Millimeter wave radar sleep behavior identification method based on adaptive synchronous compression transformation

Abstract

The invention provides a millimeter wave radar sleep behavior identification method based on adaptive synchronous compression transformation. The invention relates to the technical field of sleep monitoring, a millimeter wave radar obtains sleep behavior intermediate frequency signals, and background clutters are suppressed by using a sliding window method. And carrying out fast Fourier transform on the intermediate frequency signal after the background interference is removed to generate a distance image, and realizing target detection and distance unit positioning by using a unit average constant false alarm rate detection algorithm. At a distance unit where a target is located, adaptive synchronous compression transformation is carried out on a distance image along a slow time axis, the window length of short-time Fourier transformation is adaptively adjusted according to a dominant frequency change trend, the time-frequency energy concentration degree is improved, and signal attenuation and frequency shift offset caused by distance and angle changes are compensated. And finally, constructing a cascaded CNN-LSTM network model, extracting spatial features through the CNN, capturing a time sequence dependency relationship through the LSTM, and realizing accurate recognition of sleep behaviors under different distance and angle conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of sleep monitoring, and particularly to a method for identifying sleep behaviors using a millimeter-wave radar based on adaptive synchrosqueezing transform. Background Art

[0002] With the increasing severity of sleep disorder problems, sleep monitoring technology has received wide attention. Existing sleep monitoring methods mainly include two categories: contact type and non-contact type. The former, such as polysomnography and wearable devices, although having high measurement accuracy, are inconvenient to wear and may affect natural sleep. Non-contact methods, such as millimeter-wave radar, can achieve remote and non-intrusive monitoring using the micro-Doppler effect and have broad application prospects.

[0003] In practical applications, millimeter-wave radar monitoring still faces challenges. For example, background clutter interference affects the extraction of micro-motion signals, and the non-stationarity of signals makes it difficult for traditional short-time Fourier transform (STFT) to balance time-frequency resolution. In addition, Doppler features are greatly affected by changes in target angle and distance. Especially at different azimuth angles, signal attenuation may lead to a decrease in the accuracy of behavior recognition. Summary of the Invention

[0004] The present invention proposes a method for identifying sleep behaviors using a millimeter-wave radar based on adaptive synchrosqueezing transform. This method adaptively adjusts the STFT window length, dynamically optimizes the time-frequency resolution according to the main frequency change, effectively suppresses the frequency dispersion phenomenon of non-stationary signals, and enhances the expression of Doppler features. A cascaded CNN-LSTM network is used to fully utilize spatial features and temporal information to achieve high-precision recognition of sleep behaviors and improve the stability and adaptability of the system.

[0005] The technical solution proposed by the present invention is as follows:

[0006] A method for identifying sleep behaviors using a millimeter-wave radar based on adaptive synchrosqueezing transform, comprising the following steps:

[0007] Step 1: Initialize the millimeter-wave radar system, set the radar working parameters, and collect intermediate-frequency signals of different sleep behaviors to provide data for subsequent behavior recognition;

[0008] Step 2: Use the sliding window averaging method to perform dynamic background clutter suppression on the intermediate-frequency signals. Estimate the background by calculating the mean value within the window and perform differential operations to remove static clutter, improving the clutter suppression rate while ensuring the integrity of micro-Doppler features;

[0009] Step 3: Subsequently, perform a fast Fourier transform on the filtered signal along the fast time axis to obtain a range image; set a global detection threshold according to the spectral mean of the range image, and use the cell averaging constant false alarm rate algorithm CA-CFAR to determine whether there is a target; after detecting the target, determine the range cell where the target is located by searching for the maximum value position of the range image;

[0010] Step 4: At the target range cell, perform an adaptive synchronous compression transform ASST on the range image along the slow time axis to obtain a time-frequency representation with enhanced Doppler characteristics;

[0011] Step 5: Repeatedly perform the adaptive synchronous compression transform ASST on all radar data frames, extract the instantaneous main frequency of each frame and splice them in chronological order to generate an enhanced Doppler time-frequency map EDTM, thereby fully characterizing the change of the Doppler characteristics of the target action over time;

[0012] Step 6: Input the EDTM into a cascaded CNN-LSTM network and train the model to recognize sleep behaviors;

[0013] Step 7: Use the trained model to recognize five sleep behaviors in the validation set, including sitting up, lying down, turning over, waving, and falling out of bed.

