TWI845871B - Data pre-processing method and exercise vital signs radar - Google Patents

Abstract

A data pre-processing method is provided. A power distribution parameter set obtained by beamforming scanning can increase the signal-to-noise ratio of the signal and automatically generate a bounding box so as to improve the tracking effect of the object.

Description

Data pre-processing method and sports physiological sensing radar

The present invention relates to radar signal processing technology, in particular to a radar signal data pre-processing method and a sports physiological sensing radar using the method for detection.

There are many wearable or direct contact physiological parameter measurement devices that can monitor physiological parameters (such as heart rate) in daily life activities. However, wearing wearable or contact devices for a long time will make the subjects feel uncomfortable. Although there are still non-contact measurement methods, when the subject is in motion, the shaking of his body is likely to interfere with the measurement and affect the measurement accuracy.

In view of this, according to some embodiments, a data pre-processing method is performed by a processor in a signal processing device, including: obtaining an energy distribution parameter set and a digital signal obtained by beamforming scanning, the digital signal corresponding to a reflected radar signal of a motion physiological sensing radar; using the energy distribution parameter set to search for a target by filtering background noise; according to the energy distribution parameter set, weighting the digital signal to obtain an optimized signal; analyzing the optimized signal to extract one or more target phase data corresponding to the target from the optimized signal; and inputting the one or more target phase data into a machine learning model to obtain a physiological parameter prediction result.

According to some embodiments, a motion physiological sensing radar includes: a transmitting unit, a receiving unit and a signal processing module. The transmitting unit transmits an incident radar signal. The receiving unit receives a reflected radar signal. The signal processing module controls the transmitting unit and the receiving unit to perform beamforming scanning to obtain an energy distribution parameter set, and obtains a corresponding digital signal according to the reflected radar signal, and uses the energy distribution parameter set to search for a target by filtering background noise, and weights the digital signal according to the energy distribution parameter set to obtain an optimized signal, and analyzes the optimized signal to extract one or more target phase data corresponding to the target from the optimized signal, and inputs the one or more target phase data into a machine learning model to obtain a physiological parameter prediction result.

According to some embodiments, a data pre-processing method is performed by a processor in a signal processing device, including: obtaining an energy distribution parameter set and a digital signal obtained by beamforming scanning, wherein the digital signal corresponds to a reflected radar signal of a motion physiological sensing radar; using the energy distribution parameter set to search for a target by filtering background noise; analyzing the digital signal to extract one or more target phase data corresponding to the target from the digital signal; dividing the one or more target phase data into a plurality of sub-bands by wavelet transform; performing statistical analysis on each sub-band to obtain a statistical feature set; and inputting the statistical feature set into a machine learning model to obtain a physiological parameter prediction result.

According to some embodiments, a set of energy distribution parameters within a period is statistically calculated to determine a detection distance area covering the target activity range, and then the optimized signal located in the detection distance area is analyzed.

According to some embodiments, before the target phase data is input into the machine learning model, the target phase data is further subjected to signal processing including phase difference calculation and pulse noise removal.

According to some embodiments, a phase map and a vibration frequency map are obtained according to the optimization signal; at least one candidate position having an energy intensity exceeding an energy threshold is selected from the vibration frequency map; and a target position is selected from the candidate positions, the target position being one of the candidate positions having a vibration frequency that meets a physiological parameter range and has the largest energy intensity; and then one or more target phase data within a distance range in the phase map are obtained according to the target position. The phase map presents energy distribution with distance changes and phase changes relative to the motion physiological sensing radar, and the vibration frequency map presents energy distribution with distance changes and vibration frequency changes relative to the motion physiological sensing radar.

According to some embodiments, a set of energy distribution parameters within a period is statistically calculated to determine a detection distance area covering the target activity range, wherein the step of selecting a candidate position is to select the detection distance area from the vibration frequency map.

According to some embodiments, the optimized signal is fast Fourier transformed to obtain a distance profile map; each distance delay change of the distance profile map is subjected to DC bias elimination, IQ imbalance compensation, inverse tangent and phase unfolding to obtain a phase map; and the phase distribution at each distance of the phase map is fast Fourier transformed to obtain a vibration frequency map. The distance profile map presents energy distribution with distance changes and time changes relative to the motion physiological sensing radar.

According to some embodiments, an energy threshold is calculated for each distance bar in the phase map, and the energy threshold is determined based on an energy average or an energy maximum of the corresponding distance bar. Furthermore, the energy value of each phase on each distance bar is compared with the energy threshold of the corresponding distance bar to select a candidate position that exceeds the energy threshold.

