A human respiration recognition method based on a millimeter wave body area network

Abstract

The present application relates to the technical field of radio waves, and particularly relates to a human respiration recognition method based on a millimeter wave body area network, which comprises the following steps: adopting the movement state of a chest cavity and an abdominal cavity of a human body during various respiration movements; tagging the collected movement state to obtain a training data set; constructing a neural network model and training the neural network model by using the training data set; collecting the movement state of the chest cavity and the abdominal cavity during a to-be-recognized respiration movement and inputting the movement state into the trained neural network model, so that the neural network model recognizes the type of respiration; the present application fully utilizes the advantage of fixing a millimeter wave antenna on a body, does not rely on the reflection of electromagnetic waves to collect respiration data, and the collected respiration data will not change with the change of the environment, and meanwhile, the present application can recognize various different forms of respiration modes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radio waves, in particular to a human respiration recognition method based on a millimeter wave body area network. BACKGROUND

[0002] Human activity classification has many applications in life, most of which are in the field of medical rehabilitation. With the continuous iteration and update of sensor technology, more and more human activity recognition has been explored by relevant researchers, from the initial large-amplitude limb movement to the small-amplitude breathing movement. Because the amplitude of human respiratory movement is small, and the movement law is relatively single, at the same time, the movement has a high requirement for the sensitivity of the sensor, and some sensors used for monitoring human activity cannot provide accurate resolution. Therefore, researchers usually use radar, Wi-Fi and other wireless devices to emit electromagnetic waves and reflect multiple times with the human body or the surrounding environment to collect breathing data about the human body.

[0003] Chinese patent document CN108553108A discloses a method for detecting multiple human actions and breathing based on CSI signals in Wi-Fi, which extracts CSI signal data of different action postures and breathing of the human body at different positions from the Wi-Fi signal, and performs spatial monitoring according to the signal; Chinese patent document CN112754431A discloses a breathing and heartbeat monitoring system based on millimeter wave radar and lightweight neural network, which uses radar signals for breathing and heartbeat monitoring.

[0004] The above-mentioned radar, Wi-Fi and other devices detect and recognize human respiration based on the principle of radio wave propagation, and use a non-contact measurement method with the antenna placed outside the body. Because other movements of the human body will interfere with the propagation of the electromagnetic wave and thus affect the accuracy of the breathing data, when using a non-contact data collection method, the human body is usually required to remain stationary, which increases the difficulty of data collection. In addition, non-contact measurement has high requirements for the environment, and sudden changes in the external environment will directly affect the correctness of the breathing data, thereby reducing the accuracy of breathing detection and recognition.

[0005] Chinese patent document CN115137773A discloses a chest imaging and respiration monitoring and measuring device, which places multiple antennas on the human chest, detects the internal tissues of the human body through electromagnetic waves, and when the human body performs breathing movement, the fluctuation of the human chest will affect the electromagnetic wave signal, and the abdomen of the human body also has obvious changes when breathing. Therefore, using the above-mentioned method, the changes of the abdomen cannot be collected in the collection of breathing data, and the electromagnetic wave will penetrate into the human body, which will make the user worry about the safety hazard, and the number of antennas used is large, and the cost of the measuring device is high. SUMMARY

[0006] To address the above problems, this invention proposes a human respiration recognition method based on millimeter-wave body area networks.

[0007] It uses the movement of the chest and abdominal cavities during various respiratory movements of the human body;

[0008] The collected motion states are labeled to obtain the training dataset;

[0009] Build a neural network model and train the neural network model using a training dataset;

[0010] The neural network model is trained by collecting the motion states of the chest cavity and abdominal cavity during the respiratory movement to be identified, and the neural network model identifies the type of breathing.

[0011] Furthermore, the motion states of the thoracic and abdominal cavities are collected by a pair of millimeter-wave directional antennas respectively placed on the chest and abdomen of the human body. This pair of millimeter-wave directional antennas forms a human body local area network. During the human body's breathing movements, the signal transmission of the human body local area network is affected. The signal transmission over a period of time is collected by a vector network analyzer as the motion state of the thoracic and abdominal cavities during the breathing movements.

[0012] Furthermore, millimeter-wave directional antennas are high-frequency directional antennas.

[0013] Furthermore, one of the pair of millimeter-wave directional antennas serves as the receiver and the other as the transmitter. The millimeter-wave directional antenna installed on the chest of the human body has its opening facing downwards, while the millimeter-wave directional antenna installed on the abdomen has its opening facing upwards. The openings of this pair of millimeter-wave directional antennas are opposite each other and on the same plane.

[0014] Furthermore, the vector network analyzer is set to continuous-time mode, radiating the signal along the chest-abdomen link via a directional antenna.

[0015] Furthermore, the vector network analyzer collects signals representing the power changes of the antenna signal.

[0016] Furthermore, the signals acquired by the vector network analyzer are preprocessed before being input into the neural network model. The preprocessing includes: randomly cropping the data using a fixed-length window and standardizing the cropped data to 0-1.

