Method for evaluating the number of breaths of a human body at a long distance based on a LoRa signal

Abstract

The present application belongs to the field of wireless sensing, and relates to a long-distance human respiratory frequency evaluation method based on LoRa radio frequency signals. The method comprises three steps: (1) data collection: covering the position where the human body is located through radio frequency signals, and collecting reflection signals of different individuals in different positions and orientations; (2) data processing and enhancement: eliminating random phase noise in the signals through ratio operation on the signals, carrying out denoising and smoothing processing using a filter, and realizing data enhancement through signal rotation, and finally generating a time-frequency graph using short-time Fourier transform; (3) construction of the relationship between signal change and respiratory frequency: based on the generated time-frequency graph, a visual model is used to establish a relationship model between signal change and respiratory frequency during detection. The present application can realize long-distance detection of the respiratory frequency of a user without interfering with the user, thereby providing an effective basis for intrusion detection, disaster rescue and the like.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of wireless sensing, and particularly relates to a long-distance human respiratory frequency evaluation method based on a LoRa radio frequency signal, which is mainly used to solve the problem that the existing wireless sensing technology has a small sensing range and is difficult to accurately measure the respiratory frequency of a target user in a long-distance and through-wall scene. BACKGROUND

[0002] Human vital sign detection has been widely applied to various scenarios such as building energy management systems, intruder detection, and disaster rescue. Accurate human vital sign detection provides a reliable basis for security and emergency rescue scenarios. There are currently various methods to achieve non-contact human vital sign detection, including vision, sensors, and radio frequency signals. Vision-based systems use a large number of cameras to identify human targets, but face problems such as light, line of sight, and privacy protection. Sensor-based methods require expensive equipment and regular maintenance, increasing the cost of use. In contrast, radio frequency signals can achieve non-line-of-sight human target detection, enabling high coverage in practice and achieving a good balance between accuracy and privacy protection. However, existing radio frequency signal-based methods, such as WIFI, RFID, Bluetooth, and acoustic technologies, have a key limitation: limited sensing distance, usually only within a few meters for wireless sensing. This makes them only suitable for small-range vital sign detection, making it difficult to meet the needs of certain practical application scenarios.

[0003] LoRa is a low-power wide-area network communication technology that uses Chirp Spread Spectrum (CSS) technology to significantly improve communication distance while maintaining low power consumption. In suburban environments, its transmission distance can reach several kilometers. In addition, LoRa devices have a low price and have taken on the function of data transmission in fields such as environmental monitoring, factory management, and intelligent agriculture. Compared to WIFI, RFID, and other signals, LoRa has less transmission loss and is more suitable for long-range sensing of human behavior. Therefore, using LoRa to achieve long-distance human vital sign detection has strong feasibility and high application value. SUMMARY

[0004] The present application addresses the shortcomings of the prior art and provides a long-distance human respiratory frequency evaluation method based on LoRa signals, which can achieve human target respiratory frequency detection in long-distance and through-wall scenarios.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] The long-distance human respiratory frequency evaluation method based on LoRa signals comprises the following steps:

[0007] 1) The data acquisition process is as follows:

[0008] The LoRa node cooperates with the directional antenna to emit a LoRa radio frequency signal S(t) to the area to be detected;

[0009] The target is allowed to sit or stand in a sector area with the vertical bisector of the two receiving antennas as the reference, and the area ranges from-35° to 35° in horizontal angle and 5 to 15 meters away from the front of the antenna;

[0010] The LoRa gateway is started, and the target reflection signals R1(t) and R2(t) are collected using two directional antennas, and the signals are down-sampled to reduce the subsequent data processing and enhance the time cost;

[0011] 2) The data processing and enhancement process is as follows:

[0012] The signals collected by the two directional antennas are subjected to ratio operation to obtain a signal ratio S(t) = R1(t) / R2(t), so as to eliminate random phase offsets in the received signals and ensure the availability of signal phase information;

[0013] The obtained signal ratio is applied to a low-pass filter to remove high-frequency noise caused by the environment and multipath;

[0014] The Savitzky Golay filter is used to smooth the signal ratio;

[0015] The signal ratio is rotated at a fixed step angle on the complex plane to obtain a signal ratio candidate set, and data enhancement is realized;

[0016] The short-time Fourier transform is used on the signal ratio candidate set to obtain the time-frequency graph corresponding to each signal ratio candidate, and the time-frequency graph set is combined.