[0014] Furthermore, in the above Step 1, the initialization of the millimeter-wave radar includes setting the transmission frequency, bandwidth, sampling rate, and the number of transceiver antennas; the acquisition includes intermediate frequency signals of five sleep behaviors, including sitting up, lying down, turning over, waving, and falling out of bed.

[0015] Furthermore, in the above Step 2, for the acquired intermediate frequency signal, use the sliding window method to calculate the local background average value within a fixed window for each Chirp one by one, and subtract the average background value from the current Chirp, thereby effectively suppressing static interference. After completing background suppression, perform a one-dimensional Fourier transform on the filtered signal along the fast time axis to convert the time-domain signal to the frequency domain and generate a range image.

[0016] In the above Step 3, the present invention regards the entire spectrum of the range image of the current data frame as the reference cell without a guard cell, calculates the mean value of the spectrum within the reference cell to obtain an estimated value of the global background noise energy, and multiplies the estimated value of the global background noise energy by a scaling factor to obtain the global detection threshold, detect the entire range image through the detection threshold, determine whether the range image contains a target according to whether there is a range cell greater than the detection threshold, and in the case of containing a target, the range cell r where the target is located can be obtained by searching for the maximum value position of the range image.

[0017] In step 4, at the target range cell, perform an Adaptive Synchronous Compression Transform (ASST) on the range profile along the slow time axis to obtain a time-frequency representation with enhanced Doppler characteristics. The specific process is as follows:

[0018] (4.1) From the range profiles generated by the radar, select the range cell r where the target is located, and extract the signal sequence in its slow time dimension , where t represents the slow time axis. Then perform a Fourier transform on the signal to calculate the frequency distribution, and calculate the frequency distribution :

[0019]

[0020] Based on the peak value of the amplitude spectrum, determine the main frequency of the current signal and its changing trend.

[0021] (4.2) Dynamically adjust the length of the STFT window function based on the frequency band range where the main frequency of the signal is located. If the main frequency of the signal is higher than the preset frequency domain threshold , use a short window function , with a window length of , to improve the time resolution by shortening the time window and ensure the capture of transient characteristics of high-frequency components; if the main frequency of the signal is lower than the preset frequency domain threshold , use a long window function , with a window length , to improve the frequency resolution by extending the time window and ensure the spectral refinement of low-frequency components. Perform STFT calculation on the signal using the selected window function:

[0022]

[0023] Output the adaptive time-frequency distribution , where is the window function , is the angular frequency.

[0024] (4.3) Calculate the phase derivative of the STFT result to obtain the instantaneous frequency estimate value

[0025]

[0026] where is the phase value of the STFT calculation result. According to the estimated instantaneous frequency, redistribute the energy of the STFT to the estimated instantaneous frequency position to form a synchronous compressed time-frequency representation , and the redistribution process can be expressed as:

[0027]

[0028] Among them, represents the Dirac function. Through the energy redistribution of SST, a compressed time-frequency representation is obtained, which has a higher time-frequency resolution.

[0029] In step 5, based on the processing of a single-frame radar signal in steps 3 and 4, the above processing flow is sequentially executed for all data frames collected by the radar. Specifically, first for each frame of signal calculate its corresponding adaptive synchrosqueezing transform time-frequency representation . The obtained for each frame are arranged and spliced in chronological order to form a frame-by-frame continuous feature representation and smoothed at the boundaries between adjacent frames to obtain a time-frequency map EDTM with enhanced Doppler features, with time as the horizontal axis and Doppler frequency shift as the vertical axis, and the numerical value representing the signal intensity, characterizing the change of the Doppler features of the target action over time.

