Method for Detecting Human Respiration and Heartbeat Based on Terahertz and Recurrent Neural Network

Abstract

The present application discloses a method for detecting human respiration and heartbeat based on terahertz and recurrent neural network. This method obtains the human physiological perception signal through terahertz radar signal, preprocesses the information by means of noise reduction, then marks the processed signal data, and trains it through the constructed recurrent neural network model to obtain the relevant model after training. The technical key point is that the technical solution provided by the present application, on the basis of noise reduction processing of the terahertz echo signal, adopts the method of empirical mode decomposition to decompose the original signal into intrinsic mode functions at different frequencies, learns the data through the designed multi-layer recurrent neural network, constructs an identification model for human heartbeat and respiration information, can effectively remove the influence of noise on the terahertz echo signal, and has good robustness; by means of feature fusion of high-frequency signals and low-frequency signals, the detection accuracy of the signal is better improved.

Description

Technical Field

[0001] The present invention belongs to the field of human respiration and heartbeat monitoring, and specifically relates to a method for detecting human respiration and heartbeat based on terahertz and recurrent neural network. Background Art

[0002] For continuous monitoring of human heartbeat and respiration, the currently adopted solutions usually require users to wear special devices, such as electrocardiogram detectors, smart bracelets or chest straps, etc. These devices need to be in contact with the human body during the detection process, which is neither convenient nor very comfortable, and the applicable scenarios are also relatively limited; with the development of the Internet of Things, signal processing and artificial intelligence technologies, there are more and more methods and means for detecting human physiological characteristics, and some scholars have proposed non-contact physiological characteristic perception of the human body through electromagnetic waves.

[0003] The principle of non-contact perception is mainly to estimate the respiration and heartbeat states of the human body through the electromagnetic waves reflected by the human body, and extract the frequencies of respiration and heartbeat and the judgment of other states through the changes in the phase, frequency, etc. of the reflected electromagnetic waves; currently, the main non-contact human perception technologies mainly focus on low-frequency microwaves and millimeter waves (2.4 GHz - 70 GHz); however, due to bandwidth limitations, the spatial resolution of these frequency bands is often insufficient; for example, the resolution of a 20 GHz sensing system is only 21 cm, and the range resolution is 7.5 cm (at a distance of 5 meters); however, the amplitude of respiration is usually less than 12 mm, and the amplitude of heartbeat is usually less than 1 mm. Sensing using the microwave and millimeter wave frequency bands is not sensitive enough, making it difficult to achieve accurate detection;

[0004] The terahertz frequency band is an electromagnetic wave with a frequency of 0.1 THz - 10 THz, and its wavelength is between 30 and 3 mm. It has characteristics such as a large available bandwidth and being harmless to the human body, and can better sense the subtle changes in heartbeat and respiration, achieving more accurate detection; currently, there are not many studies on detecting human physiological characteristics through terahertz, and mainly focus on the detection of heartbeat and respiration frequencies. And the abnormal information of heartbeat and respiration has more extensive application values in fields such as medicine and biology. However, there are still many problems in using terahertz to sense human physiological characteristics; first, terahertz signals attenuate more severely in space and are easily affected by noise and interference, which causes certain difficulties in extracting the signals of heartbeat and respiration, and reasonable noise reduction algorithms need to be used to process the signals; second, the terahertz frequency band is high and the bandwidth is large, which also poses certain difficulties in extracting the characteristics of heartbeat and respiration. Using artificial intelligence technology to analyze the data can better extract relevant characteristics;

[0005] In existing technologies, relatively simple methods mainly obtain the frequencies of heartbeat and respiration by performing time-frequency transformation on the received terahertz echo signals, such as using operations like Fourier transform. However, such methods are relatively simple in processing and do not handle noise and interference signals. Further methods will suppress interference on the terahertz echo signals through filters or related algorithms, such as using operations like wavelet transform. However, such methods often eliminate high-frequency signals, and the abnormal features of heartbeat and respiration often lie in high-frequency signals. In addition, there are also some methods that use convolutional neural networks to classify abnormalities in respiration and heartbeat. However, this method is more suitable for image processing and not for the analysis of time-series signals, and relevant research has not comprehensively utilized the advantages of high-frequency and low-frequency signals for processing, resulting in low accuracy. Summary of the Invention

[0006] (1) Technical Problems to be Solved

[0007] The problem to be solved by the present invention is to provide a method for detecting human respiration and heartbeat based on terahertz and recurrent neural networks in view of the characteristics of severe signal attenuation, high frequency, and large bandwidth in the terahertz band. The method obtains human physiological perception signals through terahertz radar signals, preprocesses the information by means of noise reduction, signal decomposition, etc., constructs an adaptive noise suppression method, then marks the processed signal data, and trains through the constructed recurrent neural network model. After training, the relevant model can be used to detect the human respiration and heartbeat frequency conditions.

