CN114145730B - Vital sign monitoring action removing method based on deep learning and radio frequency sensing - Google Patents
Vital sign monitoring action removing method based on deep learning and radio frequency sensing Download PDFInfo
- Publication number
- CN114145730B CN114145730B CN202111665237.8A CN202111665237A CN114145730B CN 114145730 B CN114145730 B CN 114145730B CN 202111665237 A CN202111665237 A CN 202111665237A CN 114145730 B CN114145730 B CN 114145730B
- Authority
- CN
- China
- Prior art keywords
- vital sign
- data
- action
- radio frequency
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000009471 action Effects 0.000 title claims abstract description 102
- 238000012544 monitoring process Methods 0.000 title claims abstract description 62
- 238000000034 method Methods 0.000 title claims abstract description 51
- 238000013135 deep learning Methods 0.000 title claims abstract description 29
- 238000012545 processing Methods 0.000 claims abstract description 10
- 238000013528 artificial neural network Methods 0.000 claims description 64
- 238000012549 training Methods 0.000 claims description 50
- 238000005070 sampling Methods 0.000 claims description 39
- 239000011159 matrix material Substances 0.000 claims description 32
- 230000006870 function Effects 0.000 claims description 26
- 238000012360 testing method Methods 0.000 claims description 18
- 230000003068 static effect Effects 0.000 claims description 13
- 238000007781 pre-processing Methods 0.000 claims description 11
- 230000000052 comparative effect Effects 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 5
- 238000001514 detection method Methods 0.000 claims description 5
- 230000029058 respiratory gaseous exchange Effects 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 230000000241 respiratory effect Effects 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 6
- 238000000926 separation method Methods 0.000 abstract description 4
- 230000007547 defect Effects 0.000 abstract description 3
- 230000004044 response Effects 0.000 description 12
- 230000000694 effects Effects 0.000 description 10
- 238000004458 analytical method Methods 0.000 description 7
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000012880 independent component analysis Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000037081 physical activity Effects 0.000 description 3
- 230000000704 physical effect Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000036962 time dependent Effects 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000001747 exhibiting effect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000012847 principal component analysis method Methods 0.000 description 2
- 101100498818 Arabidopsis thaliana DDR4 gene Proteins 0.000 description 1
- 206010011409 Cross infection Diseases 0.000 description 1
- 206010029803 Nosocomial infection Diseases 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 239000008280 blood Substances 0.000 description 1
- 210000004369 blood Anatomy 0.000 description 1
- 230000036772 blood pressure Effects 0.000 description 1
- 230000036760 body temperature Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000001168 carotid artery common Anatomy 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000009532 heart rate measurement Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008450 motivation Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000001020 rhythmical effect Effects 0.000 description 1
- 238000006748 scratching Methods 0.000 description 1
- 230000002393 scratching effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves
- A61B5/0507—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves using microwaves or terahertz waves
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
- A61B5/1118—Determining activity level
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Surgery (AREA)
- Evolutionary Computation (AREA)
- Animal Behavior & Ethology (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Pathology (AREA)
- Heart & Thoracic Surgery (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physiology (AREA)
- Signal Processing (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Psychiatry (AREA)
- Computing Systems (AREA)
- Radiology & Medical Imaging (AREA)
- Software Systems (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Fuzzy Systems (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Dentistry (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a vital sign monitoring action removing method based on deep learning and radio frequency sensing. By applying the method, the influence of motion on non-contact vital sign monitoring can be eliminated, fine-granularity vital sign waveforms can be recovered correctly, the robustness of vital sign monitoring on motion is improved, and the defect that the waveform separation technology at the present stage is limited to linear signals is overcome. In addition, the method samples the sensing data captured by the radar on the fast and slow time scales, converts the data captured by various radars into a unified format, and is convenient for subsequent signal processing, so that the vital sign monitoring method provided by the invention is independent of the bottom hardware.
Description
Technical Field
The invention relates to the technical field of artificial intelligence deep learning. In particular to a vital sign monitoring action removing method based on deep learning and radio frequency sensing.
Background
Vital signs or signs include various indicators related to the health of the human body such as respiration, body temperature, pulse, blood pressure, etc. Currently, the vital sign monitoring technology based on radio frequency sensing has two important branches of contact type and non-contact type vital sign detection technology, in the vital sign monitoring technology based on radio frequency sensing, a radio frequency signal transmitting source transmits a radio frequency signal to a monitored person, and micro motion caused by the vital sign of the body surface of the monitored person can influence the amplitude and the phase of a reflected signal, that is, vital sign information of the monitored person is superimposed in the reflected signal, so that the vital sign of the monitored person can be detected through analysis processing of the reflected signal. Although the radio frequency is more tolerant to background noise than the application of camera, ultrasonic and other means to monitor vital signs of a person, weak vital sign signals may be severely disturbed by intense physical movement or even submerged, and accurate vital sign parameters cannot be obtained. In addition, radio frequency reflected signals affected by body movements and vital sign activities exhibit complex statistical features that cannot be easily separated by a single type of algorithm. Most of the non-contact vital sign monitoring at the present stage requires that the testee is in a relatively static state, the testee cannot normally move when the vital sign monitoring is carried out, and long-time keeping of the static state not only brings discomfort to the testee, but also causes psychological stress to the testee, so that continuous monitoring is difficult. Therefore, how to eliminate the influence of the motion on the non-contact vital sign monitoring, correctly recover the fine-grained vital sign waveform, complete the motion removal, and improve the robustness of the non-contact vital sign monitoring on the motion becomes a problem to be solved urgently.