[0030] In step 6, the time-frequency map EDTM with enhanced Doppler features is used as the input and fed into a cascaded CNN-LSTM network for training. In this network structure, the convolutional neural network (CNN) module is responsible for extracting the spatial features in EDTM, while the long short-term memory network (LSTM) module is used to capture the temporal correlation information in the feature sequence.

[0031] During the training process, the labeled sleep behavior data is used as the supervision signal, the cross-entropy loss function is used to calculate the error between the predicted category of the model and the true label, and the network parameters are updated based on the error backpropagation. The Adam algorithm is selected as the optimizer, and the learning rate is adaptively adjusted to accelerate convergence and minimize the loss function. After the training is completed, the optimal weights of the network are saved for subsequent inference.

[0032] In step 7, the trained cascaded CNN-LSTM network is used to perform identification tests on the validation set. The specific process includes: first loading the saved model weights, then inputting the Doppler time-frequency map (EDTM) in the validation set into the network, and outputting the corresponding category prediction results. Finally, the identification accuracy is calculated, and the confusion matrix is drawn to comprehensively evaluate the identification performance of the model.

[0033] The present invention uses the adaptive synchrosqueezing transform to process the radar intermediate-frequency signal to generate a Doppler time map, and adopts a cascaded CNN-LSTM neural network to extract the spatial and temporal features of sleep behavior, improving the identification accuracy.

[0034] The beneficial effects of the present invention are:

[0035] 1. The present invention adopts the Adaptive Synchronous Compression Transform (ASST) to effectively compensate for the weakening of Doppler signals caused by changes in the target azimuth angle or distance, improve the time-frequency energy focusing degree, thereby enhancing the signal resolvability and reducing the impact of signal attenuation on behavior recognition.

[0036] 2. The present invention adopts a cascaded CNN-LSTM network. It uses 2D-CNN to extract the spatial information of the Doppler time map and uses LSTM to capture the time series characteristics of sleep behaviors, accurately distinguish different sleep behaviors, and significantly improve the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is the overall flowchart of the sleep behavior recognition method of the present invention;

[0038] Figure 2 is the flowchart of the radar data processing of the present invention;

[0039] Figure 3 The five sleep behaviors defined in the embodiments of the present invention are based on the Doppler time map of the short-time Fourier transform, where (a) lying down, (b) sitting up, (c) turning over, (d) waving, (e) falling out of bed;

[0040] Figure 4 is the time-frequency map of Doppler feature enhancement based on the Adaptive Synchronous Compression Transform for the five sleep behaviors defined in the embodiments of the present invention, where (a) lying down, (b) sitting up, (c) turning over, (d) waving, (e) falling out of bed;

[0041] Figure 5 is the network structure diagram of the present invention, where (a) is the overall cascaded CNN-LSTM network structure diagram, (b) is the 2D-CNN network structure, (c) is the feature flattening network structure, (d) is the LSTM network structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The following is a detailed description of the preferred embodiments of the present invention with reference to the accompanying drawings, so as to more clearly elaborate the advantages and features of the present invention and make it easier for those skilled in the art to understand the essence of the present invention. Through the description of specific embodiments, it is intended to further clarify the protection scope of the present invention and provide a basis for defining the claims.

[0043] Reference Figures 1 - 5 , the millimeter-wave radar sleep behavior recognition method based on the Adaptive Synchronous Compression Transform includes the following steps:

[0044] Step 1: Initialize the millimeter-wave radar system and set the radar working parameters, including the transmission frequency, bandwidth, sampling rate, etc., to ensure that the system can meet the capture requirements of different sleep behavior characteristics;

[0045] In the acquisition environment, install the radar device at the ceiling position to ensure coverage of the test area, and place the subject in the test area directly below the radar to collect data on five sleep behaviors: sitting up, lying down, turning over, waving, and falling out of bed. The radar emits a frequency-modulated continuous wave signal, receives the echo signal reflected by the target, and obtains an intermediate-frequency signal through mixing processing, which is expressed as:

[0046]

[0047] where, is the intermediate-frequency frequency that varies with time and contains Doppler feature information caused by target movement, is the instantaneous phase. The acquired intermediate-frequency signal provides a data basis for subsequent sleep behavior recognition.