According to some embodiments, a phase map and a vibration frequency map are obtained according to the optimization signal, and at least one candidate position with energy intensity exceeding an energy threshold is selected from the vibration frequency map; and N target positions are selected from the candidate positions, and N is greater than 1, wherein the N target positions to be measured are those having a vibration frequency that meets a physiological parameter range and having the top N energy intensities among the candidate positions; and then according to each target position, one or more target phase data corresponding to a distance range in the phase map are obtained. The phase map presents energy distribution with distance changes relative to the motion physiological sensing radar and phase changes, and the vibration frequency map presents energy distribution with distance changes relative to the motion physiological sensing radar and vibration frequency changes.

In summary, according to the data pre-processing method and sports physiological sensing radar of some embodiments, the physiological parameters of the subject in motion can be accurately sensed and the intensity of the exercise can be detected. In some embodiments, the signal-to-noise ratio can be improved by weighting the digital signal. In some embodiments, the computational complexity can be reduced and the object tracking effect can be improved by automatically generating the detection distance area. In some embodiments, the noise interference can be reduced by reducing the noise through signal processing. In some embodiments, the model training and prediction speed can be accelerated by using statistical feature sets for machine learning prediction.

As used herein, the term “connected” means that two or more elements are in direct physical or electrical contact with each other, or are in indirect physical or electrical contact with each other.

Referring to FIG. 1 , it is a schematic diagram of the use status of the sports physiological sensing radar 10 according to some embodiments. The sports physiological sensing radar 10 transmits a radar signal (hereinafter referred to as "incident radar signal FH"). The incident radar signal FH is transmitted to the target 90 and is modulated by the movement of the target 90 (such as the subject) and reflected back to the sports physiological sensing radar 10. The reflected radar signal is hereinafter referred to as "reflected radar signal FN". Therefore, one or more information of the target 90 can be detected by analyzing the reflected radar signal FN. The information can be, for example, speed, distance, direction, physiological information (such as heartbeat, breathing), etc.

In some embodiments, the sports physiological sensing radar 10 may be a frequency modulated continuous wave (FMCW) radar, a continuous wave (CW) radar, or an ultra-wideband (UWB) radar. The following description will be made using the frequency modulated continuous wave radar as an example.

Referring to FIG. 2 , FIG. 2 is a schematic diagram of an example radar signal, wherein the upper half shows the change of the amplitude of the incident radar signal FH over time, and the lower half shows the change of the frequency of the incident radar signal FH over time. The incident radar signal FH includes a complex chirp signal SC. For the sake of clarity, FIG. 2 only shows one chirp pulse SC. Here, the chirp pulse SC is a linear frequency modulated pulse signal, which refers to a sine wave whose frequency increases linearly with time. In some embodiments, the frequency of the chirp pulse SC increases in a nonlinear manner. For the sake of convenience, the following description is given in a linear manner. As shown in FIG2 , within a duration Tc (e.g., 40 microseconds), the chirped pulse SC increases linearly from a starting frequency (e.g., 77 GHz) to an ending frequency (e.g., 81 GHz) according to a slope S. The starting frequency and the ending frequency can be selected from the millimeter wave band (i.e., 30 GHz to 300 GHz). The difference between the starting frequency and the ending frequency is the pulse bandwidth B.

Please refer to FIG. 3 and FIG. 4. FIG. 3 is a block diagram of a frequency modulated continuous wave radar 10' according to some embodiments. FIG. 4 is a schematic diagram illustrating an incident radar signal FH and a reflected radar signal FN. The frequency modulated continuous wave radar 10' includes a transmitting unit 11, a receiving unit 12, a demodulating unit 13, an analog-to-digital converter 14, and a processing unit 15. The transmitting unit 11 is used to transmit the incident radar signal FH, and includes a transmitting antenna and a signal synthesizer. The signal synthesizer is used to generate the incident radar signal FH including a chirp pulse Ct, and transmit it via the transmitting antenna. The receiving unit 12 includes a receiving antenna, which is used to receive the reflected radar signal FN including at least one chirp pulse Cr. The chirped pulse Cr can be regarded as a delayed version of the chirped pulse Ct. The demodulation unit 13, the analog-to-digital converter 14 and the processing unit 15 are used to process the received reflected radar signal FN, and can be collectively referred to as a signal processing module 16. The demodulation unit 13 is connected to the transmitting unit 11 and the receiving unit 12, and includes a mixer and a low-pass filter. The mixer couples the chirped pulse Ct of the incident radar signal FH and the chirped pulse Cr corresponding to the reflected radar signal FN, and can generate two coupled signals, namely, the sum of the frequencies of the two chirped pulses Ct and Cr and the difference in frequency. The low-pass filter performs low-pass filtering on the coupled signal to remove high-frequency components, and obtains a coupled signal of the frequency difference between the two chirped pulses Ct and Cr, which is hereinafter referred to as an "intermediate frequency (IF) signal SI". The analog-to-digital converter 14 converts the intermediate frequency signal SI into a digital signal. The processing unit 15 performs digital signal processing on the digital signal. The processing unit 15 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), or other programmable general-purpose or special-purpose microprocessors, digital signal processors (DSP), programmable controllers, application specific integrated circuits (ASIC), programmable logic devices (PLD), or other similar devices, chips, integrated circuits, and combinations thereof.