[0008] (2) Technical Solutions

[0009] To achieve the above object, the present invention is realized through the following technical solutions:

[0010] A method for detecting abnormalities in human respiration and heartbeat based on terahertz radar, comprising:

[0011] This method obtains human physiological signals through terahertz radar signals, preprocesses the information by means of noise reduction, then marks the processed signal data, and trains through the constructed recurrent neural network model. After training, the relevant model is obtained.

[0012] Further, it includes the following specific detection steps:

[0013] Step 1: Transmit the modulated terahertz signal onto the human body through a terahertz signal generating device, collect the human echo signal through a receiving device, demodulate it, digitize the information, and store it in a memory.

[0014] Specifically,

[0015] The terahertz signal receiving device can receive single or multiple terahertz signals, and subsequent algorithms can process different terahertz signals separately.

[0016] Step 2: Based on the Doppler principle, using the transmitted signal and echo signal of terahertz, calculate the displacement signal of the human chest cavity due to breathing and heartbeat;

[0017] Specifically,

[0018] Taking a linear frequency modulated continuous wave radar as an example, the distance information from the transmitter to the chest cavity can be obtained through the following formula:

[0019]

[0020] where Δf 1 and Δf 2 are the frequency differences between the transmitted signal and the echo during the rising and falling edges respectively, and K r is the signal frequency modulation slope. Then, based on the distance information at different times, the chest cavity displacement signal caused by human breathing and heartbeat can be simply calculated. The specific formula is as follows:

[0021] ΔR(t) = R(t) - R(t - 1).

[0022] Step 3: Denoise the collected signal through wavelet packet transform to obtain the human physiological characteristic signal;

[0023] Specifically,

[0024] Step 3 - 1, select a suitable wavelet basis and the number of wavelet decomposition layers j, and perform wavelet packet transform on the calculated displacement signal. The wavelet packet decomposition algorithm is as follows:

[0025]

[0026] where d represents the wavelet packet decomposition coefficient, n represents the wavelet node number, j represents the decomposition layer number, h and g represent the filter coefficients, and l and k represent the wavelet node numbers of each layer; according to different usage scenarios, the wavelet basis function can be selected including but not limited to Haar, Daubechies, Coiflets, and Symlets. The number of wavelet decomposition layers can be selected as 12, but not limited to this;

[0027] Step 3 - 2, perform threshold processing on the decomposed wavelet coefficients, and filter out noise by selecting a threshold; the minimax criterion threshold is used for selection here, and its definition is:

[0028]

[0029] where the noise level σ can be expressed as:

[0030] σ = middle(W 1，k |, 0 ≤ k ≤ 2 j-1 -1) / 0.6745

[0031] where W 1，k is the wavelet coefficient with a scale of 1;

[0032] Step 3-3: Perform an inverse transform on the wavelet coefficients after threshold processing to re-obtain the time series signal. The reconstruction algorithm is as follows:

[0033]

[0034] Step 4: Add N groups of positive and negative white noise signals to the signal after noise reduction in Step 3 to obtain N groups of new signals, where the amplitudes of every two groups of white noise signals are the same;

[0035] Specifically,

[0036] The signal group after adding N groups of Gaussian white noise can be expressed by the following formula:

[0037] x j (t) = ΔR(t) + (-1) j εν j (t)

[0038] where ΔR(t) represents the original signal, j = 1, 2,... N, ε is the standard deviation of white noise, and ν j is a Gaussian white noise signal that satisfies the standard normal distribution.

[0039] Step 5: Use empirical mode decomposition to perform empirical mode decomposition on the N groups of signals with added white noise in Step 4 to obtain N groups of integrated intrinsic mode functions;

[0040] Specifically,

[0041] In the said Step 5,

[0042] The steps for performing empirical mode decomposition on the N groups of human chest displacement signals with added Gaussian white noise are as follows:

[0043] Step 5-1: Find all the extreme points in the signal x j (t);

[0044] Step 5-2: Obtain the envelope line fitted by the extreme points and calculate the average value m j (t), and obtain the difference between the signal and the average value of the envelope line through the following formula

[0045]

[0046] Step 5-3: Judge Whether the criteria of the intrinsic mode function are met. If so, extraction is performed and denoted as If not, then is used as the decomposed signal, and the above steps are repeated;

[0047] Step 5-4: Repeat the above steps until the remaining signal is only a monotonic sequence or a constant sequence. Finally, the original signal is decomposed into a linear superposition of multiple groups of intrinsic mode functions:

[0048]

[0049] where a represents the number of decomposed intrinsic mode functions, and r a (t) is the final remaining monotonic sequence or constant sequence.