[0017] This invention fully utilizes the advantages of fixing a millimeter-wave antenna to the body, collecting respiratory data without relying on the reflection of electromagnetic waves. The collected respiratory data does not change with changes in the environment. At the same time, this invention can identify a variety of different breathing patterns. Attached Figure Description

[0018] Figure 1This is a flowchart of the steps of a human respiration recognition method based on millimeter-wave body area network according to the present invention;

[0019] Figure 2 These are amplitude and phase diagrams of five types of respiratory movements in a preferred embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the clipping window in a preferred embodiment of the present invention;

[0021] Figure 4 This is a schematic diagram of a one-dimensional convolutional neural network structure in a preferred embodiment of the present invention;

[0022] Figure 5 This is a schematic diagram of a 5-fold cross-data partitioning in a preferred embodiment of the present invention;

[0023] Figure 6 This is a 5-fold cross-validation accuracy graph in a preferred embodiment of the present invention;

[0024] Figure 7 This is a confusion matrix diagram of a test set in a preferred embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention proposes a method for human respiration recognition based on millimeter-wave body area networks.

[0027] It uses the movement of the chest and abdominal cavities during various respiratory movements of the human body;

[0028] The collected motion states are labeled to obtain the training dataset;

[0029] Build a neural network model and train the neural network model using a training dataset;

[0030] The neural network model is trained by collecting the motion states of the chest cavity and abdominal cavity during the respiratory movement to be identified, and the neural network model identifies the type of breathing.

[0031] Furthermore, the motion states of the thoracic and abdominal cavities are collected by a pair of millimeter-wave directional antennas respectively placed on the chest and abdomen of the human body. This pair of millimeter-wave directional antennas forms a human body local area network. During the human body's breathing movements, the signal transmission of the human body local area network is affected. The signal transmission over a period of time is collected by a vector network analyzer as the motion state of the thoracic and abdominal cavities during the breathing movements.

[0032] Furthermore, millimeter-wave directional antennas are high-frequency directional antennas.

[0033] Furthermore, one of the pair of millimeter-wave directional antennas serves as the receiver and the other as the transmitter. The millimeter-wave directional antenna installed on the chest of the human body has its opening facing downwards, while the millimeter-wave directional antenna installed on the abdomen has its opening facing upwards. The openings of this pair of millimeter-wave directional antennas are opposite each other and on the same plane.

[0034] In this embodiment, a pair of 28GHz horn antennas are selected as millimeter-wave directional antennas. Horn antennas have excellent propagation characteristics in a certain direction. The transmitting and receiving antennas are fixed to the chest and abdomen of the human body, respectively. The ends of the antennas are connected to a vector network analyzer via coaxial transmission lines. The vector network analyzer is set to continuous-time mode with a sampling frequency of 200Hz and a sampling duration of 20s. When the antenna emits electromagnetic waves, a local body area network is formed on the surface of the human body. Even slight movements of the human body will cause changes in the signal power on the body area network. Therefore, we record different forms of respiratory data of the human body based on these signal changes.

[0035] In this embodiment, five human breathing patterns were selected: breath-holding, light coughing, normal breathing, rapid breathing, and deep breathing. This means the collected data has five different labels. The amplitude and phase diagram of the collected respiratory data is shown below. Figure 2 As shown. Figure 2 The graphs (a)-|S21| to (e)-|S21| represent the amplitude information for five breathing types: breath-holding, light coughing, normal breathing, rapid breathing, and deep breathing, respectively. (a)-Phase to (e)-Phase represent the phase information of the motion corresponding to these five breathing types. As can be seen from the graphs, when holding one's breath, the signal amplitude changes very little, possibly due to slight body swaying or heartbeat. When a person coughs lightly, the signal shows rapid fluctuations within the period because a slight vibration occurs in the chest and abdomen, which interferes with the signal transmission path and alters the signal power. The difference between normal breathing, rapid breathing, and deep breathing lies in the different periods of the signal. Figure 2It can be seen that rapid breathing is characterized by a relatively short period on the amplitude graph; while deep breathing is characterized by a longer period, which is also reflected in the amplitude graph. The amplitude-phase graph shows that the differences between different breathing patterns are quite obvious and consistent with the body's response during actual testing. Eight people participated in this experiment, each with five sets of actions, each set repeated four times. Therefore, this embodiment obtained a total of 160 sets of 20-second data related to human respiration.

[0036] Because the collected data is too long to be easily fed into the network model for iterative training, this implementation uses a fixed-size time window to randomly prune the data several times, resulting in a discrete time series of fixed duration. The data pruning method is as follows: Figure 3 As shown. In this embodiment, we collected a total of 160 sets of data with a duration of 20 seconds. We selected a time window of 3 seconds and cropped the data 20 times. That is, we used a 3-second time window to randomly crop the data 20 times. After cropping, we obtained a total of 3200 data sets related to 5 different breathing modes.

[0037] The processed data is a discrete time series, similar to text data. One-dimensional convolutional neural networks (CNNs) are commonly used for feature extraction from text data. Therefore, this embodiment chooses to use a one-dimensional CNN to extract features from the discrete data, and then uses fully connected layers to classify the extracted features. In this embodiment, the network structure mainly consists of four one-dimensional convolutional layers and three fully connected layers. The loss function chosen is the cross-entropy loss function, the learning rate is 0.00008, and the number of iterations is 1000. The specific network structure is implemented using PyTorch, and a schematic diagram of the specific network structure is shown below. Figure 4 As shown.