Referring to Figure 4, the frequency of the intermediate frequency signal SI It can be expressed as formula 1, S is the slope, τ is the delay time between sending the incident radar signal FH and receiving the reflected radar signal FN. Therefore, τ can be expressed as formula 2, d is the distance from the radar transmitting antenna to the target 90, and c is the speed of light. Substituting formula 2 into formula 1, we can get formula 3. From formula 3, we can know that the frequency of the intermediate frequency signal SI is Implicitly contains distance information (i.e. the distance between the FMR 10' and the target 90).

…Formula 1

...Formula 2

...Formula 3

Refer to FIG5 , which is a schematic diagram of signal processing according to some embodiments. Here, the chirp pulses SC are numbered C1, C2, C3, ..., Cn in sequence, where n is a positive integer. The analog-to-digital converter 14 converts the received intermediate frequency signals SI corresponding to the chirp pulses C1 to Cn into digital signals SD (represented as number sequences D1, D2 ..., Dn, where n is a positive integer), and each chirp pulse Cx (x=1~n) has a corresponding number sequence Dx (x=1~n). The number sequence Dx (x=1~n) of the digital signal SD can be represented as a one-dimensional array (row matrix). These horizontal arrays Dx (x=1~n) are arranged vertically in sequence to form a two-dimensional matrix A1. It is understandable that the digital signal SD can also be arranged into a column array, and these column arrays are arranged horizontally in sequence to obtain another two-dimensional matrix. The values of the two-dimensional matrix A1 represent the signal strength (amplitude). The index value x of the column of the two-dimensional matrix A1 corresponds to the order of the chirp pulse SC. The index value of the horizontal column of the two-dimensional matrix A1 has the meaning of time, that is, the horizontal array of the two-dimensional matrix A1 is a time domain signal (a set of digital data related to time).

The processing unit 15 performs a fast Fourier transform (FFT) (hereinafter referred to as "distance Fourier transform") on each row matrix of the two-dimensional matrix A1 (i.e., the two-dimensional matrix A1 formed by the digital signal SD) to obtain a frequency domain signal SP (respectively represented as P1, P2..., Pn, n is a positive integer), i.e., the two-dimensional matrix A2. Therefore, the row matrix of the two-dimensional matrix A2 is equivalent to the spectrum distribution corresponding to a chirped pulse Cx. As mentioned above, the frequency of the intermediate frequency signal SI implicitly contains distance information. That is, the index value of the row of the two-dimensional matrix A2 has a distance meaning. The values of the two-dimensional matrix A2 represent the intensity of each frequency on the spectrum, and can present the intensity of the radar signal reflected at different distances from the FMCW radar 10'. As shown in FIG5, the filled box in the two-dimensional matrix A2 is the peak value (i.e., the value exceeds a threshold value), indicating that there is a target 90 at the distance corresponding to this frequency. The distance between the FMCW radar 10' and the target 90 can be calculated from the frequency at the peak. Furthermore, according to the distance change of the specific target 90 calculated at different time points, large-scale motion information (such as average speed) can be calculated.

Although the above description is based on an example that the transmitting unit 11 has one transmitting antenna and the receiving unit 12 has one receiving antenna, the transmitting unit 11 has multiple transmitting antennas to transmit multiple incident radar signals FH, and the receiving unit 12 has multiple receiving antennas to receive reflected radar signals FN respectively to perform beamforming.

Please refer to Figures 6 and 7 together. Figure 6 is a flow chart of a radar signal data preprocessing method according to some embodiments, which illustrates a data preprocessing process that can be used to predict physiological parameters using a machine learning model. Figure 7 is a block diagram of a signal processing device 60 according to some embodiments. The signal processing device 60 includes a processor 61 and a storage device 62. The storage device 62 is a computer-readable storage medium for storing a program 63 executed by the processor 61 to execute the data preprocessing method. In some embodiments, the signal processing device 60 is the aforementioned frequency modulated continuous wave radar 10', and the processor 61 is the aforementioned processing unit 15. In some embodiments, the signal processing device 60 is an edge device or a cloud server, that is, after the frequency modulated continuous wave radar 10' obtains the digital signal SD, the digital signal SD will be transmitted to the edge device or the cloud server, and the edge device or the cloud server will perform digital signal processing.