[0050] Step 6: Calculate the average value of the N groups of integrated intrinsic mode functions obtained in Step 6 to obtain the final component of the intrinsic mode function group, and filter out the intrinsic mode functions with frequencies higher than 50 Hz;

[0051] Specifically,

[0052] In the said Step 6,

[0053] The calculation of the average value of the N groups of integrated intrinsic mode functions can be obtained by the following formula:

[0054]

[0055] Step 7: Mark the components of the intrinsic mode function group obtained in Step 6 according to different physiological states of people, and indicate the corresponding human heartbeats and breathing types of different signals.

[0056] Step 8: Input the multiple intrinsic mode function components obtained in Step 7 into a pre-built recurrent neural network for training to obtain the final recurrent neural network model, which includes information on the weights, activation values, and abnormal types corresponding to the eigenvalues of each link;

[0057] Specifically,

[0058] In the said Step 8,

[0059] The recurrent neural network is a multi-layer structure, where the network structure of each layer consists of an input, a hidden layer, and an output. The input is each group of intrinsic mode functions, and the input is connected to the hidden layer, and the connection parameter is W IMFc ; in the hidden layer are various different neuron nodes, and each neuron node is also connected to each other, and the connection parameter is W cc ; the hidden layer is connected to the output, and each neuron node has an output, and the connection parameter is During the transmission process, as time goes by, the information change process between neuron nodes in the hidden layer is as follows:

[0060] c(t) = g(W cc ·c(t - 1)+W IMFc ·IMF i (t)+b c )

[0061] And the output of each neuron node can be expressed as:

[0062]

[0063] The number of layers of the multi-layer recurrent neural network in step 8 depends on the number of intrinsic mode functions decomposed. The hidden layer neuron nodes between different layers are interconnected, and the connection parameter W′ cc , and the change of information transfer between neuron nodes at different levels is as follows:

[0064]

[0065] Therefore, the output of each neuron node can be modified as:

[0066]

[0067] where l represents the serial number of different levels. That is, it can be expressed as the eigenvalue of different abnormal states.

[0068] Step 9: Load the trained recurrent neural network model into the system to identify the abnormal information of the heartbeat and breathing of different human bodies.

[0069] Specifically,

[0070] The main steps of the method for discriminating abnormal information of human breathing and heartbeat are as follows:

[0071] Step 9-1: Repeat the work of steps 1 to 6 on the collected terahertz echo signal of the human body to obtain each intrinsic mode function component of the human chest displacement signal during this time period;

[0072] Step 9-2: Perform operations on the intrinsic mode function components in step 9-1 through the multi-layer recurrent neural network model obtained in step 8 to obtain the corresponding output, and judge the abnormal state according to the output being closer to the characteristics of a certain abnormality.

[0073] (3) In summary, the present invention includes at least one of the following beneficial technical effects:

[0074] Based on the noise reduction processing of the terahertz echo signal, the technical solution provided by the present invention uses the empirical mode decomposition method to decompose the original signal into intrinsic mode functions at different frequencies, and learns the data through a designed multi-layer recurrent neural network to construct an identification model for human heartbeat and breathing information;

[0075] The present invention effectively removes the influence of noise on the terahertz echo signal and has good robustness;

[0076] The present invention better improves the detection accuracy of the signal by fusing the features of high-frequency signals and low-frequency signals;

[0077] The present invention can distinguish different states of heartbeat and breathing without manual judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is the overall flowchart of the present invention;

[0079] Figure 2 is a schematic diagram of the terahertz signal acquisition system of the present invention;

[0080] Figure 3 is a schematic diagram of the original signal and the intrinsic mode function after wavelet transform noise reduction and empirical mode decomposition of the present invention;

[0081] Figure 4 is the structure diagram of the multi-layer recurrent neural network constructed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0082] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. In addition, for the convenience of description below, the "upper", "lower", "left", "right", etc. cited are consistent with the upper, lower, left, right, etc. of the accompanying drawings themselves. The "first", "second", etc. in the following text are used for description and distinction, and have no other special meanings.