Disclosure of Invention
The invention aims to solve the defects in the prior art, and provides a vital sign monitoring action removing method based on deep learning and radio frequency sensing, which can eliminate the influence of motion on non-contact vital sign monitoring and correctly recover fine-granularity vital sign waveforms. The current phase of non-contact vital sign monitoring mostly requires that the subject be in a relatively stationary state, because the body movements of the subject may affect the outcome of the vital sign monitoring, and vital sign signals may be severely disturbed by the body movements or even submerged. The vital sign monitoring action removing method overcomes the defect, improves the robustness of vital sign monitoring to movement, and has certain practicability.
The aim of the invention can be achieved by adopting the following technical scheme:
a vital sign monitoring action removal method based on deep learning and radio frequency sensing, the vital sign monitoring action removal method comprising the steps of:
S1, acquiring a radio frequency signal, wherein the radio frequency signal is from a radar and comprises vital sign information; sampling the radio frequency signals in a fast time dimension and a slow time dimension to obtain a radio frequency data matrix H in a preset format.
In radar-based vital sign monitoring, the rf reflected signal acquired from the radar, i.e. the radar impulse response CIR signal, is a modulated time-frequency signal, which needs to be sampled first in order to process the impulse response CIR signals from different radars in a uniform data format, in particular as follows:
In sampling the impulse response CIR from radar, there are typically two different scales of sampling times: t and t ', t' are small-scale sampling times, the sampling rate is usually in the picosecond level, which is called fast time, and can be used for representing the condition of reflected waves at different distances; t=t' +it, a large-scale sampling time, called slow time, where i=1, 2,..n, a time index of slow time sampling, N, a given maximum number of slow samples, T, a sampling period of large-scale sampling, t=1/fs, a slow time sampling frequency, and sampling the radar impulse response CIR in the two dimensions of the fast time and the slow time respectively with preset sampling times, so that a uniform-format radio frequency data matrix H which is an N multiplied by M matrix and is convenient for subsequent processing can be obtained, wherein the matrix is expressed as H i,j, j=1, 2, M represents a sequence index of the fast time, and M is a given maximum fast sampling time.
S2, preprocessing the radio frequency data matrix: carrying out noise reduction treatment on the radio frequency data matrix by using a two-dimensional moving average method, obtaining vital sign data containing action information, removing a static background of the vital sign data by using an average method, and obtaining vital sign data to be treated; processing the vital sign data to be processed by using a constant false alarm detection method to obtain action data; the method comprises the following steps:
S21, intercepting a data block with a corresponding size from a matrix H by using a sliding window with a preset size as first operation data, wherein the width of the sliding window is M, the length of the sliding window is N/N1, and N1 is a preset positive integer, so that the first operation data is a matrix with the size of N/N1 multiplied by M;
s22, calculating a column average value of the first operation data to obtain a first average value vector;
s23, calculating a difference value between the first operation data and the first average value, and removing a static background to obtain second operation data;
S24, calculating the absolute value of the second operation data, and then calculating the column average value for each column to obtain a second average value vector: setting a threshold based on a maximum value of the second mean vector element; scanning the second average value vector based on a threshold value, and finding out a start index and an end index which are larger than the threshold value; taking M as the length of the second sliding window, taking the difference value between the start index and the end index as the width, intercepting a smaller two-dimensional matrix from the second operation data, and extracting the action data to obtain a third operand.
S25, sliding the sliding window along the slow time dimension by a preset step length, and repeating the steps S21 to S24 until all the preprocessed radio frequency signals are processed. Splicing the second operand obtained each time along the slow time dimension, keeping the array number unchanged, obtaining vital sign data with static background removed, and marking the vital sign data as vital sign data to be processed; correspondingly, all third operands are spliced to obtain action data.
S3, identifying body action classification based on the action data, and acquiring action types:
Common body movements can be divided into three types, i.e. stationary, cyclostationary and non-stationary, and any body movement can be formed by a combination of these three types of movements. Essentially, both smooth and cyclosmooth-like actions can be regarded as smooth, i.e. having time-dependent properties, in particular exhibiting repeatability. Thus, stationary and non-stationary types of motion can be distinguished by autocorrelation, the autocorrelation value of the stationary signal has little attenuation, whereas the attenuation of the autocorrelation value of the non-stationary signal is significant. The specific operation is as follows:
Acquiring body motion data from the step S2, and obtaining an autocorrelation value for each line of the motion data;
Calculating autocorrelation attenuation values, calculating the average value of all autocorrelation attenuation values, and taking the average value of the attenuation values as a final autocorrelation attenuation value;
a predetermined attenuation threshold is set, a hypothesis test is constructed, which is a non-stationary motion when the autocorrelation attenuation value is greater than the attenuation threshold, and a stationary motion otherwise.
Outputting action types, wherein the action types comprise non-stationary actions and stationary actions;
s4, constructing a depth comparison self-learning neural network based on action types;
In the case of contactless vital sign monitoring using radar, vital signs are difficult to detect in the presence of body motion, because the radar echo is highly nonlinear in its composition, and therefore, based on source separation algorithms, in particular the independent component analysis ICA method, in which the source signal superposition is linear superposition, vital sign signals cannot be separated from the radar echo in the presence of body motion. Given a set of source signals X (t), the function Y (t) =f (X (t)) may be highly complex and nonlinear, where f represents a nonlinear function. If the inverse function f -1 of f can be found, the source signal Y (t) can be derived from the nonlinear function Y (t); however, trying to solve the inverse function of f under non-linear conditions may lead to an infinite number of solutions, so that traditional signal separation methods based on linearity are no longer applicable when there is body motion.
Recent advances in deep learning have shown that under certain specific conditions, with deep contrast learning, it is possible to find the inverse function f -1 of f when there is body motion, in particular when the signal sources X (t) are independent and time dependent, i.e. the body motion is a cyclostationary motion. The purpose of contrast learning is to approximate the inverse function f -1 by maximizing the difference between the observed and constructed instances.