In step S200, as described above, the analog-to-digital converter 14 can convert the received intermediate frequency signal SI corresponding to each chirp pulse Cx into a digital signal SD, so that the processor 61 can obtain the digital signal SD corresponding to the reflected radar signal FN. In addition, after receiving the digital signal SD, the frequency modulated continuous wave radar 10' scans the field in a beamforming manner to calculate the signal strength at different distances and azimuths to obtain an energy distribution parameter set. The energy distribution parameter set includes parameters such as angle, distance, power, etc., based on which a two-dimensional spatial spectrum signal strength map can be established. As shown in Figure 8, it is a two-dimensional spatial spectrum signal strength map according to some embodiments. The horizontal axis is distance, the vertical axis is angle, and the power (energy intensity) is presented here in shades of color. In addition to using fast Fourier, beamforming algorithms can also use other self-adjusting beamforming methods, such as multi-signal classification (MUSIC), Capon, ESPRIT, CBF algorithms, etc.

In step S202, the energy distribution parameter set is used to filter the background noise to search whether the target 90 exists in the field. The background noise filtering method may be, for example, a constant false alarm rate (CFAR) filtering method. If a peak is found through this calculation (as shown by the dashed box in FIG8 ), it means that the target 90 exists.

In step S204, the digital signal SD is weighted according to the energy distribution parameter set to obtain an optimized signal, as shown in Formula 4. Y _k is the optimized signal, Xs is the digital signal, w _k ( ) is based on the distance r and angle The weight calculated by the parameter. The weighted weight is calculated by combining the distance r and angle of the energy distribution parameter. The parameters are substituted into the Capon Beamforming weight formula to calculate. In this way, the signal in a specific area (i.e. the area close to the target 90) can be optimized to improve the signal-to-noise ratio.

...Formula 4

In step S206, the optimized signal can be used for analysis to extract the target phase data corresponding to the target 90. After obtaining the target phase data, it can be used to input into the machine learning model 64 to predict physiological parameters (step S208). For example, predict the corresponding breathing rate or heart rate. In some embodiments, the target phase data is normalized and then input into the machine learning model 64. In one embodiment, the machine learning model 64 adopts the MobileNetV3 model. The usage sample is to collect usage data of two kinds of sports equipment (bicycles and elliptical machines). 30 person-times of radar data are collected for each type of sports equipment, for a total of 60 person-times of radar data, of which 50 are used for training and 10 are used for prediction. Each radar data includes data of four exercise intensities (resting, slow, medium, and fast), and each exercise intensity lasts two minutes. The frequency modulated continuous wave radar 10' is set up at a height of 1~2.5 meters and 0.5~1.5 meters away from the subject, but not limited to this. During the collection process, the subject wears a heart rate monitor to synchronously obtain the real-time heart rate as a marked sample. Next, we will explain how to analyze and optimize the signal to obtain the target phase data.

Referring to FIG9 , a signal analysis flow chart according to some embodiments is shown. First, in step S701, the optimized signal is subjected to the aforementioned range Fourier transform to obtain a range profile map (step S702). As shown in FIG10 , a schematic diagram of a range profile map according to some embodiments is shown. The range profile map presents the energy distribution of the range change (horizontal axis) and the time change (vertical axis) of the frequency modulated continuous wave radar 10 'relative to the frequency modulated continuous wave radar 10 ', where the energy difference is presented by the depth of the color.

In addition to obtaining a distance profile map, a phase map and a vibration frequency map can be further obtained based on the optimized signal. In step S703, the distance profile map is subjected to DC removal, ellipse correction, arctangent and phase unwrapping to obtain a phase map (step S704). As shown in FIG11 , it is a schematic diagram of a phase map according to some embodiments. The phase map presents the energy distribution of the distance change (horizontal axis) and the phase change (vertical axis) of the frequency modulated continuous wave radar 10 'relative to the frequency modulation, and the energy difference is presented here by the depth of the color. Next, in step S705, the phase distribution at each distance of the phase map (i.e., range bin) is fast Fourier transformed to obtain a vibration frequency map (step S706). As shown in FIG12, it is a schematic diagram of a vibration frequency map according to some embodiments. The vibration frequency map presents the energy distribution of the distance change (horizontal axis) and the vibration frequency change (vertical axis) of the frequency modulated continuous wave radar 10' relative to the frequency modulation, and the energy difference is presented here by the depth of color.

After the vibration frequency map is obtained, in step S707, at least one candidate position having an energy intensity exceeding an energy threshold is selected from the vibration frequency map (step S708). As shown in FIG. 13, it is a schematic diagram of the vibration frequency distribution of the distance bar block according to some embodiments. FIG. 13 shows a peak exceeding the energy threshold _Vth , so the distance bar block is selected as the candidate position. In other words, step S707 compares each distance bar block in the phase map with the energy threshold _Vth , and if it exceeds the energy threshold _Vth , the corresponding distance bar block is selected as the candidate position.