[0083] This example specifically describes the technical solution of the present invention by using 2000 groups of collected human terahertz echo signals and the corresponding abnormal states of heartbeat and breathing.

[0084] As Figure 1 shown, the method for detecting abnormal breathing and heartbeat of the human body based on terahertz radar includes the following steps:

[0085] Step 1, receive and read the terahertz echo signal,Figure 2 It is a schematic diagram of a terahertz signal acquisition system, mainly composed of a terahertz signal source, a terahertz receiver, and an industrial computer. The terahertz signal source emits terahertz signals onto the human body, and the terahertz receiver receives the terahertz echo signals and stores them in the industrial computer;

[0086] Step 2: Use the terahertz echo signal and the transmitted signal to calculate the displacement signal from the transmitter to the human chest cavity. Taking a linear frequency modulation continuous wave radar as an example, its calculation method is as follows:

[0087]

[0088] Where Δf 1 and Δf 2 are the frequency differences between the transmitted signal and the echo during the rising and falling edges respectively, and K r is the signal frequency modulation slope. According to the distances in different periods, the displacement signals of the human chest cavity caused by breathing and heartbeat can be obtained:

[0089] ΔR(t) = R(t) - R(t - 1)

[0090] Step 3: Perform noise reduction processing on each group of displacement signals through wavelet packet transform (the subsequent steps take a group of displacement signals as an example). The wavelet basis function uses the Daubechies function, and the number of wavelet decomposition layers is 12. The specific steps of noise reduction are as follows:

[0091] Step 3-1: Decompose the signal by wavelet packet according to the following formula:

[0092]

[0093] Where d represents the wavelet packet decomposition coefficient, n represents the wavelet node number, j represents the decomposition layer number, h and g represent the filter coefficients, and l and k represent the wavelet node numbers of each layer.

[0094] Step 3-2: Perform threshold processing on the decomposed wavelet coefficients, and filter the noise through the threshold selection function. The threshold selection function is defined as follows:

[0095]

[0096] Where the noise level σ can be expressed as

[0097] σ = middle(W 1，k |, 0 ≤ k ≤ 2 j-1 -1) / 0.6745

[0098] Where W 1，k is the wavelet coefficient with a scale of 1.

[0099] Step 3-3: Perform inverse transformation on the wavelet coefficients after threshold processing to re-obtain the time series signal. The reconstruction algorithm is as follows:

[0100]

[0101] Step 4: Add Gaussian white noise to the obtained reconstructed signal, a total of 20 groups are added, including 10 groups of negative signals and 10 groups of positive signals. The amplitudes of a pair of positive and negative signals are the same, resulting in 20 groups of signals with Gaussian white noise. The addition formula is as follows:

[0102] x j (t) = ΔR(t) + (-1) j εv j (t)

[0103] where ΔR(t) represents the original displacement signal, j = 1, 2,... 20, ε is the standard deviation of white noise, and v j is a Gaussian white noise signal that satisfies the standard normal distribution.

[0104] Step 5: Perform empirical mode decomposition on the 20 groups of signals with added Gaussian white noise. The specific steps are as follows:

[0105] Step 5-1: Find all the extreme points in the signal x j (t);

[0106] Step 5-2: Obtain the envelope line fitted by the extreme points and calculate the average value m j (t). Obtain the difference between the signal and the average value of the envelope line through the following formula

[0107] h j (t) = x j (t) - m j (t)

[0108] Step 5-3: Determine whether h j (t) meets the criteria of the intrinsic mode function. If it meets the criteria, extract it and denote it as If it does not meet the criteria, take h j (t) as the decomposed signal and repeat the above steps;

[0109] Step 5-4: Repeat the above steps until the remaining signal is only a monotonic sequence or a constant sequence. Finally, decompose the original signal into a linear superposition of multiple groups of intrinsic mode functions:

[0110]

[0111] where a represents the number of decomposed intrinsic mode functions, and r a (t) is the last remaining monotonic sequence or constant sequence.

[0112] Step 6: Average the 20 groups of intrinsic mode functions, and the solution formula is as follows:

[0113]

[0114] Finally, multiple groups of intrinsic mode functions of the original signal are obtained. For example, Figure 3 in the first group of signals is the original signal, and the following 10 groups are the set of intrinsic mode functions of the original signal.