[0017] 3) Relationship between signal change and breathing frequency:

[0018] The time-frequency graph set obtained through data processing and enhancement is used as the input layer variable of the visual model, the breathing frequency is used as the output layer variable of the model, and the true value of the target breathing frequency is used as the label;

[0019] The model input layer, output layer, adapter training parameters and network structure are set, the input layer variable data and the true value of the target breathing frequency are used to train the input layer, output layer and adapter parameters, so as to fine-tune the visual model;

[0020] The network model obtained through fine-tuning is used as the relationship model between signal change and breathing frequency.

[0021] To optimize the above technical solutions, the following specific measures are taken:

[0022] Further, the specific steps of the data collection are as follows:

[0023] S1: The LoRa node is matched with a directional antenna, and is configured to transmit signal S(t) = exp{j2πf c t+jπk 2} at 915 MHz frequency band with 125 kHz bandwidth continuously, where f c is the center frequency, k = B / T is the scanning rate of the linear frequency modulation pulse, B is the frequency bandwidth, and T is the linear frequency modulation pulse duration;

[0024] S2: The target is allowed to sit or stand in the range of -35°-35° in front of the LoRa transmitting antenna, at a distance of 5-15 m from the transmitting antenna;

[0025] S3: The LoRa gateway is constructed based on USRP B210 and GNU Radio, and the reflected signals of the target during the detection period are continuously received through two LoRa receiving antennas and where θ c is the carrier frequency offset caused by the clock inconsistency of the transmitting node and the receiver, θ s is the phase error caused by the sampling frequency offset, and τ n (t) is the propagation delay of the nth path;

[0026] S4: The signal is saved after down-sampling.

[0027] Further, the specific steps of the data processing and enhancement are as follows:

[0028] S1: From the vector angle, the received signal is composed of a static vector and a dynamic vector, where the static vector can be approximated as a single static vector in a limited time, and the dynamic vector can be regarded as a function of the moving distance of the target, so the received signal can also be expressed as and where d(t) is the length of the dynamic path caused by the target movement;

[0029] S2: Since the two antennas share the same clock, the two antennas have the same carrier frequency offset and sampling frequency offset, and the frequency offset can be eliminated by performing ratio operation on the received signals of the two antennas to obtain the signal ratio:

[0030]

[0031] where Δs is the dynamic path length difference caused by the position difference of two receiving antennas, which can be regarded as a constant in a short time compared with d(t). The signal ratio can be regarded as the Mobius transform of the signal received by one antenna, which does not change the shape and direction of the signal trajectory on the complex plane, but only changes the position and size of the trajectory. Therefore, the signal ratio still contains target motion information;

[0032] S3: According to the breathing characteristics of the human body, determine the frequency range, and apply a low-pass filter to the signal ratio to remove high-frequency noise caused by the environment and multipath effects;

[0033] S4: Apply a Savitzky Golay filter to further smooth the signal ratio;

[0034] S4: The signal is stored in complex form, and a single sampling point can be represented as: z=a+bi or Ae iφ , where a and b are the real and imaginary parts, respectively, and A and φ are the amplitude and phase. When the signal ratio is represented in real and imaginary parts, the real and imaginary parts are orthogonal to each other and exhibit complementary trends, that is, when the real part is not conducive to wireless sensing at a certain position, the imaginary part is conducive to sensing, and vice versa, so the real and imaginary parts can be combined for complementary sensing. In order not to be limited to the two dimensions of the real and imaginary parts of the signal, the signal ratio is data enhanced, and the signal ratio is rotated in the complex plane to generate multiple signal ratio candidates. The original signal ratio is stored as Z=(z1, z2,..., z n , and a signal ratio candidate is obtained by rotating the whole by θ:

[0035] Z'=(z'1, z'2,..., z' n )=(z1·e iθ , z2·e iθ ,..., z n ·e iθ )

[0036] θ is changed from 0 to 2π by a fixed step to generate different signal ratio candidates;

[0037] S5: In order to preserve time and frequency information, the short-time Fourier transform is used to extract the time-frequency features of the real and imaginary parts of the signal ratio candidates one by one, and the above features are presented in the form of a set of time-frequency graphs through visualization methods.

[0038] Further, the relationship between the signal change and the breathing rate is constructed as follows:

[0039] S1: The time-frequency graph set obtained by data processing and enhancement is used as the input layer variable of the fine-tuning visual model, the breathing rate is used as the output layer variable, and the true value of the breathing rate is used as the label required for model fine-tuning;

[0040] S2: add a low-parameter adapter to the visual model backbone network according to the use scene, and freeze the parameters of the backbone network except the adapter;

[0041] S3: count the breathing times of the measured target by itself as the true value, and fine-tune the set visual model by using the corresponding time-frequency graph set of the signal ratio and the true value of the breathing times;

[0042] S4: input the fine-tuned visual model as the relationship model between signal change and breathing times, and input the corresponding time-frequency graph set of the signal ratio to obtain the breathing times of the target during detection.