In some embodiments, the energy threshold V _th is a floating threshold. A respective energy threshold V _th is calculated for each distance strip in the phase map. The energy threshold V _th is determined based on the energy average or energy maximum of the corresponding distance strip. For example, the energy threshold V _th is the sum of a times the energy average and b times the energy maximum, where a+b=1, and a and b are positive numbers. For another example, the energy threshold V _th is a times the energy average, where a is a positive number.

The candidate positions obtained in the aforementioned step S708 may be multiple, so it is necessary to further determine which one should be selected to eliminate the interference signal. In step S709, one or more of the candidate positions are selected to obtain one or more target positions (step S710). The target position is one of the candidate positions that has a vibration frequency that meets a physiological parameter range. The physiological parameter range can be, for example, a breathing frequency range (e.g., 10 to 20 beats per minute), a heart rate range (e.g., 60 to 100 beats per minute), etc.

Specifically, in some embodiments, there is a target 90 in the detection field. Find each candidate position with a vibration frequency that meets the physiological parameter range, and select the one with the largest vibration frequency range energy. The selected candidate position (distance) is the location of the target 90 (i.e., the target position).

In some embodiments, there are multiple targets 90 in the detection field. N target positions are selected from the candidate positions, and N is greater than 1, wherein the N target positions are those with vibration frequencies that meet the physiological parameter range and have the top N energy intensities among the candidate positions. These target positions are the locations where the targets 90 are located (i.e., target positions).

After determining the location of one or more targets, the corresponding one or more target phase data can be retrieved (step S711). Considering that object detection in motion may result in misjudgment and deviation. In step S711, according to each target location, a target phase data within a corresponding distance range in the phase map is obtained (step S712). In some embodiments, according to each target location, a target phase data within a corresponding distance range in the phase map is obtained, wherein the target phase data includes a distance bar block of the target location. In other embodiments, based on each target position, multiple target phase data within a corresponding distance range in the phase map are obtained, wherein these target phase data include not only the distance bar block of the target position, but also one or more distance bar blocks adjacent to the target position. For example, taking the distance bar block of the target position as the center and taking two distance bar blocks on both sides, the target phase data includes five distance bar blocks.

The content of the aforementioned step S208 is explained here. In step S208, the target phase data of each target position to be measured is input into the machine learning model 64 to obtain the physiological parameter prediction result. For example, the corresponding breathing frequency or heart rate is predicted. In some embodiments, the target phase data is normalized and then input into the machine learning model 64.

Referring to FIG. 14 , it is a flow chart of another data pre-processing method according to some embodiments. Compared with FIG. 6 , step S205 is further included before step S206. In step S205, a set of energy distribution parameters within a period (e.g., 10 to 20 seconds) can be continuously collected and counted to analyze the activity state of the target 90, and a detection distance area (bounding box) covering the activity range of the target 90 is determined accordingly, such as the dotted frame shown in FIG. 8 . Accordingly, in step S206, only the optimized signal of the detection distance area (within a specific range) can be analyzed. That is, in the aforementioned step S707, only the distance bar within the detection distance area needs to be monitored, and the candidate position is selected from the detection distance area in the vibration frequency map. In this way, the amount of calculation and the calculation time can be saved. The detection distance area can be updated regularly (for example, 30 seconds). The update cycle can be dynamically adjusted according to the analyzed disturbance rate of the target 90 in the detection distance area. For example, when the target 90 moves violently, the update cycle can be shortened; conversely, when the target 90 moves and slows down, the update cycle can be extended to reduce the calculation complexity. In other embodiments of the present disclosure, a plurality of detection distance areas may be selected to meet the needs of multi-target detection.

In some embodiments, as shown in FIG. 14 , compared to FIG. 6 , before executing step S208, the target phase data is further subjected to signal processing (step S207). Referring to FIG. 15 , it is a schematic diagram of signal processing according to some embodiments. The upper diagram of FIG. 15 is a schematic diagram of a distance bar block, and after the phase difference is calculated, the middle diagram of FIG. 15 is presented. Phase difference calculation refers to subtracting the previous and subsequent values of two adjacent phases. Then, after removing the pulse noise, the lower diagram of FIG. 15 is presented. The specific method may be, for example, that when the phase difference is too large and exceeds a preset threshold, it is replaced by 0. In this way, noise interference can be reduced.