[0115] Step 7: Mark the corresponding data, and the marking information can be directly stored in the file names of these intrinsic mode functions;

[0116] Step 8: Train the obtained data through the constructed multi-layer recurrent neural network model. Assume that after empirical mode decomposition, the original signals in this example all generate 10 groups of intrinsic mode functions. Then, the number of layers of the constructed multi-layer recurrent neural network is also 10. The specific structure is as Figure 4 shown. The input of each neuron node is the amplitude of the intrinsic mode function at a certain moment. The input connects to the neuron node, and its connection parameter is W IMFc , and the neuron node is connected to the output, and its connection parameter is In addition, the neuron nodes at the same level are connected in sequence, and its connection parameter is W cc , and the neuron nodes at different levels are also connected in sequence, and its connection parameter is W′ cc . The change in the information transfer between neuron nodes at different levels is:

[0117]

[0118] Therefore, the output of each neuron node can be expressed as:

[0119]

[0120] where l represents the serial number of different levels. According to the output of the neural network, determine as the final output result. Based on this result, different types of heartbeat and breathing abnormalities can be determined through linear classification. During the training process, the algorithm will automatically test the connection parameters to find the best connection parameters.

[0121] Based on the noise reduction processing of the human chest displacement signal, the technical solution provided by the present invention uses the method of empirical mode decomposition to decompose the calculated displacement signal into intrinsic mode functions at different frequencies, and learns the data through the designed multi-layer recurrent neural network to construct an identification model for human heartbeat and breathing abnormality information.

[0122] Step 9. After the model is established, the human body can be detected. After collecting the terahertz echo signal of the human body, the set of intrinsic mode functions of the collected signal can be obtained according to Steps 1-6. The relevant data is imported into the trained multi-layer recurrent neural network to obtain the output value, and the best matching classification is searched in the linear classification model in Step 8 to obtain the abnormal types of the human heartbeat and respiration detected.

[0123] Based on the noise reduction processing of the human chest displacement signal, the technical solution provided by the present invention uses the empirical mode decomposition method to decompose the calculated displacement signal into intrinsic mode functions at different frequencies, and learns the data through the designed multi-layer recurrent neural network to construct an identification model for human heartbeat and respiration abnormal information.

[0124] The present invention effectively removes the influence of noise on the human chest displacement signal and has good robustness.

[0125] The present invention better improves the detection accuracy of the signal by fusing the features of the high-frequency signal and the low-frequency signal.

[0126] The present invention can automatically distinguish different heartbeat and respiration states without manual judgment.

[0127] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for detecting human respiration and heartbeat based on terahertz and recurrent neural network, characterized in that, this method acquires human physiological perception signals through terahertz radar signals, preprocesses the information by means of noise reduction, then marks the processed signal data, and trains through a constructed recurrent neural network model to obtain a relevant model; including the following specific detection steps: Step 1: Transmit the modulated terahertz signal to the human body through a terahertz signal generating device, and collect the human body echo signal through a receiving device, demodulate and digitize the information, and store it in a memory; Step 2: According to the Doppler principle, use the transmitted signal and echo signal of terahertz to calculate the displacement signal of the human chest due to respiration and heartbeat; Step 3: Perform noise reduction processing on the calculated displacement signal through wavelet packet transform to obtain a relatively pure human physiological characteristic signal; Step 4: Add N groups of positive and negative white noise signals to the signals after noise reduction processing in Step 3 respectively to obtain N groups of new signals, where the amplitudes of every two groups of white noise signals are the same; Step 5: Use empirical mode decomposition to perform empirical mode decomposition on the N groups of signals with added white noise in Step 4 to obtain N groups of integrated intrinsic mode functions; Step 6: Calculate the average value of the N groups of integrated intrinsic mode functions obtained in Step 5 to obtain the final component of the intrinsic mode function group, and filter out the intrinsic mode functions with frequencies higher than 50 Hz; Step 7: Mark the component of the intrinsic mode function group obtained in Step 6 according to different human respiration and heartbeat abnormalities, and indicate the corresponding abnormal types of different signals; Step 8: Input the multiple intrinsic mode function components and marking information obtained in Step 7 into a pre-built recurrent neural network for training to obtain a final recurrent neural network model, which contains the weights, activation values of each link, and information on the abnormal types corresponding to the eigenvalues; Step 9: Load the trained recurrent neural network model into the system to be used for identifying the heartbeat and respiration abnormal information of different humans.

2. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network as claimed in claim 1, characterized in that: In the said Step 1, the terahertz signal receiving device receives single-channel or multi-channel terahertz signals, and subsequent algorithms can process different terahertz signals separately.

3. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network as claimed in claim 1, characterized in that: In the said Step 2, according to the Doppler principle, the distance information from the emission source to the chest can be calculated. Taking a linear frequency modulated continuous wave radar as an example, the formula is as follows: ； Wherein and are the frequency differences between the transmitted signal and the echo during the rising and falling edges respectively, is the signal frequency modulation slope. By simply calculating based on the distance information at different times, the thoracic displacement signal caused by human respiration and heartbeat can be obtained. The formula is as follows: 。 4. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network as claimed in claim 1, characterized in that: In the said Step 3, the steps of performing noise reduction processing on the calculated displacement signal through wavelet packet transform are: Step 3-1: Select an appropriate wavelet basis and the number of wavelet decomposition levels , and perform wavelet packet transform on the calculated displacement signal. The wavelet packet decomposition algorithm is as follows: ； Among them, represents the wavelet packet decomposition coefficient, represents the wavelet node number, represents the decomposition level, and represent the filter coefficients, and represent the numbers of wavelet nodes at each layer; according to different usage scenarios, the wavelet basis function selects Haar, Daubechies, Coiflets, and Symlets, and the wavelet decomposition level selects 12; Step 3-2, perform threshold processing on the decomposed wavelet coefficients, and filter out noise by selecting a threshold; the minimum-maximum criterion threshold is used for selection here, and its definition is: ； where the noise level can be expressed as: ； Among them is the wavelet coefficient with a scale of 1; Step 3-3: Perform inverse transformation on the wavelet coefficients after threshold processing to re-obtain the time series signal. The reconstruction algorithm is as follows: 。 5. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network according to claim 1, characterized in that: In the said step 4, The signal group after adding N groups of Gaussian white noise can be expressed by the following formula: ； Among them represents the original displacement signal, , is the standard table of white noise, is a Gaussian white noise signal that satisfies the standard normal distribution.

6. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network according to claim 1, characterized in that: In the said step 5, The steps of performing empirical mode decomposition on N groups of human chest displacement signals after adding Gaussian white noise are: Step 5-1, find all the extreme points in the signal ; Step 5-2, obtain the envelope line fitted with extreme points and calculate the average value , obtain the difference between the signal and the average value of the envelope line through the following formula ； Step 5-3, determine whether meets the criteria of the intrinsic mode function. If it meets, extract it and denote it as . If it does not meet, use as the decomposed signal and repeat the above steps; Step 5-4: Repeat the above steps until the remaining signal is only a monotonic sequence or a constant sequence. Finally, decompose the original signal into a linear superposition of multiple groups of intrinsic mode functions: ； Among them, represents the number of the decomposed intrinsic mode functions, is the last remaining monotonic sequence or constant sequence.

7. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network according to claim 6, characterized in that: In the said step 6, The average value of the N groups of integrated intrinsic mode functions can be obtained by the following formula: 。 8. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network according to claim 1, characterized in that: In the said step 8, The recurrent neural network has a multi-layer structure, where the network structure of each layer consists of an input, a hidden layer, and an output. The input is each group of intrinsic mode functions, and the input is interconnected with the hidden layer, and its connection parameter is ; In the hidden layer are various different neuron nodes, and each neuron node is also interconnected, and its connection parameter is ; The hidden layer is interconnected with the output, and each neuron node has an output, and its connection parameter is ; During the transmission process, as time goes by, the information change process between the neuron nodes in the hidden layer is as follows: ； The output of each neuron node can be expressed as: ； The number of layers of the multi-layer recurrent neural network in step 8 depends on the number of decomposed intrinsic mode functions. The hidden layer neuron nodes between different layers are interconnected, and their connection parameters , and the changes in the information transfer of neuron nodes between different levels are as follows: ； Therefore, the output of each neuron node can be modified to: ； Among them represents the serial numbers of different levels that is, the characteristic values that can represent different abnormal states 9. The method for detecting human respiration and heartbeat based on terahertz and recurrent neural network according to claim 8, characterized in that: In the said step 9, The main steps of the discrimination method for human respiration and heartbeat are: Step 9-1: Repeat the work of steps 1 to 6 on the collected human terahertz echo signal to obtain each intrinsic mode function component of the human chest displacement signal during this time period; Step 9-2: Perform operations on the intrinsic mode function components in step 9-1 through the multi-layer recurrent neural network model obtained in step 8 to obtain the corresponding output, and judge the abnormal state according to the output being closer to the characteristics of a certain abnormality.

Images