[0048] Step 2: To reduce the interference of the static background on the detection of the target signal, perform background noise suppression processing on the intermediate-frequency signal. The specific method is sliding window filtering: Let be the time-domain intermediate-frequency echo of the i-th Chirp signal ( is the fast time), take the sliding window length N (N = 20), and the background estimate value at time t is the local mean of the previous N Chirps:

[0049]

[0050] where, , is the chirp signal period. Subtract the background estimate value from the current Chirp signal to obtain the signal with static clutter suppressed:

[0051]

[0052] After completing the background suppression, perform FFT on the background-removed signal of each Chirp along the fast time axis to convert it from the time domain to the frequency domain and generate a range image for characterizing the echo intensity distribution at different range cells. Its expression is:

[0053]

[0054] where, represents the Fourier transform performed on the fast time variable , and r is the corresponding range index.

[0055] Based on the range image , a global background estimation method without a protection unit is used for target detection. The amplitude spectrum of the entire range image is used as the reference unit, and its mean value is calculated as the global background noise estimation value. , and its formula is:

[0056]

[0057] where n represents the total number of reference units, represents the index of the i-th range unit.

[0058] According to the estimated background value and the set scaling factor α, the global detection threshold is calculated:

[0059]

[0060] Traverse each range unit. If its reflection intensity satisfies:

[0061]

[0062] then it is determined that there is a target signal in this unit. In the case where multiple units meet the conditions, the unit with the maximum reflection intensity is selected as the final range unit r where the target is located:

[0063]

[0064] Step 4: At the range unit where the target is located, perform a Fourier transform along the slow-time dimension to obtain the main frequency distribution of the current signal. Use different window sizes to perform a short-time Fourier transform on different main frequency components. After obtaining the result of the short-time Fourier transform, calculate the instantaneous frequency of the STFT, and finally use the synchrosqueezing transform to redistribute the energy of the STFT result.

[0065] Step 5: Based on the processing of the single-frame radar signal in Step 3 and Step 4, repeat the target detection and adaptive synchrosqueezing transform for all radar data frames to obtain a time-frequency map (EDTM) with enhanced Doppler characteristics. The characteristic maps of different sleep behaviors are as Figure 4 shown;

[0066] To further verify the performance advantages of the adaptive synchrosqueezing transform (ASST) proposed in the present invention in sleep behavior recognition, performance comparisons are made for the traditional short-time Fourier transform (STFT), continuous wavelet transform (CWT), and synchrosqueezing transform (SST).

[0067] In terms of the evaluation of time-frequency energy focusing, the Rényi entropy and the main frequency bandwidth are used as quantization indicators. The lower the Rényi entropy value, the more concentrated the time-frequency energy distribution; the narrower the main frequency bandwidth, the higher the frequency resolution. The results are shown in Table 1:

[0068] Time-frequency analysis method Rényi entropy Main frequency bandwidth (Hz) STFT 3.30 15.2 CWT 2.75 9.4 SST 2.53 7.6 ASST 2.13 5.1

[0069] Table 1

[0070] As can be seen from Table 1, the method of the present invention (ASST) is superior to the traditional time-frequency methods in terms of time-frequency energy concentration and frequency resolution, and can more clearly characterize the Doppler feature changes during the sleep behavior process.

[0071] Furthermore, to verify the stability and robustness of the present invention in extracting Doppler features under different signal-to-noise ratio conditions, this embodiment conducts a simulation test on the signal-to-noise ratio change situations caused by target distance change, azimuth angle change, and human body posture change in millimeter-wave radar applications. Referring to the commonly used low signal-to-noise ratio settings (−30 dB, −20 dB, −10 dB) in existing literature, the radar echo signal attenuation situations in actual scenarios such as long-distance detection and off-axis angle detection are respectively simulated.