Because the non-stationary motion is generated in a burst form, the repeatability similar to stationary motion does not exist, the same neural network structure cannot be adopted for removing stationary and non-stationary motion, two deep contrast learning neural networks h and h 'are respectively constructed based on body motion types, a multi-layer perceptron MLP model is used for the two deep contrast learning neural networks h and h', a six-layer perceptron is used for the neural network h suitable for stationary motion, the output of the h is connected with a classifier g (h) formed by the two-layer perceptrons, and the last two-layer classifier uses cross entropy as a loss function; the neural network h ' suitable for non-stationary action is provided with 5 layers of perceptrons, the output of h ' is connected with four classifiers g (h '), the four classifiers are also composed of two layers of perceptrons, and the last four classifiers use cross entropy as a loss function. To increase the nonlinearity, an activation function leakyReLU is set after the output layer of each layer of sensor.
S5, training a depth contrast self-learning neural network;
s51, preparing a training and testing sample set according to action types:
constructing a sample set suitable for smooth actions:
construction of observation sample Z (n): Wherein A (n-1) is a time unit for delaying A (n), and A (n) is a vital sign data matrix containing action information obtained by sampling and preprocessing radar echo signals;
Next, a comparative example sample Z (n) was constructed, the order of the A (n) sequences was all disordered, labeled A (n), and A (n-1) was replaced to obtain The observed samples were very different from the comparative examples in the lower half, each observed sample was labeled 1 and each comparative sample was labeled-1.
The data of the observation sample and the contrast sample are constructed to form a sample set, and the training sample and the test sample are divided according to a certain proportion so as to train the neural network, for example, hundred thousand groups of sample data are constructed, three hundred thousand groups of data are divided into a test set according to a proportion of 30%, and the rest of the hundred thousand groups of data are divided into a training set.
Constructing a sample set suitable for non-stationary actions:
construction of observation sample Z' (n): z ' (n) = [ A ' (n) ], unlike the observed sample Z (n) suitable for smooth motion, the method does not need data expansion, does not need a comparative sample, and is a vital sign data matrix containing motion information obtained by sampling and preprocessing radar echo signals, wherein A ' (n) is the same as A (n);
Next, the observation samples were split into T aliquots, e.g., t=4, resulting in a total of 4 observation samples, directly labeled into 4 different categories. According to the invention, thirty-thousand groups of samples are respectively generated for each category, a total of 12 ten-thousand groups of samples form a sample set, thirty-thousand groups of data are divided into test sets according to a proportion of 30%, and the rest ninety-thousand groups of data are training sets.
S52, training the depth contrast self-learning neural network according to the training sample set:
According to different action types, respectively sending training sample data corresponding to the action types into the built depth contrast self-learning neural networks, and training the neural networks on a Tensorflow platform: the neural network parameters and weights are initialized first, then the random gradient descent and error back propagation method is used to minimize the loss function, and the neural network parameters and weights are updated to achieve the optimal network parameters. And (3) sending a group of data training data each time during training until all sample data are input, and finishing the training. The parameters set include the number of samples selected for one training Batch size, learning rate LEARNING RATE, momentum, decay step DECAY STEP, and decay factor.
S6, applying the trained deep learning neural network to finish action removal: and sending the vital sign data to be processed containing vital information into a trained deep learning neural network, completing action removal, and recovering fine-granularity vital sign signals, wherein the vital signs contain respiratory signals and/or heart rate signals.
Further, in the step S2, the vital sign data and the motion data to be processed may also be obtained by a nonlinear ICA principal component analysis method.
Further, in the step S4, the depth contrast learning neural network may be further constructed based on a CNN convolutional neural network model.
Further, in the step S4, a minimum mean square error MSE may be used as the loss function.
Further, in the step S4, the activation function may also use a Sigmoid function.
Further, in the step S5, the Xavier method may be further applied to train the neural network, and initialize the parameters and weights of the neural network.
Further, in the step S5, a small-batch sample gradient descent method may be further applied to minimize the loss function and update the neural network parameters and weights.
Compared with the prior art, the invention has the following advantages and effects:
The invention provides a vital sign monitoring action removing method based on deep learning and radio frequency sensing, which can well eliminate the interference of movement on vital sign monitoring and correctly recover fine-granularity vital sign waveforms by applying deep contrast learning to non-contact vital sign monitoring. The robustness of non-contact vital sign monitoring is improved. The testee gets rid of the constraint of wearing equipment when carrying out the vital sign monitoring, also need not to keep static when accepting the vital sign monitoring, both saved wearing equipment's trouble of wearing and taking off, avoided wearing equipment uncomfortable sense and cross infection's risk to a great extent, can carry out normal activity when accepting the monitoring again, love the testee and just accomplish the monitoring when normal motion carelessly, whole monitoring process is comfortable, automatic, intelligent, experience sense is splendid, monitoring result can also real-time remote monitoring.
In particular, the advantages and effects of the present invention are exhibited in the following aspects:
(1) The method for monitoring vital signs aiming at motion robustness is provided and realized, and information acquisition and processing are realized in a non-contact mode based on a radio frequency sensing and deep contrast learning technology;
(2) Sampling radar radio frequency signals on a fast time scale and a slow time scale, converting sensing data captured by various radars into a unified format, facilitating subsequent signal processing, and enabling the vital sign monitoring system provided by the invention to be independent of a bottom hardware platform;
(3) The motion is further classified by analyzing the relation between the human motion and the tiny vibration excited by vital signs, different body motion types are distinguished, a neural network is constructed and trained, and the effectiveness of motion removal in life monitoring is improved;
(4) The method expands a recently developed deep contrast learning framework, separates nonlinear signal components in a self-supervision mode, completes action removal, and improves the robustness of vital sign monitoring to movement.
