest algorithm, Xgboost algorithm, LightGBM algorithm, ANN algorithm, CNN algorithm, DNN algorithm and neural network algorithm, based on the leave-one-out method, and grid search to find the model with the best f1 index in the current training set.

[0113] In this step, the training data is divided into training set and validation set in a ratio of 8:2. Based on the logistic regression classification algorithm, grid search is used to select the model with the best classification index f1 value as the optimal model output.

[0114] h. Package the pulse classification model in the specified format and upload it to the platform for backup.

[0115] In this step, the pulse classification model is packaged in joblib format and uploaded to the platform for standby. The cloud platform binds the pulse classification model to the sampling terminal. During the application process, the cloud platform can directly call the pulse classification model according to the information of the sampling terminal to complete the pulse category detection of the patient to be detected, thereby obtaining the result value of the patient's pulse category.

[0116] The above embodiment is based on flexible pressure sensor detection technology, using the time-frequency domain feature fusion discrimination method, completing the training set sample data collection through the doctor's precise pulse diagnosis, and creating a pulse classification model to complete the pulse detection of the patient. Compared with the existing detection scheme, this method can effectively improve the reliability and stability of the sampling data. The introduction of the time-frequency domain feature fusion discrimination method can also greatly improve the accuracy of the model, which has a positive role in promoting the inheritance and standardization of traditional Chinese medicine.

[0117] The present invention discloses a detection system and method for distinguishing TCM pulse based on the fusion of time-frequency domain features. The detection system solves the problems of inaccurate positioning of airbag pressurization equipment and mechanical solidification of floating, middle and sinking pressures, resulting in low detection accuracy. This detection system greatly compensates for the differences between patients and doctors through the TCM pulse diagnosis method, can effectively ensure the validity and stability of pulse data, and greatly improves the accuracy of pulse detection. The detection method for distinguishing TCM pulse based on the fusion of time-frequency domain features further improves the validity of pulse data and the objectivity of detection results, which can help the inheritance and sustainable development of TCM.

[0118] Although the present invention is described herein with reference to the illustrative embodiments of the present invention, the above embodiments are only preferred embodiments of the present invention, and the embodiments of the present invention are not limited to the above embodiments. It should be understood that those skilled in the art can design many other modifications and embodiments, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A detection method for distinguishing TCM pulse based on time-frequency domain feature fusion, characterized in that: include: Collect the pulse wave fluctuation pressure and the timing signal of the doctor's pulse pressure during the doctor's pulse diagnosis process; Preprocess the collected time series signals and extract representative feature information of the time series signals for fusion; The TCM calibration data and the fused feature data are combined into a training set, and the pulse classification model is built and trained based on the artificial intelligence classification and recognition algorithm, so that the pulse classification model can be used to distinguish the pulse.

2. The detection method for distinguishing TCM pulse based on time-frequency domain feature fusion according to claim 1 is characterized in that: The construction of the pulse classification model; the specific method includes: Step a, calibrating the collected time series signal data and the pulse type of traditional Chinese medicine diagnosis by calibration values ​​one by one to form complete training data; Step b, extracting and sorting the collected pulse wave fluctuation pressure signals of Cun, Guan and Chi and the doctor's pulse feeling finger pressure signals; Step c, respectively aligning the time sequence of the doctor's pulse pressure signals of Cun, Guan and Chi with the pulse wave fluctuation signals, and at the same time taking the derivative of the pulse wave fluctuation signals, separating the floating, middle and sinking pulse wave fluctuation signals in combination with the gradient change of the pulse wave fluctuation signals, and finally generating nine independent signals; Step d, denoising, baseline calibration and stable periodic signal extraction for each signal; Step e, the stable periodic signal extracted in step d is analyzed in time domain and frequency domain to extract characteristic data that can reflect pulse information; Step f, separately splice and fuse the characteristic data of each signal in step e to form fused characteristic data that can reflect the current patient's pulse type, so that the calibration value of the TCM calibration data and the fused characteristic data together form a complete training set data, and the complete training set is formed by the diagnostic sampling of several patients.

3. The detection method for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: In the step b, the three pulse wave fluctuation pressure signals are sorted out, and the specific method is: effectively analyzing each signal collected, and taking the mean value of the effective signal as the pulse wave fluctuation pressure signal at the current position; The validity analysis refers to calculating the energy spectrum of the pulse wave signal. When the energy value reaches a set threshold and the time domain characteristics show quasi-periodicity, it is determined to be a valid signal. The quasi-periodicity refers to the fact that the pulse wave period values ​​present a certain similarity or alternating regularity, that is, it is determined to be a quasi-periodic signal.