[0072] To quantify the extraction performance of weak Doppler features, this paper introduces the normalized index Q value of the root mean square error of instantaneous frequency (defined as the ratio of the square of the recognized instantaneous frequency error to the square of the theoretical frequency), and the formula is:

[0073]

[0074] where is the recognized instantaneous frequency, is the theoretical instantaneous frequency. The smaller the Q value, the higher the instantaneous frequency recognition accuracy and the better the feature extraction effect.

[0075] The experimental results show that under different conditions such as signal-to-noise ratios of −30 dB, −20 dB, and −10 dB, the adaptive synchrosqueezing transform (ASST) method proposed by the present invention always maintains the lowest Q value, and has better instantaneous frequency extraction ability compared with the traditional short-time Fourier transform (STFT), continuous wavelet transform (CWT), and synchrosqueezing transform (SST). The specific comparison is shown in Table 2:

[0076] Signal-to-noise ratio condition STFT CWT SST ASST −30 dB 1.39 2.41 1.31 0.88 −20 dB 1.21 1.83 1.20 0.76 −10 dB 0.95 1.32 0.91 0.65

[0077] Table 2

[0078] As can be seen from the table, in the case of low signal-to-noise ratio (such as −30 dB), the Q value of the ASST method is reduced by about 36.7% compared with STFT, by about 63.5% compared with CWT, and by about 32.8% compared with SST; as the signal-to-noise ratio increases, the ASST method can still continuously maintain the lowest value, demonstrating excellent signal extraction ability.

[0079] Step 6: Use the EDTM feature map as input data and feed it into the designed cascaded CNN-LSTM model. The model adopts a supervised learning strategy, using five types of labeled sleep behavior tags (including sitting up, lying down, turning over, waving, and falling out of bed) as supervision signals, calculating the error between the predicted output and the true label using the cross-entropy loss function, and iteratively updating the network parameters through the Adam optimizer. Finally, save the network weight file with the best performance on the validation set.

[0080] Step 7: Load the optimal model weights trained and saved in Step 6, input the EDTM feature map in the validation set into the model for inference and prediction, and obtain the classification results of each sample. To further verify the effectiveness of the model architecture, the present invention conducts a systematic comparison of the cascaded CNN model, LSTM model, and CNN-LSTM model. As can be seen from the comparison results in Table 3, the CNN-LSTM model can fully combine the high-dimensional spatial feature extraction ability of the CNN module and the time series modeling advantage of the LSTM module, thus significantly improving the classification accuracy and system robustness in complex behavior scenarios.

[0081] Model Accuracy CNN 90.93% LSTM 82.72% CNN-LSTM 95.91%

[0082] Table 3

[0083] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the specification and drawings of the present invention, or directly or indirectly applied to other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A millimeter wave radar sleep behavior recognition method based on adaptive synchronous compression transform, characterized in that: The following steps are involved: Step 1: Initialize the millimeter wave radar system, set radar working parameters, collect intermediate frequency signals of different sleep behaviors, and provide data for subsequent behavior recognition; Step 2: Use the sliding window averaging method to suppress dynamic background clutter on the intermediate frequency signal. The mean value in the window is calculated as the background estimate, and a differential operation is performed to remove static clutter, thereby improving the clutter suppression rate while ensuring the integrity of the micro-Doppler feature. Step 3: Then, the filtered signal is subjected to a fast Fourier transform along the fast time axis to obtain a range image. The global detection threshold is set according to the spectrum mean of the range image, and the unit average constant false alarm algorithm CA-CFAR is used to determine whether a target exists. After the target is detected, the range unit r where the target is located is determined by searching for the maximum position of the range image. Step 4: At the target range unit, perform adaptive synchronous compression transform ASST on the range image along the slow time axis to obtain a time-frequency representation with Doppler feature enhancement; Step 4.1: Obtain the slow-time signal sequence of the target distance unit r , then the signal Perform Fourier transform to obtain the frequency distribution , and its calculation formula is as follows: ； Among them, t represents the slow time axis. By analyzing the spectrum Amplitude distribution, determine the main frequency of the current signal ; Step 4.2: Adaptively adjust the window function length in STFT according to the frequency range of the signal main frequency; if the signal main frequency Below the preset frequency domain threshold When using the long window function Otherwise, the short window function is used. , use the selected window function to Perform STFT operation, the calculation formula is as follows: ； in, is the adaptive time-frequency distribution, is the selected window function, is the angular frequency; Step 4.3: Further analysis of the STFT results Processing is performed to calculate the derivative of its phase with respect to time to obtain an estimate of the instantaneous frequency , and its calculation formula is as follows: ； According to the estimated instantaneous frequency, the energy of STFT is Reassign to the estimated instantaneous frequency position to form a synchronous compressed time-frequency representation , the redistribution process can be expressed as: ； in, is the Dirac function; the energy is redistributed through synchronous compression transformation SST to obtain the compressed time-frequency representation and enhance the micro-Doppler feature expression of the target; Step 5: Steps 3 and 4 constitute a processing flow for a single-frame radar signal. This processing flow is repeated for all collected radar data frames, and the adaptive synchronous compressed time-frequency representation is extracted frame by frame. , splice the time-frequency representations of each frame in sequence according to the time sequence to generate the Doppler feature enhanced time-frequency map EDTM; Step 6: Input EDTM into the cascaded CNN-LSTM network and train the model to recognize sleep behavior; Step 7: Use the trained model to identify five sleep behaviors in the validation set, including sitting up, lying down, turning over, waving hands, and falling out of bed.