Drawings
FIG. 1 is a general flow chart of a vital sign monitoring action removal method based on deep learning and radio frequency sensing disclosed in the present embodiment;
Fig. 2 is a TI IWR 1443 beamforming spectrogram disclosed in the present embodiment;
FIG. 3 is a schematic diagram of the relative errors in conducting vital signs while stationary and playing a smart phone game according to the prior art;
FIG. 4 is a graph of relative error of heart rate under various physical actions as disclosed in this example;
FIG. 5 is a graph of normalized correlation coefficients of vital sign waveforms under various body actions as disclosed in this embodiment;
Fig. 6 is a schematic diagram of a neural network structure suitable for smooth motion disclosed in this embodiment;
FIG. 7 is a schematic diagram of a neural network suitable for non-stationary motion as disclosed in this embodiment;
Fig. 8 is a neural network training flowchart disclosed in this embodiment.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Example 1
The deep learning concept is derived from the research of an artificial neural network, is a new field in machine learning research, and has the motivation of building and simulating a neural network for analysis learning of the human brain, and imitates the mechanism of the human brain to interpret data. The deep neural network comprises a neural network structure of a plurality of hidden layers. The layers are typically arranged in order, each layer consisting of a number of original non-linear operations, and the input representation of the upper layer may be passed to the next layer and converted to a more concise representation. Deep learning forms a more abstract high-level representation by combining low-level features to find a distributed feature representation of the data. In the research process, researchers consider that deep contrast learning is applied to non-contact vital sign monitoring, so that the influence of movement on vital sign monitoring can be well eliminated, fine-granularity vital sign waveforms can be correctly recovered, and the robustness of non-contact vital sign monitoring on movement is improved.
In order to deepen the further understanding of the invention, the analysis of the radio frequency response of the radar, the analysis of the body action type, the influence of the body action on the vital sign monitoring and the analysis of the movement on the signal space are described as follows, and it is to be noted that the description is only for better understanding of the invention and not for limiting the invention.
(1) Analysis of radio frequency response of radar
The radio frequency sensor is a key component for non-contact vital sign monitoring, and in the monitoring process, the radio frequency generator transmits radio frequency signals to a monitored person, the radio frequency signals can be reflected back when encountering the monitored person, and the limb movement and vital sign activity of the monitored person can influence the amplitude and the phase of the reflected signals. A radio frequency sensor, such as a radar, is a detection device operating in a radio frequency band, and is used to convert detected information into an electrical signal or other information output in a desired form according to a certain rule. The radio frequency response may be represented by an impulse response CIR, and the time delay due to the distance change is a major factor affecting the radio frequency channel impulse response. The distance is expressed as follows:
Wherein t is the time of the time, D b(t)、dr(t)、dh (t) is the average distance from the radar to the measured object, and is the distance change caused by body movement, respiration and heartbeat respectively.
For a given transmit waveform s (t), the received signal y (t) is expressed as:
Where c is the speed of the radio wave, f c is the carrier frequency, and α (t) is the channel gain.
For reflected wave signals from any radar, such as IR-UWB radar or FMCW radar, the impulse response CIR can be extracted from y (t)/s (t), which contains any information reflected by the subject, including vital sign information and body motion information.
Both IR-UWB radar and FMCW radar are equipped with single-antenna and multiple-antenna systems. The multi-antenna system can provide more spatial diversity relative to a single-antenna system, so that the wave beam of electromagnetic waves is narrower and has better performance. The action removing technology provided by the invention can be well used under the condition of a single antenna, and can achieve better effect by combining a multi-antenna system.
As shown in the effect diagram of millimeter wave multi-antenna beamforming in fig. 2, it can be seen that the beam becomes narrower as the antennas increase.
(2) Body action type analysis
We can divide common physical actions into three types: i.e. stationary, cyclostationary and non-stationary, any body motion may be formed by some combination of these types. Stationary motion is similar to gaussian noise but is not gaussian noise, and playing games using a cell phone is a typical stationary motion, which also includes typing and leg jitter. The cyclostationary type generally includes various rhythmic body movements, such as swinging the body, and walking on a treadmill is one of the cyclostationary type movements. Whereas sudden standing or sitting movements are classified as jerky movements due to their sudden nature, another typical jerky movement is a flip of the body during sleep.
(3) Influence of body movements on vital sign monitoring
To study the effect of physical activity on vital sign monitoring we measured heart rates at six typical physical activities using a template matching method and calculated the respective relative errors as shown in figure 3, where all physical activities produce significant errors in heart rate measurement.
Vital signs are largely independent of body motion and if they are superimposed linearly, the use of a template matching method should enable correct extraction of vital signals. In order to verify whether linearity is established, a person to be tested wears a wearable device, synchronously receives signals of a radar, and removes noise from the signals obtained through the wearable device by using a singular value decomposition method to obtain a reference waveform. If the superposition of body motion and vital signs is linear, the correlation coefficient of the waveform derived from the radar signal with the reference waveform should be high. We calculate the correlation coefficients of the vital sign waveforms of all the contact devices under body movement with the waveforms of the radar signals and normalize them with the respective static correlation coefficients; the results of fig. 4 show that the correlation coefficient is close to 0.1 in most cases, thus proving that the radar signal superposition caused by human body motion is far from linear. Therefore, the existing waveform separation method based on the linear superposition of the signal components is supposed to fail, that is, the reflected signal caused by the body movement may exhibit various statistical characteristics and cannot be easily separated by a single type of algorithm.
(4) Influence of motion on signal space
The signal frequencies caused by vital signs are low, typically below 2Hz, and cannot be captured in a fast time, but are hidden in some slow time sequence, mainly affecting the phase of the radar signal. The signal caused by the body motion has wider bandwidth and higher frequency, can be captured in a fast time and a slow time at the same time, and mainly affects the amplitude of the radar signal. Studies have shown that the effect of body movements on radar radio frequency signals is much more complex than the effect of vital signs on radar radio frequency signals. The present invention uses extended versions of RF-SCG waveform extraction techniques to monitor vital signs of a subject while stationary and playing a smart phone game on IR-UWB radar and FMCW radar, respectively. Relative error as shown in fig. 5, it can be seen from fig. 5 that the performance of vital sign monitoring is significantly reduced when the subject plays a smart phone game, which indicates that physical actions may result in reduced effectiveness of vital sign monitoring.