4. The detection method for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: In the step c, the timing is aligned, and the specific method is: aligning the finger pressure signal and the pulse wave signal based on the horizontal coordinate of the sampling time, and determining the waveform intervals of the floating, middle and sinking signals with reference to the finger pressure signal waveform trend; The pulse wave fluctuation signal is separated by taking the first-order derivative of the pulse wave based on the determined waveform interval, and the point where the waveform slope fluctuates abnormally is used as the cutoff point for separation.

5. The detection method for distinguishing Chinese medicine pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: In the step d, each signal is subjected to denoising, baseline calibration and stable periodic signal extraction; the specific method is: The signal denoising uses a bandpass filter to filter out environmental noise greater than 30 Hz and noise introduced by human movement or breathing less than 0.5 Hz; The baseline calibration is to perform baseline calibration on periodic signals with different pulse wave fluctuation pressures by extreme point detection combined with partial least squares fitting; The stable periodic signal extraction will introduce a large number of interference signals due to the doctor's flexibly adjusting the finger position and the interference of unknown factors such as the patient and the surrounding environment during the sampling process. It is necessary to analyze the distribution of the collected pulse signals based on the waveform periodic data and eliminate the signals that seriously deviate from the periodic distribution law.

6. The detection method for distinguishing TCM pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: In the step e, the extracted time domain characteristic data include the main wave amplitude h1, the dicrotic wave amplitude h3, the pulse output SV and the average pulse period t; Among them, pulse output SV = coef × t / (K × K); coefficient coef = 0.283; difference K = Ps-Pd, Ps and Pd are the highest peak value and baseline value of the pulse wave respectively.

7. The detection method for distinguishing TCM pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: In the step e, the extracted frequency domain feature data includes the reflection coefficient Rf, the propagation velocity PWV and the wavelet coefficient Wf; Wherein, the reflection coefficient Rf = (P1-Pmin) / P1; P1 is the spectrum amplitude corresponding to the first harmonic in the spectrum diagram after the signal is Fourier transformed, and Pmin is the spectrum amplitude corresponding to the first minimum value of the harmonic; The propagation velocity PWV = 4 × fmin × L; fmin is the frequency corresponding to the first minimum value of the harmonic; L is the distance that the reflected wave is transmitted from the terminal to the radial artery; the wavelet coefficient Wf is the high and low frequency coefficients obtained after the signal is transformed by wavelet.

8. The detection method for distinguishing TCM pulse based on time-frequency domain feature fusion according to claim 2 is characterized in that: The construction of the pulse classification model; the specific method also includes: g. Based on the machine learning qualitative analysis modeling algorithm, the training set is used to find the best model through cross-validation, and the optimal model is used as the final pulse classification model; h. Package the pulse classification model in the specified format and upload it to the platform for backup.

9. A detection system for distinguishing TCM pulse based on time-frequency domain feature fusion, characterized in that: The detection method for realizing the TCM pulse condition identification based on the fusion of time-frequency domain features as claimed in claim 1 comprises: A data acquisition module is used to collect the pulse wave fluctuation pressure and the timing signal of the pressure of the doctor's finger pressure during the doctor's pulse diagnosis process; A signal processing module, used for preprocessing the collected time series signal; Feature extraction module, used to extract representative feature information of time series signals and fuse them; Intelligent discrimination module, which is used to combine TCM calibration data and feature fusion data into a training set, and to build and train the pulse classification model based on artificial intelligence classification and recognition algorithm; The data storage module is implemented based on the cloud platform and is used to store the original collected pulse wave fluctuation pressure and the pressure data of the doctor's pulse feeling and finger pressure, the pulse classification model and the classification recognition results; The control display module is used to control the parameter configuration, real-time waveform display and pulse condition identification result display of the data acquisition module.

10. The detection system for distinguishing TCM pulse based on time-frequency domain feature fusion according to claim 9 is characterized in that: The data acquisition module has a structural form including a finger sleeve and a wristband, a signal processing unit is integrated in the wristband, a flexible pressure sensor is integrated at the finger tip of the finger sleeve, and a transmission line between the signal processing unit and the flexible pressure sensor is distributed on the back of the hand in a wiring manner with adjustable length.

Images