2. The method for millimeter wave radar sleep behavior recognition based on adaptive synchronous compression transformation according to claim 1, characterized in that: The millimeter wave radar initialization in step 1 includes setting the transmission frequency, bandwidth, sampling rate and number of transceiver antennas; collecting intermediate frequency signals of five sleeping behaviors, namely, sitting up, lying down, turning over, waving hands and falling out of bed, directly below the radar, and ensuring that the collection time of each action is sufficient to cover the complete action cycle to ensure the integrity of the signal.

3. The method for millimeter wave radar sleep behavior recognition based on adaptive synchronous compression transformation according to claim 1, characterized in that: In step 2, when processing the collected intermediate frequency signal, the sliding window method is first used to calculate the local background average value in the previous fixed window for each chirp signal Chirp, and the average background value is subtracted from the current Chirp to effectively suppress static interference. After the background suppression is completed, the filtered intermediate frequency signal is subjected to a one-dimensional Fourier transform along the fast time axis to convert the time domain signal into the frequency domain, thereby generating a range image.

4. The method for recognizing sleeping behavior using millimeter wave radar based on adaptive synchronous compression transform according to claim 1, characterized in that: In step 3, the entire spectrum of the current radar data frame range image is used as a reference unit without a protection unit to prevent the target signal from being filtered by the protection unit. The reference unit spectrum is averaged to obtain a global background noise energy estimate, which is then multiplied by the scaling factor. , get the global detection threshold, and use the detection threshold to determine whether the range image contains the target: if there is a distance unit exceeding the threshold, it indicates that there is a target, and the distance unit where the target is located is determined by searching the maximum position of the range image.

5. The method for millimeter wave radar sleep behavior recognition based on adaptive synchronous compression transformation according to claim 1, characterized in that: In the step 6, the Doppler time map with Doppler feature enhancement is input into the cascaded CNN-LSTM network for behavior recognition training, and the process is as follows: EDTM is obtained as a data set, which is divided into a training set and a test set, and the training set is input into the cascaded CNN-LSTM network for training; wherein the 2D-CNN module is composed of three alternating 2D convolutional layers and pooling layers, the kernel size of the convolutional layer is 3, the pooling layer adopts maximum pooling, the kernel size is 2, the step size is 2, and the ReLU activation function is configured after the convolutional layer to accelerate convergence, and the Dropout layer is introduced to prevent overfitting; then, the spatial features extracted by the 2D-CNN are converted into a one-dimensional vector sequence through a flattening operation; the LSTM module is composed of two layers of LSTM networks, each layer has 64 hidden units, and a Dropout layer is introduced to enhance generalization ability to capture the temporal correlation between spatial features; finally, the high-level spatiotemporal features output by the LSTM are input into the classifier.

Images