The invention discloses a vital sign monitoring action removing method based on deep learning and radio frequency sensing. The radar radio frequency echo signals containing vital sign information are sampled and preprocessed on a fast scale and a slow scale to obtain vital sign data and action data to be processed; identifying a type of action based on the action data and the autocorrelation; constructing sample sets applicable to different action types based on vital sign data to be processed, and scratching a training sample set and a test sample set according to a certain proportion, for example, dividing 70% of the sample sets into training sample sets and dividing 30% of the sample sets into test sample sets; and taking the training sample as the total input of the whole neural network, and updating and adjusting the weight and the parameters of the neural network by using a random gradient descent and error back propagation method to complete the training of the neural network. And finally, inputting the test sample into a trained deep learning neural network, completing action removal, and outputting fine-granularity vital sign signals including respiratory signals and/or heart rate signals.
In the practice of the invention, radar is preferably placed in front of the subject, since the pulse BVP of blood volume associated with the heart beat is likely to be caused by the common carotid artery, while the respiratory signal is mainly dependent on chest vibration. It has been shown that aiming the radar at the body side misses the breathing signal to a large extent, but not the heartbeat signal.
Example 2
The embodiment discloses a method for recovering vital sign monitoring waveforms based on deep learning and radio frequency sensing, which comprises two parts of model training and model testing.
The implementation of the invention will be described in detail below using IR-UWB radar X4M05 as an example.
In order to obtain sufficient sample data, a total of 12 healthy subjects, aged 15-64 years, with a weight of between 50-80 kg, were enrolled for 6 men and 6 women. Subjects were required to maintain a quasi-static sitting position in a daily living environment, or to perform 7 common human actions: playing a mobile phone, typing, swinging the body, shaking the legs, walking on a running machine, standing/sitting, turning over (while sleeping). The rf-sensing radar is placed in a range of 0.5 to 2m from the subject, the precise range may vary from person to person. Data collection was performed using different time spans, but ensuring that the total time per subject was approximately the same: including one minute, one hour of typing, and one night of sleep monitoring on a treadmill, for a total of 80 hours of RF data, including about 330k heart beat cycles and 68k breath cycles. 30% of the data are used as training sample data for offline training of the deep learning module, and the remaining 70% of the data are used as test sample data for online recovery of vital sign waveforms by using the training module. According to different body actions, training sample data and test sample data are applied to respectively perform experiments, wherein the experiments are performed on a PC (personal computer) based on Python 3.7 and TensorFlow 2.0.0, and the PC is provided with an i9-10900KF (3.7 GHz) CPU, a 16GB DDR4 RAM and a GeForce RTX 2070 display card. The clocks between hardware components are synchronized based on a precision time protocol using ethernet. Novelda's IR-UWB radar X4M05 operates at 7.3 or 8.7GHz, with a bandwidth of 1.5GHz; it has a pair of tx-rx (transmitter, receiver) antennas, a field of view (FoV) 65 ° azimuth and elevation.
The whole implementation flow is shown in fig. 1, and the specific implementation steps are as follows:
s1, acquiring a radio frequency signal from a radar, sampling the radio frequency signal in a fast time dimension and a slow time dimension, and acquiring a radio frequency data matrix in a preset format;
in sampling the impulse response CIR from radar, there are typically two different scales of sampling times: t and t ', t' are small-scale sampling times, the sampling rate is usually in the picosecond level, which is called fast time, and can be used for representing the condition of reflected waves at different distances; t=t' +it, a large-scale sampling time, called slow time, where i=1, 2,..n, a time index of slow time sampling, N, a given maximum number of slow samples, T, a sampling period of large-scale sampling, t=1/fs, a slow sampling frequency, and so on, and sampling the radar impulse response CIR in the two dimensions of the fast time and the slow time respectively with preset sampling times to obtain a radio frequency data matrix H which is an N multiplied by M matrix and is convenient for subsequent processing, wherein the matrix is expressed as H i,j, j=1, 2. In this embodiment, n=2000, m=100, fs=512 Hz.
S2, preprocessing the radio frequency data matrix sampled and output from the step S1: the application method comprises the steps of carrying out noise reduction on the radio frequency data matrix, obtaining vital sign data containing action information, removing the static background of the vital sign data by using an average value method, and obtaining vital sign data to be processed; processing vital sign data to be processed by using a constant false alarm detection method to obtain action data; the method comprises the following steps:
S21, a sliding window with the width of 100 and the length of 200 is used for intercepting a data block with the corresponding size from a matrix H as first operation data, so that the first operation data is a matrix with the size of 200 multiplied by 100;
S22, calculating a column average value of the first operation data to obtain a first average value vector containing 100 elements;
S23, calculating a difference value between the first operation data and the first average value, and removing a static background to obtain second operation data, wherein the second operation data is a matrix with the size of 200 multiplied by 100;
S24, calculating the absolute value of the second operation data, and then calculating the column average value for each column to obtain a second average value vector: acquiring a threshold value based on the maximum value of the second average vector element, wherein in the embodiment, the threshold value=maximum value/5; scanning the second average vector based on a threshold value, finding a start index and an end index greater than the threshold value, such as 35 and 50; taking 200 as the length of the second sliding window and the difference value 15 of the start index and the end index as the width, intercepting a smaller two-dimensional matrix from the second operation data, extracting the action data with the matrix size of 200 x 15, and obtaining a third operand;
S25, sliding the sliding window along the slow time dimension by the step length 50, and repeating the steps S21 to S24 until all the preprocessed radio frequency signals are processed. Splicing the second operand obtained each time along the slow time dimension, keeping the array number unchanged, obtaining vital sign data with static background removed, and marking the vital sign data as vital sign data to be processed; correspondingly, all third operands are spliced to obtain action data.