Referring to FIG. 16 , it is a schematic diagram of the physiological parameter prediction results of executing the process shown in FIG. 14 . The accuracy is 90.49%, the root mean square error is 12.72 (times/minute, bpm), and the standard error is 7.43 (times/minute, bpm). It can be seen that the predicted heart rate value change is consistent with the actual heart rate change, which can effectively judge the intensity of exercise. As a comparison, if the optimization signal is not used and the digital signal SD is used to execute the process shown in FIG. 9 and input it into the machine learning model 64 (the signal processing of step S207 is not executed), the accuracy is 86.88%, the root mean square error is 20.04 (times/minute, bpm), and the standard error is 14.69 (times/minute, bpm). It can be seen that the prediction results of some embodiments of the present disclosure are improved by about 4% accuracy. Referring to FIG. 17 , it is a Bland-Altman diagram based on some embodiments to compare the difference in prediction results of these two methods.

Referring to FIG. 18 , it is a flow chart of another data pre-processing method according to some embodiments. Compared with FIG. 14 , steps S300 to S307 are substantially the same as the aforementioned steps S200 to S207, except that the present embodiment does not directly input the target phase data into the machine learning model 64. In step S308 , each distance strip block in the target phase data is divided into a plurality of sub-bands by wavelet transformation, and statistical analysis is performed on each sub-band to obtain a statistical feature set. For example, for each sub-band, there are 14 statistical features: entropy, skewness, kurtosis, variance, standard deviation, mean, median, 5th percentile value, 25th percentile value, 75th percentile value, 95th percentile value, root mean square value, zero crossing rate, and mean crossing rate. If a fifth-order wavelet decomposition is performed to obtain a total of 6 sub-bands, the amount of data for a distance strip block can be reduced from 500 to 84 feature parameters. In this way, the computational burden can be reduced.

In step S309, the statistical feature set is input into the machine learning model to obtain a physiological parameter prediction result. Referring to FIG. 19, it is a schematic diagram of the physiological parameter prediction result of executing the process shown in FIG. 18. The accuracy is 88.76%, the root mean square error is 15.02 (times/minute, bpm), and the standard error is 8.58 (times/minute, bpm). Referring to FIG. 20, it is a Bland-Altman diagram according to some embodiments, which is used to compare the results of executing the processes shown in FIG. 15 and FIG. 18. It can be seen that although the accuracy is slightly worse, the prediction performance is not much different.

The non-contact motion physiological sensing method is to obtain and process the digital signal SD in a sliding window manner. In some embodiments, the window size is 10 seconds and the time steps are 1 second.

In summary, according to the radar signal data pre-processing method and the sports physiological sensing radar 10 of some embodiments, the physiological parameters of the subject in motion can be accurately sensed and the intensity of the exercise can be detected. In some embodiments, the signal-to-noise ratio can be improved by weighting the digital signal. In some embodiments, the calculation complexity can be reduced and the object tracking effect can be improved by automatically generating a detection distance area. In other embodiments, multiple detection distance areas are automatically generated to meet the needs of multi-target detection. In some embodiments, noise interference can be reduced by reducing noise through signal processing. In some embodiments, model training and prediction speed can be accelerated by using statistical feature sets for machine learning prediction.

10: Sports physiological sensing radar 10': Frequency modulation continuous wave radar 11: Transmitter 12: Receiver 13: Demodulator 14: Analog digital converter 15: Processor 16: Signal processing module 60: Signal processing device 61: Processor 62: Storage device 63: Program 64: Machine learning model 90: Target FH: Incident radar signal FN: Reflected radar signal A1, A2: Two-dimensional matrix Array B: pulse bandwidth Ct, Cr: chirp pulse C1, C2, C3, Cn: chirp pulse D1, D2, Dn: sequence P1, P2, Pn: frequency domain signal S200~S208: steps S300~S309: steps S701~S712: step S: slope SC: chirp pulse SD: digital signal SI: intermediate frequency signal SP: frequency domain signal Tc: duration τ: delay time V _th : energy threshold

[Figure 1] is a schematic diagram of the use status of a sports physiological sensing radar according to some embodiments. [Figure 2] is a schematic diagram of an example radar signal. [Figure 3] is a block diagram of a frequency modulated continuous wave radar according to some embodiments. [Figure 4] is a schematic diagram of an incident radar signal and a reflected radar signal. [Figure 5] is a schematic diagram of signal processing according to some embodiments. [Figure 6] is a flow chart of a data pre-processing method according to some embodiments. [Figure 7] is a block diagram of a signal processing device according to some embodiments. [Figure 8] is a two-dimensional spatial spectrum signal intensity diagram according to some embodiments. [Figure 9] is a signal analysis flow chart according to some embodiments. [Figure 10] is a schematic diagram of a distance topography map according to some embodiments. [Figure 11] is a schematic diagram of a phase map according to some embodiments. [Figure 12] is a schematic diagram of a vibration frequency map according to some embodiments. [Figure 13] is a schematic diagram of the vibration frequency distribution of a distance strip block according to some embodiments. [Figure 14] is a flow chart of another data pre-processing method according to some embodiments. [Figure 15] is a schematic diagram of signal processing according to some embodiments. [Figure 16] is a schematic diagram of the physiological parameter prediction results of executing the process shown in Figure 14. [Figure 17] is a Bland-Altman diagram according to some embodiments. [Figure 18] is a flow chart of another data pre-processing method according to some embodiments. [Figure 19] is a schematic diagram of the physiological parameter prediction results of executing the process shown in Figure 18. [Figure 20] is a Bland-Altman plot according to some embodiments.