[0043] Further, in step S3, the model training cycle number is set to 200, the learning rate is set to 0.001, and if the root mean square error of the verification set does not decrease for 10 consecutive times in the model fine-tuning iteration process, the iteration is stopped.

[0044] The beneficial effects of the present application are: a long-distance human breathing times evaluation method based on LoRa signal is designed, LoRa signals are transmitted to the detection area, and two receiving antennas are used to continuously receive reflected signals, the carrier frequency offset and sampling frequency offset of the received signals of the two receiving antennas are eliminated by ratio operation, the signal ratio is denoised and smoothed by using a filter, and the signal ratio is rotated to obtain a signal ratio candidate with better sensing capability, then short-time Fourier transform and visualization method are used to generate a time-frequency graph set corresponding to the signal ratio candidate, finally, the time-frequency graph set is input, and the breathing times true value is labeled to fine-tune the visual model, thereby establishing a relationship model between signal change and breathing times, and realizing accurate evaluation of the target breathing times. The present application can continuously detect the breathing times of the user at a long distance without disturbing the user, and provides an effective basis for intrusion detection, disaster rescue and other scenes. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 Framework diagram of the long-distance human breathing times evaluation method based on LoRa signal

[0046] Figure 2 Adapter-based visual model parameter fine-tuning schematic diagram Specific implementation method

[0047] The present application will now be further described in detail in conjunction with the accompanying drawings.

[0048] As Figure 1The illustrated long-distance human respiratory frequency evaluation method based on LoRa signal, the main idea is: the method is divided into three steps: data collection, data processing and enhancement, and the relationship between signal change and respiratory frequency is constructed. The first two steps realize the collection of the target reflection signal in the detection area through LoRa node and LoRa gateway, and the signal data is preprocessed through signal division, low-pass filter, Savitzky Golay filter, while the signal rotation produces signal ratio candidates, and finally the visual method is used to realize the visual presentation of time-frequency characteristics. The third step is to establish the relationship model between signal change and respiratory frequency during detection by fine-tuning the visual model.

[0049] The specific steps of data collection are as follows:

[0050] 1) Place two receiving antennas and one transmitting antenna at the same vertical height (1.2 m) in turn, and all face the detection area, connect the LoRa node with the transmitting antenna, and use a 10000mAh mobile power supply to power the LoRa node;

[0051] 2) Let the target user sit or stand in the fan area with the vertical bisector of the two receiving antennas as the reference, the range of which is horizontal angle -35° to 35°, and the distance from the front of the antenna is 5 to 15 m;

[0052] 3) Connect the PC equipped with Inter Core i7-4790 CPU and 12GB memory to USRP B210 through USB interface, and connect the two RX interfaces of USRP B210 to the two receiving antennas respectively. At the same time, draw the data flow chart based on GNU Radio, and collect the reflection signal data of the detection area based on the flow chart, and save the signal data in the form of real part and imaginary part after downsampling;

[0053] The specific steps of data processing and enhancement are as follows:

[0054] 1) Because the transmitting node and the receiver are not synchronized, the received signal has phase shift caused by carrier frequency offset and sampling frequency offset, and the received signals of the two antennas are respectively represented as And From the vector angle, the two receiving antennas can be represented as And Where d(t) is the length of the dynamic path caused by target movement. By dividing the received signals of the two antennas, the phase shift in the signal can be eliminated, and the signal ratio obtained is:

[0055]

[0056] Compared with d(t), the Δs caused by the difference of the two receiving antenna positions can be regarded as a constant. The signal ratio can be regarded as the Mobius transform of the signal received by a single receiving antenna in an ideal case, which does not change the shape and direction of the original signal trajectory on the complex plane, but only changes the position and size of the signal trajectory, and the target motion information is still contained therein;

[0057] 2) Based on the upper limit of the normal human breathing frequency, the theoretical frequency maximum of the received signal ratio is calculated, and a low-pass filter is applied to the signal ratio to remove high-frequency noise caused by the environment and multipath effect;

[0058] 3) On the basis of applying low-pass filter, further use Savitzky Golay filter to smooth the signal ratio;

[0059] 4) The signal ratio is composed of multiple sampling points, each sampling point is stored in the form of complex value, and each sampling point in the signal ratio is rotated by an angle θ in turn, so as to realize the rotation of the signal ratio by θ, and the result after rotation is a signal ratio candidate. By changing θ from 0 to 2π with a fixed step, multiple signal ratio candidates are generated;

[0060] 5) Use short-time Fourier transform to extract time-frequency features of the real part and the imaginary part of the signal ratio candidate one by one, to present time and frequency information at the same time. At the same time, the time-frequency features are saved as a set of time-frequency diagrams through visualization technology.