S3, identifying body actions based on the action data, and acquiring action types;
Common body movements can be divided into three types, i.e. stationary, cyclostationary and non-stationary, and any body movement can be formed by a combination of these three types of movements. Essentially, both smooth and cyclosmooth-like actions can be regarded as smooth, i.e. having time-dependent properties, in particular exhibiting repeatability. Thus, stationary and non-stationary types of motion can be distinguished by autocorrelation, the autocorrelation value of the stationary signal has little attenuation, whereas the attenuation of the autocorrelation value of the non-stationary signal is significant. The specific operation is as follows:
acquiring body motion data from the step S2, and obtaining an autocorrelation value for each line of the motion data;
Calculating autocorrelation attenuation values, calculating the average value of all autocorrelation attenuation values, and taking the average value of the attenuation values as a final autocorrelation attenuation value;
setting the attenuation threshold to 0.45, constructing a hypothesis test, and when the autocorrelation attenuation value is greater than 0.45, performing non-stationary motion, otherwise performing stationary motion.
Outputting an action type, wherein the output action type is a stable action or a non-stable action;
s4, constructing a depth comparison self-learning neural network based on action types;
since the non-stationary motion is generated in a burst form, there is no repeatability similar to the stationary motion, and the same neural network structure cannot be adopted for the removal of the stationary and non-stationary motion. As shown in fig. 6, the neural network suitable for smooth motion is composed of a neural network h and a classifier g (h), the neural network h uses six layers of sensors, the output of h is connected with the two layers of sensors g (h), g (h) is composed of two layers of sensors, and the last layer of sensors of the two layers of sensors uses cross entropy as a loss function; the neural network suitable for the non-stationary action is composed of a neural network h ' and a four-layer classifier g (h '), as shown in fig. 7, the neural network h ' is provided with 5-layer perceptrons, the output of the h ' is connected with the four-layer classifier g (h '), the four-layer classifier is also composed of two-layer perceptrons, and the last-layer perceptrons of the four-layer classifier use cross entropy as a loss function. To increase the nonlinearity, an activation function leak ReLU is set after the output layer of the last layer of sensor of each layer.
S5, training a depth contrast self-learning neural network;
551. preparing training and testing sample sets according to action types:
constructing a sample set suitable for smooth actions:
First, an observation sample Z (n) is constructed: Wherein A (n-1) is a time unit for delaying A (n), and A (n) is a vital sign data matrix containing action information obtained by sampling and preprocessing radar echo signals;
Next, a comparative example sample Z (n) was constructed: the order of the a (n) matrix sequences is totally shuffled, marked as a (n), and replaced with a (n-1), resulting in The observed samples were very different from the comparative examples in the lower half, each observed sample was labeled 1 and each comparative sample was labeled-1. The invention constructs a hundred thousand data sample for training the neural network, wherein thirty thousand data are test sets and seventeen thousand data are training sets.
Constructing a sample set suitable for non-stationary actions:
First, an observation sample Z' (n) is constructed: z ' (n) = [ A ' (n) ], wherein A ' (n) is the same as A (n), and is a vital sign data matrix containing action information obtained by sampling and preprocessing radar echo signals, and is different from an observation sample Z (n) suitable for smooth actions, and the data expansion is not needed, and a comparison example sample is not needed;
Next, the observation samples were split into T aliquots, e.g., t=4, resulting in a total of 4 observation samples, directly labeled into 4 different categories. The invention generates thirty-thousand groups of samples for each category, and the total of the thirty-thousand groups of samples. Wherein ninety-thousand sets of data are test sets and thirty-thousand sets of data are training sets.
S52, training the depth contrast self-learning network according to the training sample set:
According to different action types, respectively sending training sample data corresponding to the action types into the built deep comparison learning networks, training the neural network on a Tensorflow platform, wherein the specific training flow is shown in fig. 8, firstly, initializing the neural network parameters and weights, and then, using a random gradient descent and error back propagation method to minimize a loss function, and updating the network parameters and the weights so as to achieve the optimization of the network parameters. When training the neural network, forward propagation and backward propagation are interdependent, parameters and weights of each layer of the neural network are calculated and stored according to the forward propagation sequence, namely, the sequence from an input layer to an output layer, and the calculation of the neural network parameter gradient is performed according to the backward propagation sequence, namely, the sequence from the output layer to the input layer. And (3) sending a group of data training data each time during training until all sample data are input, and finishing the training. In this embodiment, the Batch size is set to 512, LEARNING RATE, momentum, DECAY STEP, and decay factor to 0.001, 0.9, 5e5, and 0.999, respectively.
S6, applying the trained depth contrast self-learning neural network to finish action removal: and sending the vital sign data to be processed containing the vital information into a trained deep learning neural network, completing action removal, and recovering fine-granularity vital sign signals, wherein the vital sign signals contain breathing signals and/or heart rate signals.
In particular, in the above step S2, vital sign data and motion data to be processed may also be obtained by a nonlinear ICA principal component analysis method; in the step S4, the depth comparison self-learning neural network can also be constructed based on a CNN convolutional neural network model; in step S4, sigmoid may also be used as an activation function; in step S4, the minimum mean square error MSE may also be used as a loss function; in step S5, the Xavier method can be also applied to initialize the parameters and the weight of the neural network; in step S5, a small batch sample gradient descent method can be applied to minimize a loss function and update the parameters and weight of the neural network; radio frequency signals may also be acquired by FMCW radar.
The above examples are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above examples, and any other changes, modifications, substitutions, combinations, and simplifications that do not depart from the spirit and principle of the present invention should be made in the equivalent manner, and the embodiments are included in the protection scope of the present invention.