[0038] The processed data is sent to the network for iterative training. In this embodiment, the training method is 5-fold cross-validation, where 80% of the data is used as the training set and 20% as the test set, and this is repeated 5 times for iterative training. The specific data partitioning method for 5-fold cross-validation is as follows: Figure 5 As shown.

[0039] The accuracy table after training with 5-fold cross-validation is as follows: Figure 6 As shown, by Figure 6 It can be seen that the model's accuracy ranges from a minimum of about 92% to a maximum of 96%, with an average accuracy of about 95%. The confusion matrix of a certain fold is as follows: Figure 7As shown in the figure, the vast majority of breathing patterns can be predicted, while a small portion of the breathing data is incorrectly predicted as other categories. After training with 5-fold cross-validation, it can be seen that after data collection via a high-frequency antenna and simple preprocessing, a one-dimensional convolutional neural network can classify five different breathing patterns with a high level of accuracy.

[0040] This embodiment also proposes a human respiration recognition system based on millimeter-wave body area networks. The system includes a data acquisition module, a data processing module, and a prediction module. The data acquisition module includes a pair of millimeter-wave directional antennas, a vector network analyzer, and a coaxial cable. The millimeter-wave directional antennas are used to transmit and receive electromagnetic waves, and the vector network analyzer is used to record some performance indicators of the antennas. The vector network analyzer is set to continuous-time mode with a sampling frequency of 200Hz and a sampling duration of 20s. The antennas and the vector network analyzer are connected using a coaxial cable. In this pair of millimeter-wave directional antennas, one acts as a transmitter and the other as a receiver. For example, in this embodiment, the antenna acting as the transmitter is installed on the chest of the human body, and the antenna acting as the receiver is installed on the abdomen of the human body. The openings of the two antennas are opposite each other, and the two antennas are in the same plane.

[0041] As one feasible approach, a communication device is integrated with a pair of millimeter-wave directional antennas and a vector network analyzer into a wearable device. The communication device transmits the signals collected by the vector network analyzer to a server for processing.

[0042] The data processing module and the prediction module are modules on the server. After the server receives the signal collected by the vector network analyzer, it inputs it into the data processing module for preprocessing. The preprocessing includes cropping the data to an appropriate size using a fixed-length time window and standardizing the cropped data to 0-1. For the training data, it is also necessary to label the processed data to obtain the dataset used in the training process of the prediction module.

[0043] The prediction module is a classification network used to predict data labels. During training, this network takes the dataset as input, predicts the labels of the dataset data, and updates the network parameters through backpropagation based on the cross-entropy between the predicted and actual labels. In the prediction network of this embodiment, a one-dimensional convolutional neural network is used to extract respiratory feature information from the input data, and the extracted respiratory feature information is input into a fully connected network for feature classification to obtain the prediction result.

[0044] During the training of the prediction network, the preprocessed data is a discrete time series, similar to a text sequence. The preprocessed data is divided into training and test sets according to a certain ratio. The training set is used to train the model and update the model's parameters in reverse. The test set is used to test the performance of the network model without updating the network's weights and biases. The training set data is fed into the constructed network model to train the model until it converges. The test set data that was not used in the training is used to verify the performance of the current model.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for human respiration recognition based on millimeter-wave body area networks, characterized in that, Specifically, the following steps are included: This study utilizes the movement of the chest and abdominal cavities during various respiratory movements of the human body. One of a pair of millimeter-wave directional antennas acts as the receiver, and the other as the transmitter. The antenna mounted on the chest faces downwards, while the antenna mounted on the abdomen faces upwards. The antennas are positioned opposite each other and on the same plane. This pair of antennas forms a human local area network (LAN). The human body's respiratory movements affect the signal transmission of this LAN. A vector network analyzer collects signal transmission data over a period of time to represent the movement of the chest and abdominal cavities during respiratory movements. The vector network analyzer is set to continuous-time mode, radiating the signal along the chest-abdomen link through the directional antennas. The signal collected by the vector network analyzer represents the power variation of the antenna signal. The collected motion states are labeled to obtain a training dataset; there are 5 types of labels, which correspond to 5 types of breathing: breath-holding, light cough, normal breathing, rapid breathing, and deep breathing. Build a neural network model and train the neural network model using a training dataset; The motion states of the chest cavity and abdominal cavity during the respiratory movement to be identified are collected as input to complete the training of the neural network model, and the neural network model identifies the type of breathing.

2. The human respiration recognition method based on millimeter-wave body area network according to claim 1, characterized in that, Millimeter-wave directional antennas are directional antennas operating at high frequencies.

3. The human respiration recognition method based on millimeter-wave body area network according to claim 1, characterized in that, The signals acquired by the vector network analyzer are preprocessed before being input into the neural network model. The preprocessing includes: randomly cropping the data using a fixed-length window and standardizing the cropped data to 0-1.

Images