[0061] The specific steps of building the relationship between signal change and breathing frequency are as follows:

[0062] 1) Build a deep learning Linux development environment, based on the deep learning framework Pytorch, adjust the input layer and output layer of the visual model to adapt to the expected input and output. At the same time, according to the structure shown in the figure, add a low-parameter adapter inside the visual model; Figure 2

[0063] 2) During the test, the target being tested records the breathing frequency by itself as the true value of the breathing frequency. The true value and the time-frequency diagram set obtained through data processing and enhancement together constitute a data sample, wherein the former is the label and the latter is the input.

[0064] 3) Divide the data set into training data set, test data set and validation data set, the proportion is: 7:2:1:

[0065] ​4) Using the time-frequency map set corresponding to the signal ratio candidate and the true value of the respiratory rate, the visual model is fine-tuned. At the same time, the parameters of the input layer, the adapter and the output layer are updated according to the Adam optimization algorithm of Pytorch. The number of iterations is 200, the learning rate is set to 0.001, and the root mean square error (RMSE) of the validation set does not decrease for 10 consecutive times in the process of model fine-tuning iteration, then the iteration is stopped;

[0066] 5) The fine-tuned visual model is used as the relationship model between signal change and respiratory rate, and the time-frequency map set corresponding to the signal ratio obtained from the target reflection signal can obtain the target respiratory rate during detection.

[0067] It should be noted that the terms such as "up", "down", "left", "right", "front", "back" and the like cited in the invention are only for the convenience of clear description, and are not used to limit the scope of the implementation of the invention. The change or adjustment of the relative relationship is also considered as the scope of the implementation of the invention without substantial change of the technical content.

[0068] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the scope of the present application is within the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application are considered as the protection scope of the present application.

Claims

1. A method for long-distance human respiratory rate assessment based on LoRa signals, characterized in that, Includes the following steps: 1) LoRa signal preprocessing: S1.1: The ratio of the LoRa signals collected by the two receiving antennas is calculated to obtain the signal ratio. The carrier frequency offset and sampling frequency offset are canceled by the shared clock of the two receiving channels to restore the phase availability. S1.2: Determine the upper limit of the signal ratio based on the theoretical human breathing frequency band, apply a low-pass filter to the signal ratio to suppress high-frequency noise caused by the environment and multipath, and further smooth it using a Savitzky-Golay filter to obtain the preprocessed signal ratio for subsequent processing. 2) Complex domain phase compensation and candidate time-frequency map set construction: S2.1: A phase compensation search strategy with a fixed step size of 15° is adopted; S2.2: Values ​​are taken at equal intervals within the range of 0-360°, and the number of phase compensation candidates is 24; S2.3: Perform short-time Fourier transform on the real and imaginary parts of each candidate signal ratio, using a Hann window with a window length of 10s, a step size of 1s, and an overlap rate of 90%; and combine the time-frequency graphs corresponding to the real and imaginary parts to form a candidate time-frequency graph set for subsequent models. 3) Model building and fine-tuning: S3.1: During training, simultaneously minimize the regression loss of breathing count and the branch consistency regularization term, wherein the consistency regularization adopts KL divergence and its loss weight is 0.30; S3.2: Model fine-tuning uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 200 training epochs; the parameters of the visual model backbone network are frozen, and low-parameter adapters are added only to the bottleneck layer of the backbone, and the input layer, output layer, and the adapters are trained to complete small-sample fine-tuning. S3.3: Terminate training early if the root mean square error of the validation set does not decrease after 10 consecutive iterations; 4) Estimate respiratory rate: The candidate time-frequency map set obtained during the detection period is input into the fine-tuned visual model to obtain continuous real-valued output within the detection window. Post-processing is performed and peak detection is adopted. A constraint of minimum peak interval of 1.2s is applied to discretize the continuous output of the model into integer breath counts.

2. The respiratory rate estimation method as described in claim 1, characterized in that, Post-processing is performed on the continuous real-valued output of the model. Peak detection is used and a constraint of minimum peak interval of 1.2s is applied to discretize the output into integer counts.

Images