S200~S208: Steps

Claims (21)

A data pre-processing method is performed by a processor in a signal processing device, comprising: obtaining an energy distribution parameter set and a digital signal obtained by beamforming scanning, wherein the digital signal is obtained by demodulating a reflected radar signal of a motion physiological sensing radar into an intermediate frequency signal and performing analog-to-digital conversion; using the energy distribution parameter set to search for a target by filtering background noise; weighting the digital signal according to the energy distribution parameter set to optimize an area adjacent to the target to obtain an optimized signal; analyzing the optimized signal to extract one or more target phase data corresponding to the target from the optimized signal; and inputting the one or more target phase data into a machine learning model to obtain a physiological parameter prediction result. The data pre-processing method as described in claim 1 further includes: statistically calculating the energy distribution parameter set within a period to determine a detection distance area covering the target activity range, wherein the step of analyzing the optimized signal is to analyze the optimized signal located in the detection distance area. The data pre-processing method as described in claim 1, wherein before the one or more target phase data are input into the machine learning model, the one or more target phase data are also subjected to a signal processing, and the signal processing includes: phase difference calculation and pulse noise removal. The data pre-processing method as described in claim 1, wherein the step of analyzing the optimized signal includes: According to the optimized signal, a phase map and a vibration frequency map are obtained, wherein the phase map presents the energy distribution with the distance change and phase change relative to the motion physiological sensing radar, and the vibration frequency map presents the energy distribution with the distance change and vibration frequency change relative to the motion physiological sensing radar ; Select at least one candidate position with energy intensity exceeding an energy threshold from the vibration frequency map; Select a target position from the at least one candidate position, the target position being the one with a vibration frequency that meets a physiological parameter range and has the largest energy intensity among the at least one candidate position; and Based on the target position, obtain the one or more target phase data within a distance range in the phase map. The data pre-processing method as described in claim 4 further includes: statistically analyzing the energy distribution parameter set within a period to determine a detection distance area covering the target activity range, wherein the step of selecting the at least one candidate position is to select from the detection distance area in the vibration frequency map. The data pre-processing method as described in claim 4, wherein the steps of obtaining the phase map and the vibration frequency map include: performing fast Fourier transformation on the optimized signal to obtain a distance profile map, wherein the distance profile map presents energy distribution with distance changes and time changes relative to the motion physiological sensing radar; performing DC bias elimination, IQ imbalance compensation, inverse tangent and phase expansion on each distance delay change of the distance profile map to obtain the phase map; and performing fast Fourier transformation on the phase distribution at each distance of the phase map to obtain the vibration frequency map. The data pre-processing method as described in claim 4, wherein the step of selecting the at least one candidate position from the vibration frequency map includes: calculating the energy threshold for each distance bar in the phase map, the energy threshold being determined according to an energy average value or an energy maximum value of the corresponding distance bar; and comparing the energy value of each phase on each distance bar with the energy threshold of the corresponding distance bar to select the at least one candidate position exceeding the energy threshold. A sports physiological sensing radar includes: a transmitting unit, transmitting an incident radar signal; a receiving unit, receiving a reflected radar signal; and a signal processing module, controlling the transmitting unit and the receiving unit to perform beamforming scanning to obtain an energy distribution parameter set, and demodulating the reflected radar signal into an intermediate frequency signal and obtaining a corresponding digital signal through analog-to-digital conversion, and using the energy distribution parameter set to obtain a corresponding digital signal. Parameter set, searching for a target by filtering background noise, and weighting the digital signal according to the energy distribution parameter set to optimize the area near the target to obtain an optimized signal, and analyzing the optimized signal to extract one or more target phase data corresponding to the target from the optimized signal, and inputting the one or more target phase data into a machine learning model to obtain a physiological parameter prediction result. The sports physiological sensing radar as described in claim 8, wherein the signal processing module counts the energy distribution parameter set within a period to determine a detection distance area covering the target activity range, so as to analyze the optimized signal located in the detection distance area. The motion physiological sensing radar as described in claim 8, wherein the signal processing module also performs a signal processing on the one or more target phase data before inputting the one or more target phase data into the machine learning model, and the signal processing includes: phase difference calculation and pulse noise removal. The motion physiological sensing radar as described in claim 8, wherein the signal processing module obtains a phase map and a vibration frequency map according to the optimized signal, selects at least one candidate position with energy intensity exceeding an energy threshold from the vibration frequency map, and selects a target position from the at least one candidate position, and obtains the one or more target phase data within a distance range in the phase map according to the target position; wherein the phase map presents energy distribution with distance changes and phase changes relative to the motion physiological sensing radar, and the vibration frequency map presents energy distribution with distance changes and vibration frequency changes relative to the motion physiological sensing radar, and the target position is the one of the at least one candidate position that has a vibration frequency that meets a physiological parameter range and has the largest energy intensity. The motion physiological sensing radar as described in claim 11, wherein the signal processing module statistics the energy distribution parameter set within a period to determine a detection distance area covering the target activity range, so as to select the at least one candidate position from the detection distance area in the vibration frequency map. The motion physiological sensing radar as described in claim 11, wherein the signal processing module performs a fast Fourier transform on the optimized signal to obtain a distance profile map, and performs DC bias elimination, IQ imbalance compensation, inverse tangent and phase expansion on each distance delay change of the distance profile map to obtain the phase map, and performs a fast Fourier transform on the phase distribution at each distance of the phase map to obtain the vibration frequency map; wherein the distance profile map presents energy distribution with distance change and time change relative to the motion physiological sensing radar. The motion physiological sensing radar as described in claim 11, wherein the signal processing module calculates the energy threshold for each distance bar in the phase map, and compares the energy value of each phase on each distance bar with the energy threshold corresponding to the distance bar to select the at least one candidate position exceeding the energy threshold, wherein the energy threshold is determined based on an energy average value or an energy maximum value of the corresponding distance bar. A data pre-processing method is performed by a processor in a signal processing device, comprising: obtaining an energy distribution parameter set and a digital signal obtained by beamforming scanning, wherein the digital signal corresponds to a reflected radar signal of a motion physiological sensing radar; using the energy distribution parameter set to search for a target by filtering background noise; analyzing the digital signal to extract one or more target phase data corresponding to the target from the digital signal; dividing the one or more target phase data into a plurality of sub-bands by wavelet transform; performing statistical analysis on each of the sub-bands to obtain a statistical feature set; and inputting the statistical feature set into a machine learning model to obtain a physiological parameter prediction result. The data pre-processing method as described in claim 15, wherein the step of analyzing the digital signal includes: weighting the digital signal according to the energy distribution parameter set to obtain an optimized signal; and analyzing the optimized signal to extract the one or more target phase data corresponding to the target from the optimized signal. The data pre-processing method as described in claim 16, wherein the step of analyzing the optimized signal includes: obtaining a phase map and a vibration frequency map based on the optimized signal, wherein the phase map presents energy distribution with changes in distance and phase relative to the motion physiological sensing radar, and the vibration frequency map presents energy distribution with changes in distance and vibration frequency relative to the motion physiological sensing radar. ; Select at least one candidate position with energy intensity exceeding an energy threshold from the vibration frequency map; Select a target position from the at least one candidate position, the target position being the one with a vibration frequency that meets a physiological parameter range and has the largest energy intensity among the at least one candidate position; and Based on the target position, obtain the one or more target phase data within a distance range in the phase map. The data pre-processing method as described in claim 17, wherein the step of selecting the at least one candidate position from the vibration frequency map comprises: For each distance bar in the phase map, the energy threshold is calculated respectively, and the energy threshold is determined according to an energy average value or an energy maximum value of the corresponding distance bar; and the energy value of each phase on each distance bar is compared with the energy threshold of the corresponding distance bar to select the at least one candidate position exceeding the energy threshold. The data pre-processing method as described in claim 17 further includes: statistically calculating the energy distribution parameter set within a period to determine a detection distance area covering the target activity range, wherein the step of analyzing the optimized signal is to analyze the optimized signal located in the detection distance area. The data pre-processing method as described in claim 16, wherein the step of analyzing the optimized signal comprises: obtaining a phase map and a vibration frequency map according to the optimized signal, wherein the phase map presents energy distribution with changes in distance and phase relative to the motion physiological sensing radar, and the vibration frequency map presents energy distribution with changes in distance and vibration frequency relative to the motion physiological sensing radar; At least one candidate position with energy intensity exceeding an energy threshold is selected in the figure; N target positions are selected from the at least one candidate position and N is greater than 1, wherein the N target positions to be measured are those having a vibration frequency that meets a physiological parameter range and having the top N largest energy intensities in the at least one candidate position; and according to each of the target positions, the one or more target phase data within a corresponding distance range in the phase map are obtained. The data pre-processing method as described in claim 15, wherein before the one or more target phase data are subjected to wavelet transformation, the one or more target phase data are also subjected to a signal processing, and the signal processing includes: phase difference calculation and pulse noise removal.

Images