AU2021107064B4

AU2021107064B4 - Non-contact human respiration detection with radar signals

Info

Publication number: AU2021107064B4
Application number: AU2021107064A
Authority: AU
Inventors: Changyang Li; Zhiyu Ning
Original assignee: Rudder Technology Pty Ltd
Current assignee: Rudder Technology Pty Ltd
Priority date: 2021-08-24
Filing date: 2021-08-24
Publication date: 2022-08-25
Anticipated expiration: 2029-08-24
Also published as: AU2021107064A4

Abstract

SA radar signal processing pipeline that processes raw radar signals for signal enhancement and noise reducing, which is used for human respiration detection. The radar sensor continuously collects I channel data and Q channel data. The two-channel data are then transformed by a convolutional transformer. The transformed data are combined as a unified series data for the detection of human breath rate using the frequency distribution method.

Description

Non-contact human respiration detection with radar signals

Field of the Invention

[0001] This invention relates to human respiration detection (breath rate) based on radar signals.

Background of the Invention

[0002] It is reported that 49% of males aged 40-69 years have problems with sleep apnea and the number increases to 62% in males aged over 70 years. Sleeping apnea is related to many health problems, such as daytime drowsiness, hypertension, heart attack, stroke, memory loss, diabetes, depression and insomnia. It may even cause early death in specific scenarios. Respiration detection during sleeping is a key to detect apnea, which could be inferred to some health issues.

[0003] There is no commercial product or solution for in-home respiration detection. The reasons can be summarized in the following two-folds. Firstly, compared to monitoring respiration, monitoring the heartbeat rate is more popular as it can be easily accessed by many wearable devices, such as smartwatches and ECG devices. Secondly, different to heartbeat rate that could be reflected by other kinds of signs, monitoring breath rate can only be accessed by counting the ups and downs of the chest. These two facts limit the development of non-contact human respiration detection.

[0004] As radar signals could reflect ups and downs of the chest, the present invention proposes to process and analyze radar signals for the detection of respiration.

[0005] The existing methods using radar signals to detect respiration are transforming radar signals into a frequency distribution, and then count the respiration within the reasonable frequency. A major challenge of such methods is that the generated frequency distribution can be easily and significantly affected by the environment, thus the frequency distribution contains many unexpected noises which impedes the accuracy of respiration detection.

Therefore, the present invention aims to propose a robust radar signal processing pipeline to minimize the variations of noise for respiration detection.

Summary of the Disclosure

[0006] The respiration detection system comprises a radar sensor and a convolutional transformer.

[0007] The collector collects continuous I channel signals and Q channel signals.

[0008] The collected I and Q channel signals are input to a convolutional transformer for signal enhancement and noise reduction.

[0009] The transformed I and Q channels are then combined as a unified series data for detection of respiration by the estimator.

[0010] The estimator uses a frequency distribution method to identify the breath rate.

Brief Description of the Drawings.

[0011] Figure 1 shows an exemplary radar-based respiration detection system in accordance with an embodiment.

[0012] Figure 2 shows exemplary data processing in the system of Figure 1 in accordance with an embodiment.

Description of Embodiments

[0013] Figure 1 shows a respiration detector 101 for a home environment, living room or the like.

[0014] Detector 101 comprises a collector 104, transformer 105 and estimator 106. The data 103 is operable in system memory 102 for interpretation and execution of the computational functionality.

[0015] In the embodiment shown, the radar device 107 is installed above ceiling 108.

[0016] The detector can be configured to detect the breath rate of a person 111 who is laying on bed 110 within a supervision area 109.

[0017] At a specific time t, the radar sensor 107 generates two kinds of signals, i.e., the I channel data 112 (denoted as It) and Q channel data 113 (denoted as Qt). By sequencing the I and Q channel data of a time period T(ti, t2 , ... , t, ), two series data could be obtained:

[0018] [Equation 1] I(T) [Itl, It2 ' -'' Itn]

[0019] [Equation 2] Q (T) [QtQt2' --- Qtn]

[0020] 1(T) and Q (T) will be then input to the proposed convolutional transformer 105 for signal enhancement and noise reduction, respectively. The convolutional transformer is composed of several ID convolution layers of which the details are disclosed in the table below:

LayerIndex Filter Size Filter Stride Pad Filter Channel

1 11 8 5 1024

2 3 1 1 2048

3 3 1 1 4096 4 3 1 1 4096

5 1 1 0 4096

6 3 1 1 4096 7 3 1 1 4096

8 3 1 1 2048

9 3 1 1 1024

10 1 1 0 1

[0021] Denoted the transformed I(T) and Q(T) as I(T) 114 and Q (T) 115, respectively. The

loss function to train the convolutional transformer is defined as below:

[0022] [Equation 3] L = z (- 7Qt_ -tL)1+ 2

[0023] Then, I(T) 114 and Q(T) 115 are combined as a unified series data 116 for the time

period T using the following equation:

[0024] [Equation 4] U(T) = arctanI( Q (T)

[0025] The unified series data U(T) 116 now can be processed into frequency distribution 117 as other methods adopted for respiration detection. In detail, the breath rate can be identified as the frequency (within 0.08Hz - 0.67Hz) that has the largest power in the frequency distribution.

[0026] The hyper-parameter n is empirically set to 60 seconds, however can be adjusted according to different environments.

[0027] The foregoing description, for purposes of explanation, used specific nomenclature to provide a thorough understanding of the invention. However, it will be apparent to one skilled in the art that specific details are not required to practice the invention. Thus, the foregoing descriptions of specific embodiments of the invention are presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed as obviously many modifications and variations are possible in view of the above teachings. The embodiments were chosen and described to best explain the principles of the invention and its practical applications, thereby enabling others skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the following claims and their equivalents define the scope of the invention.

[0028] The term "approximately" or similar as used herein should be construed as being within 10% of the value stated unless otherwise indicated.

Claims

1. A convolutional network method for breath rate detection, in operable communication with a radar device and frequency distribution method, wherein

- any radar device (for example self-injection-locking radar) that is capable of capturing the subtle change of skin by emitting continuous 1/Q channel signals is used as a collector, which sends the 1/Q channels signals to the convolution network,

- the convolutional network is comprised of ten layers, including one convolutional layer with large filter (with filter size being 11), seven convolutional layers with medium filter (with filter size being 3), and two convolutional layers with small filter (with filter size being 1),

- the 1/Q channel data (denoted as I(T) for the I channel data, and Q(T) for the Q channel data) is input to the convolutional network respectively for the generation of transformed data (denoted as I(T) for the transformed I channel data, and Q(T) for the transformed Q channel data), f(T) - the transformed data is combined as a unified series data using equation arctan

, - a frequency distribution is generated from the unified series data as normal and existing frequency distribution methods did (for example Fast Fourier Transform method), and on the frequency distribution, the frequency that has the largest power (within 0.08Hz - 0.67Hz) is regarded as the breath rate,

2 + (Qt- Qt)2 wherein the convolutional network method uses a loss function (It- t) 2n to train the ten-layer convolutional network.