Claims (9)
1. The vital sign monitoring action removing method based on deep learning and radio frequency sensing is characterized by comprising the following steps of:
S1, acquiring a radio frequency signal, wherein the radio frequency signal is from a radar and comprises vital sign information; sampling the radio frequency signals in a fast time dimension and a slow time dimension to obtain a radio frequency data matrix H in a preset format, setting the maximum slow sampling frequency as M, the maximum fast sampling frequency as N and the slow sampling frequency as a preset value, wherein M, N are positive integers;
s2, preprocessing the radio frequency data sequence to obtain vital sign data and action data to be processed;
S3, identifying body action classification based on the action data, and acquiring action types;
s4, constructing a depth comparison self-learning neural network based on the action type; the self-learning neural network is set as follows:
Two deep contrast learning neural networks h and h' are respectively constructed based on body action types, and multi-layer perceptron MLP is used for the two deep contrast learning neural networks, wherein six layers of perceptrons are used for the neural network h suitable for smooth action, the output of the h is connected with a classifier g (h) formed by two layers of perceptrons, and the cross entropy is used as a loss function for the last layer of the two classifiers; the neural network h ' suitable for non-stationary action is provided with 5 layers of perceptrons, the output of the h ' is connected with four classifiers g (h '), the four classifiers are also composed of two layers of perceptrons, the last four classifier uses cross entropy as a loss function, and in order to increase nonlinearity, an activation function leakage ReLU is arranged behind the output layer of each layer of perceptrons;
S5, training the depth comparison self-learning neural network;
S6, applying the trained deep learning self-neural network to finish action removal: and sending the vital sign data to be processed containing vital information into the depth comparison self-learning neural network, completing action removal, and recovering fine-granularity vital sign signals, wherein the vital signs contain respiratory signals and/or heart rate signals.
2. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to claim 1, wherein the preprocessing procedure in step S2 is as follows:
Noise reduction processing is carried out on the radio frequency data matrix based on a two-dimensional moving average method, and vital sign data containing action information is obtained; removing the static background of the vital sign data based on an average value method to obtain vital sign data to be processed; and processing the vital sign data to be processed based on a constant false alarm detection method to obtain action data.
3. The deep learning and radio frequency sensing based vital sign monitoring action removal method according to claim 1, wherein the identifying the body action process in step S3 is as follows:
acquiring body motion data, and solving an autocorrelation value for each row of the motion data;
Calculating an autocorrelation attenuation value;
Setting a preset attenuation threshold, and when the autocorrelation attenuation value is larger than the attenuation threshold, performing non-stationary motion, otherwise performing stationary motion;
and outputting the action type.
4. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to claim 1, wherein training the deep contrast self-learning network in step S5 comprises constructing a training sample set and a test sample set according to action types, wherein the process is as follows:
constructing a sample set suitable for smooth action, wherein the sample set consists of an observation sample and a comparison sample:
construction of an observation sample Wherein A (n-1) is a time unit for delaying A (n), and A (n) is a vital sign data matrix containing action information obtained by sampling and preprocessing radar echo signals; construction of comparative example sample/>All the sequences of a (n) are disordered, denoted as a (n), and a (n-1) is replaced by a (n); marking an observation sample as 1 and a comparison sample as-1; dividing the sample set into a training sample set and a test sample set according to a preset proportion;
constructing a sample set suitable for non-stationary actions:
Constructing an observation sample Z ' (n) = [ A ' (n) ], wherein A ' (n) is a vital sign data matrix containing action information obtained by sampling and preprocessing radar echo signals; dividing the observation sample into T equal parts, and marking the T equal parts into different T categories; generating a predetermined group of samples for each category respectively to form a sample set; the sample set is divided into a training sample set and a test sample set at a predetermined ratio.
5. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to claim 1, wherein training the deep contrast self-learning network in step S5 comprises setting and updating parameters and weights of the neural network by applying a training sample set, and comprises the following steps:
According to different action types, respectively sending training sample data corresponding to the action types into the built deep contrast learning networks to train the neural network, wherein the training sample data comprises the following steps: initializing parameters and weights of a neural network; minimizing a loss function by using a random gradient descent and error back propagation method, and updating parameters and weights of the neural network; when training, a group of training data is sent in each time until all sample data are input, and the training is finished; the parameters set include the number of samples selected for one training Batch size, learning rate LEARNING RATE, momentum, decay step DECAY STEP, and decay factor.
6. The deep learning and radio frequency aware based vital sign monitoring action removal method of claim 1, further comprising step S7: the respiration waveform and/or the heart rate waveform are output and displayed in real time.
7. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to claim 2, wherein the specific steps of preprocessing in step S2 are as follows:
S21, intercepting a data block with a preset size from a matrix H by using a sliding window with the preset size as first operation data, wherein the preset size is as follows: the width is M, the length is N/N1, and N1 is a preset positive integer, so that the first operation data is a matrix with the size of N/N1 multiplied by M;
s22, calculating a column average value of the first operation data to obtain a first average value vector;
s23, calculating a difference value between the first operation data and the first average value, and removing a static background to obtain second operation data;
s24, calculating the absolute value of the second operation data, and then calculating the column average value for each column to obtain a second average value vector: setting a threshold based on a maximum value of the second mean vector element; scanning the second average value vector based on the threshold value, and finding out a start index and an end index which are larger than the threshold value; taking M as the length of a second sliding window, taking the difference value between a start index and an end index as the width, intercepting a smaller two-dimensional matrix from second operation data, and extracting action data to obtain a third operand;
S25, sliding the sliding window along the slow time dimension with a preset step length, and repeating the steps S21 to S24 until all the preprocessed radio frequency signals are processed; splicing the second operand obtained each time along the slow time dimension, keeping the array number unchanged, obtaining vital sign data with static background removed, and marking the vital sign data as vital sign data to be processed; correspondingly, all third operands are spliced to obtain action data.
8. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to any one of claims 1 to 7, wherein in the step S1, the given slow sampling frequency fs is 512Hz, the maximum slow sampling frequency M is 2000, and the maximum fast sampling frequency N is 100.
9. The method for removing vital sign monitoring actions based on deep learning and radio frequency sensing according to any one of claims 1 to 7, wherein the radar in the step S2 is an IR-UWB radar or an FMCW radar.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111665237.8A CN114145730B (en) | 2021-12-30 | 2021-12-30 | Vital sign monitoring action removing method based on deep learning and radio frequency sensing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111665237.8A CN114145730B (en) | 2021-12-30 | 2021-12-30 | Vital sign monitoring action removing method based on deep learning and radio frequency sensing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114145730A CN114145730A (en) | 2022-03-08 |
| CN114145730B true CN114145730B (en) | 2024-05-24 |
Family
ID=80449444
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111665237.8A Active CN114145730B (en) | 2021-12-30 | 2021-12-30 | Vital sign monitoring action removing method based on deep learning and radio frequency sensing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114145730B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114903475A (en) * | 2022-06-20 | 2022-08-16 | 泉州师范学院 | Health detector based on PCANet neural network and multi-sensor information fusion and method thereof |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111568396A (en) * | 2020-04-13 | 2020-08-25 | 广西万云科技有限公司 | V2iFi is based on vital sign monitoring technology in compact radio frequency induction's car |
| CN112052986A (en) * | 2020-08-13 | 2020-12-08 | 五邑大学 | Magnetic storm prediction method and device based on self-supervision learning and storage medium |
| CN113011487A (en) * | 2021-03-16 | 2021-06-22 | 华南理工大学 | Open set image classification method based on joint learning and knowledge migration |
| CN113850204A (en) * | 2021-09-28 | 2021-12-28 | 太原理工大学 | Human body action recognition method based on deep learning and ultra-wideband radar |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10417525B2 (en) * | 2014-09-22 | 2019-09-17 | Samsung Electronics Co., Ltd. | Object recognition with reduced neural network weight precision |
| US20210374553A1 (en) * | 2020-06-02 | 2021-12-02 | Salesforce.Com, Inc. | Systems and methods for noise-robust contrastive learning |
-
2021
- 2021-12-30 CN CN202111665237.8A patent/CN114145730B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111568396A (en) * | 2020-04-13 | 2020-08-25 | 广西万云科技有限公司 | V2iFi is based on vital sign monitoring technology in compact radio frequency induction's car |
| CN112052986A (en) * | 2020-08-13 | 2020-12-08 | 五邑大学 | Magnetic storm prediction method and device based on self-supervision learning and storage medium |
| CN113011487A (en) * | 2021-03-16 | 2021-06-22 | 华南理工大学 | Open set image classification method based on joint learning and knowledge migration |
| CN113850204A (en) * | 2021-09-28 | 2021-12-28 | 太原理工大学 | Human body action recognition method based on deep learning and ultra-wideband radar |
Non-Patent Citations (2)
| Title |
|---|
| 基于神经网络的嵌套命名实体关系抽取模型;陈浩，王兴芬;情报工程;20210731;67-75 * |
| 非接触式生命体征监测技术临床应用研究进展;杨晓丹;杨冰磊;王梓涵;章依妮;周韩俊;李欣荣;曹世华;;医学信息;20180915(18);49-52 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114145730A (en) | 2022-03-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tan et al. | Exploiting WiFi channel state information for residential healthcare informatics | |
| Wang et al. | Device-free wireless sensing: Challenges, opportunities, and applications | |
| Chen et al. | Contactless electrocardiogram monitoring with millimeter wave radar | |
| Yin et al. | ECG monitoring system integrated with IR-UWB radar based on CNN | |
| US10310073B1 (en) | System and method for determining engagement level of a human being using a millimeter-wave radar sensor | |
| Garripoli et al. | Embedded DSP-based telehealth radar system for remote in-door fall detection | |
| US9532735B2 (en) | Apparatus and method for wireless monitoring using ultra-wideband frequencies | |
| CN107837078A (en) | Apparatus and method and wearable device for biometrics infomation detection | |
| Yang et al. | Unsupervised detection of apnea using commodity RFID tags with a recurrent variational autoencoder | |
| EP4442201A1 (en) | Millimeter-wave radar-based noncontact electrocardiogram monitoring method | |
| CN109009124A (en) | Respiratory rate measurement and object localization method based on ULTRA-WIDEBAND RADAR | |
| Ren et al. | Vital sign detection in any orientation using a distributed radar network via modified independent component analysis | |
| CN110520935A (en) | Learn sleep stage from radio signal | |
| CN114145730B (en) | Vital sign monitoring action removing method based on deep learning and radio frequency sensing | |
| Tang et al. | Augmenting experimental data with simulations to improve activity classification in healthcare monitoring | |
| Ding et al. | Human motion recognition with spatial-temporal-convLSTM network using dynamic range-doppler frames based on portable FMCW radar | |
| CN109091126A (en) | A kind of human body physiological state differentiation method for visualizing and equipment based on multisensor | |
| CN111568425A (en) | Non-contact type multi-person respiration detection method | |
| Fioranelli et al. | Contactless radar sensing for health monitoring | |
| US20200341115A1 (en) | Subject identification in behavioral sensing systems | |
| Le et al. | Multivariate singular spectral analysis for heartbeat extraction in remote sensing of uwb impulse radar | |
| Muaaz et al. | WiHAR: From Wi-Fi channel state information to unobtrusive human activity recognition | |
| Mosleh et al. | Monitoring respiratory motion with wi-fi csi: Characterizing performance and the breathesmart algorithm | |
| Gao et al. | Contactless sensing of physiological signals using wideband RF probes | |
| CN114098679B (en) | Vital sign monitoring waveform recovery method based on deep learning and radio frequency sensing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |