A non-contact sleep stage identification and structure analysis method and system

Abstract

The present invention discloses a non-contact sleep staging identification and structure analysis method and system, belonging to the field of sleep monitoring. The method comprises: obtaining a chest physiological motion signal to be staged, wherein the chest physiological motion signal to be staged is a number of sleep segments, including a chest breathing motion signal and a heartbeat motion signal of a human body; inputting the chest physiological motion signal to be staged into a pre-trained neural network model, and obtaining the sleep staging identification and structure analysis results output by the neural network model. The present invention has the ability to accurately estimate sleep structure parameters, which is of great significance for sleep diagnosis; at the same time, it is low-cost, easy to use, and will not cause any constraints or sleep effects on users. It can be applied to daily intelligent sleep monitoring and has application value for home health monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of non-contact sleep monitoring, and relates to a sleep stage recognition and structure analysis method and system, and in particular to a non-contact sleep stage recognition and structure analysis method and system. Background Art

[0002] Sleep structure refers to the overall characteristics and changing trends during sleep at night, including the duration, proportion and periodic distribution of each stage. Good sleep structure is closely related to sleep quality. Accurate estimation of these parameters plays an important role in patient status assessment and doctor monitoring of treatment progress.

[0003] The sleep structure can be analyzed according to the following sleep structure parameters:

[0004] Total sleep time (TST): the total time of non-awakening stage, i.e. REM, light sleep and deep sleep;

[0005] Slow wave sleep time (SWST): NREM stage, that is, the total time of light sleep and deep sleep stages;

[0006] Sleep efficiency (SE): the ratio of total sleep time to time in bed;

[0007] Sleep onset latency (SL) and REM latency (REML): the time required from going to bed and preparing for bed to actually entering sleep or REM;

[0008] Light sleep percentage (LSP), deep sleep percentage (DSP), REM percentage (REMP): the proportion of light sleep time (LST), deep sleep time (DST) or REM time (REMT) in the total sleep time.

[0009] Sleep quality is crucial to human health, but traditional sleep monitoring methods, such as polysomnography (PSG), have problems such as high cost, complex use, and inconvenient wearing, making them difficult to popularize. In recent years, millimeter-wave radar technology has been applied to non-contact sleep monitoring due to its advantages such as non-contact and penetrability, but existing methods have shortcomings in terms of accuracy, anti-interference ability, and adaptability to abnormal signs. Based on this, the present invention proposes a non-contact sleep stage recognition and structure analysis method and system, which can effectively identify sleep stages and extract sleep structure parameters, has high accuracy and adaptability, and provides a convenient tool for home health monitoring. Summary of the invention

[0010] In order to solve the above problems, the present invention provides a non-contact sleep stage identification and structure analysis method and system. The technical solution adopted by the present invention is as follows:

[0011] A non-contact sleep stage identification and structure analysis method comprises the following steps:

[0012] Acquire a chest physiological motion signal to be staged, wherein the chest physiological motion signal to be staged is a plurality of sleep segments, including a chest breathing motion signal and a heartbeat motion signal of a human body;

[0013] Inputting the chest physiological movement signal to be staged into a pre-trained neural network model to obtain sleep stage recognition and structure analysis results output by the neural network model;

[0014] Among them, the neural network model includes a CRNN layer and a NCRF layer, the CRNN layer includes a CNN network and a RNN network, the CNN network is used to capture the multi-scale short-time features in the sleep segment, and the RNN network is used to extract sleep analysis sequence knowledge based on the multi-scale short-time features; the NCRF layer is used to obtain the node probability distribution at the sleep segment and the edge probability distribution of adjacent associated segments, calculate the joint conditional probability distribution of the chest physiological movement signal to be staged, and make a global optimal decision on the sleep stage identification and structural analysis results.

[0015] Furthermore, the method for acquiring the thoracic physiological motion signal is:

[0016] The FMCW wave is emitted into the environment through the radar system and the reflected signal of the human body is received, and the chest breathing movement signal and the heartbeat movement signal in the reflected signal are extracted to obtain the chest physiological movement signal.

[0017] Furthermore, the training method of the neural network model is:

[0018] Acquire chest physiological motion signals throughout the night through a multi-transmit and multi-receive radar system, wherein the multi-transmit and multi-receive radar system includes a plurality of radar units, and the radar units include a plurality of antenna pairs;

[0019] Preprocessing the whole-night chest physiological motion signal to obtain an original sample set;

[0020] The neural network model is trained using the original sample set through a two-stage training method to obtain a trained neural network model.

[0021] Furthermore, the preprocessing of the whole night chest physiological motion signal is specifically as follows:

[0022] The whole-night chest physiological movement signal is divided into a number of sleep segments, the sleep segments are normalized, and a sleep stage label is assigned to each of the sleep segments. The sleep segments and the corresponding sleep stage labels are used as original samples, and an original sample set is generated based on multiple original samples.

[0023] Furthermore, the two-stage training method includes:

[0024] The first stage is: shuffling and randomly sampling the original sample set in units of antenna pairs to obtain a pre-training sample set, and using the pre-training sample set to pre-train the CRNN layer;

[0025] The second stage is: shuffling and randomly sampling the original sample set in units of radar units to obtain an overall training sample set, and using the overall training sample set to perform end-to-end overall training on the neural network model.

[0026] Furthermore, the pre-training of the CRNN layer uses an L1 loss function and a stochastic gradient descent method based on mini-batch gradient to optimize parameters;

[0027] The end-to-end overall training of the neural network model uses an L2 loss function and a stochastic gradient descent method based on mini-batch gradient to optimize parameters.

[0028] Furthermore, in the overall end-to-end training process of the neural network model, a smaller learning rate is used for the CRNN layer parameters, and a larger learning rate is used for the NCRF layer parameters.

[0029] Furthermore, the RNN network adopts a bidirectional long short-term memory network.

[0030] Furthermore, the CNN network input is the sleep segment of the chest physiological movement signal to be staged, and the output is the wakefulness period label and the non-awakening period label; the RNN network input is the non-awakening period sleep segment, and the output is the rapid eye movement period label, the light sleep period label and the deep sleep period label; the NCRF layer input is the entire sleep segment sequence with sleep labels, and the output is the optimized sleep staging and structure label sequence.

[0031] A non-contact sleep stage recognition and structure analysis system, the non-contact sleep stage recognition and structure analysis system comprising:

[0032] at least one processor;

[0033] and, a memory communicatively coupled to the at least one processor;

[0034] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the above-mentioned non-contact sleep stage identification and structure analysis method.

[0035] The beneficial effects of the present invention are:

[0036] (1) The present invention provides a novel non-contact sleep structure monitoring method and system. It uses artificial intelligence methods to understand the sleep state change process implicit in the radio frequency signal, and proposes a convolutional neural network (CNN)-recurrent neural network (RNN)-neural conditional random field (NCRF) model to accurately identify sleep stage changes and extract sleep structure characteristic parameters. The present invention has the ability to accurately estimate sleep structure parameters, which is of great significance for sleep diagnosis; at the same time, it is low-cost, easy to use, and will not cause any constraints or sleep effects on users. It can be applied to daily intelligent sleep monitoring and has application value for home health monitoring.

[0037] (2) The present invention has good adaptability to different objects and only requires a small amount of training and fine-tuning to adapt to the needs of new equipment.

[0038] (3) The present invention has good accuracy in estimating sleep structure parameters and has certain application potential in sleep cycle assessment and sleep event detection (such as awakening events). BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of a non-contact sleep monitoring system in an embodiment of the present invention.

[0040] Figure 2 This is an example diagram of the input signal of the neural network model in an embodiment of the present invention.

[0041] Figure 3 Schematic diagram of the neural network model in an embodiment of the present invention.

[0042] Figure 4 Schematic diagram of sleep stages in an embodiment of the present invention.

[0043] Figure 5 This is a graph of leave-one-out cross-validation results based on 6 GHz frequency band data in an embodiment of the present invention.

[0044] Figure 6 Graph showing experimental results based on 60 GHz frequency band data in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0046] A non-contact sleep stage identification and structure analysis method comprises the following steps:

[0047] Acquire a chest physiological motion signal to be staged, wherein the chest physiological motion signal to be staged is a plurality of sleep segments, including a chest breathing motion signal and a heartbeat motion signal of a human body;

[0048] Inputting the chest physiological movement signal to be staged into a pre-trained neural network model to obtain sleep stage recognition and structure analysis results output by the neural network model;

[0049] Among them, the neural network model includes a CRNN layer and a NCRF layer, the CRNN layer includes a CNN network and a RNN network, the CNN network is used to capture the multi-scale short-time features in the sleep segment, and the RNN network adopts a bidirectional long short-term memory network to extract sleep analysis sequence knowledge based on the multi-scale short-time features; the NCRF layer is used to obtain the node probability distribution at the sleep segment and the edge probability distribution of adjacent associated segments, calculate the joint conditional probability distribution of the chest physiological movement signal to be staged, and make a global optimal decision on the sleep stage identification and structural analysis results.

[0050] The CNN network input is the sleep segment of the chest physiological movement signal to be staged, and the output is the wakefulness period label and the non-awakening period label; the RNN network input is the non-awakening period sleep segment, and the output is the rapid eye movement period label, the light sleep period label and the deep sleep period label; the NCRF layer input is the entire sleep segment sequence with sleep labels, and the output is the optimized sleep staging and structure label sequence.

[0051] The method for acquiring the thoracic physiological motion signal is as follows: transmitting FMCW waves into the environment through a radar system and receiving human body reflection signals, extracting thoracic respiratory motion signals and heartbeat motion signals from the reflection signals, and obtaining the thoracic physiological motion signal.

[0052] The training method of the neural network model is as follows: obtaining the whole-night chest physiological motion signal through a multi-transmitter and multi-receiver radar system, wherein the multi-transmitter and multi-receiver radar system includes a plurality of radar units, and the radar units include a plurality of antenna pairs; preprocessing the whole-night chest physiological motion signal to obtain an original sample set; and using the original sample set to train the neural network model through a two-stage training method to obtain a trained neural network model.

[0053] The preprocessing of the whole-night thoracic physiological movement signal is specifically as follows: dividing the whole-night thoracic physiological movement signal into a number of sleep segments, normalizing the sleep segments, assigning a sleep stage label to each of the sleep segments, taking the sleep segments and the corresponding sleep stage labels as original samples, and generating an original sample set based on multiple original samples.

[0054] The two-stage training method includes:

[0055] The first stage is: shuffling and randomly sampling the original sample set in units of antenna pairs to obtain a pre-training sample set, and using the pre-training sample set to pre-train the CRNN layer;

[0056] The second stage is: shuffling and randomly sampling the original sample set in units of radar units to obtain an overall training sample set, and using the overall training sample set to perform end-to-end overall training on the neural network model.

[0057] The pre-training of the CRNN layer uses the L1 loss function and the stochastic gradient descent method based on the mini-batch gradient to optimize the parameters; the end-to-end overall training of the neural network model uses the L2 loss function and the stochastic gradient descent method based on the mini-batch gradient to optimize the parameters.

[0058] During the end-to-end overall training of the neural network model, a smaller learning rate is used for the CRNN layer parameters, and a larger learning rate is used for the NCRF layer parameters.

[0059] Example

[0060] like Figure 1 The figure shows the schematic diagram of the proposed non-contact sleep monitoring system. By placing a radar sensor next to the bed, FMCW waves are emitted into the environment and the signals reflected by the human body are received. The reflected signals contain information about breathing, heartbeat and limb movement, and the data is transmitted to the cloud server. Through the recorded sleep data throughout the night, a neural network model is designed and trained to learn the potential representation information contained in the radio frequency signal. Finally, the model is used to identify the sleep stages of new subjects for testing.

[0061] The extracted whole-night chest physiological motion signals are divided into several 30-second sleep segments, and the chest breathing motion signals and heartbeat motion signals are extracted from each 30-second sleep segment. Figure 2 (a) and (b) in the figure show examples of chest physiological motion signal acquisition during sleep at night using 6 GHz and 60 GHz frequency band sensors. Since radio frequency signals from different devices and different frequency bands may have different reflection coefficients, signal-to-noise ratios, and spatial resolution capabilities, in order to eliminate differences between sensors, environments, and individuals, the chest physiological motion signals are normalized and ultimately used as input for the sleep staging task:

[0062]

[0063]

[0064] in, For a 30-second sleep segment, , are the extracted chest respiratory motion signal and heartbeat motion signal, respectively, both are input column vectors, Indicates Z-score standardization, , are the mean and variance of the original chest respiratory motion signal and heartbeat motion signal, respectively. Figure 2 (c) and (d) show examples of normalized chest respiratory motion signals and heartbeat motion signals for a 30-second sleep segment.

[0065] The model task of this embodiment uses a multi-transmit and multi-receive radar signal as input and defines a single whole-night sleep segment sequence as ,in, For a 30-second sleep segment, is the number of 30-second segments in the whole night data; the task goal is to Output sleep stage labels, including wakefulness (A), REM (R), light sleep (L), and deep sleep (D). Therefore, the label set and sleep stage sequence are defined as and .

[0066] Figure 3 The figure shows a schematic diagram of the neural network model structure for sleep stage recognition, which consists of a CRNN layer and a NCRF layer. The CRNN acts as an encoder to learn the potential representation in the chest physiological motion signal, extract the multi-scale short-term features of the 30-second sleep fragment, and learn the sleep stage sequence knowledge. The NCRF layer establishes a sleep stage transfer model and fuses the chest physiological motion signal through a voting mechanism.

[0067] The CRNN layer includes a CNN network and a RNN network, such as Figure 3 As shown in (a), the CNN network is designed to capture multi-scale short-term features in sleep segments. In order to prevent degradation problems caused by the complexity of the overall network structure of the model and to maintain the gradient correlation in the learning process as much as possible, the CNN network consists of an input convolution layer, a maximum pooling layer, 4 convolution modules and an average pooling layer. The modules are directly connected to add the downsampled input features to the output features. The design of the input convolution layer is inspired by signal processing technology, and two one-dimensional convolution kernels of different sizes are used to extract short-term features of different frequency components. Each convolution module includes 4 one-dimensional convolution layers. Each convolution layer uses a one-dimensional convolution kernel of the same size as a feature filter, followed by a batch normalization layer and an activation layer (ReLu). The CNN network is simplified as Function, CNN output feature is defined as :

[0068]

[0069] It is further necessary to characterize the dynamic feature information and potential rules of stage changes during sleep. For example, each cycle of the light sleep stage lasts about half an hour, and then the body usually enters deep sleep, but the vital sign characteristics of light sleep and deep sleep are highly similar to each other. The features extracted by CNN can distinguish them from wakefulness or REM, but cannot distinguish the depth of sleep. Therefore, the present invention uses a CNN network to distinguish between wakefulness and non-wakefulness, and uses an RNN network to distinguish between rapid eye movement, light sleep and deep sleep in non-wakefulness sleep segments. Figure 3 As shown in (b), in this embodiment, the RNN network uses a two-layer bidirectional long short-term memory network (Bi-SRU) to process the non-awakening sleep segment:

[0070]

[0071] in, For the Gate of Forgetfulness, To reset the gate, is the hidden state, is the output feature, is the parameter matrix, is the parameter vector, and is the activation function. In this embodiment, a two-layer bidirectional SRU is used, and its structure can achieve high parallelization, greatly improving the speed of task calculation and learning. In addition, during model training and testing, the hidden states in the model at the beginning and end of each full-night sleep sequence will be set to zero.

[0072] The CRNN layer establishes a representation of the information in the thoracic physiological motion signal and outputs a concatenated feature vector:

[0073]

[0074] in, are all column vectors. Estimate the probability of each sleep stage label And after Softmax output:

[0075]

[0076] in, Represents sleep stages in the feature vector Conditional probability distribution under , node label predictor It consists of two fully connected layers. To pre-train the encoder CRNN and node predictor, minimize the cross entropy between the estimated label probability distribution and the true value :

[0077]

[0078] In the formula, is the probability distribution of the true labels of sleep stages, represents the parameter set in CRNN, Represents the weight coefficients of different labels.

[0079] The trained CRNN encoder can extract the optimal features from the chest physiological movement signal, and its sleep stage recognition result is based on the output label probability distribution of the current sleep segment. However, sleep stage has significant transformation structure limitations. For example, if you do not experience the light sleep stage, you cannot enter the deep sleep stage. Figure 3 (d) shows the transition rules of different sleep stages, and the conditional random field model is used as a probabilistic graph model for discriminant, which is usually used to model the target sequence. As shown in (e) in 3, the NCRF state transition model is established to estimate the node probability distribution at any sleep segment and the edge probability distribution of adjacent associated segments, calculate the joint conditional probability distribution of the whole night sequence, and make a global optimal decision on the sleep stage output sequence. Since the RF signals measured by different antenna pairs are subject to different multipath effects and noise, there are differences in the extracted feature vectors and the output probability distribution, so the output results of different antenna pairs need to be selected. Antenna pair eigenvector By sharing the dimension reduction layer After splicing , by voting layer Comprehensively understand the information between different antenna pairs and calculate the weights The dimension reduction layer With voting layer All are implemented through the fully connected layer. Node probability in the NCRF model and represent sleep stage shifts The marginal probability It can be expressed as:

[0080]

[0081]

[0082] In the formula, According to The node probability distribution estimated by the antenna pair data, is the joint feature vector of adjacent sleep segments, is the parameter matrix, is the Bayesian term, the marginal probability at any moment after the shape transformation For size The state transfer matrix (in this embodiment ). The final NCRF layer calculates the joint conditional probability distribution of the whole night sleep sequence for:

[0083]

[0084] In the formula, is the global normalization factor, defined as To further train the entire model end-to-end, the loss function The negative log-likelihood (NLL) function of the overnight sleep sequence is defined as:

[0085]

[0086] In the formula, is the probability distribution of the true label of the whole night sleep stage sequence, is the parameter set in the NCRF model.

[0087] In order to train the model, the present invention proposes a two-stage training method, which pre-trains the CRNN layer and performs end-to-end overall training on the model. The first stage is supervised pre-training on the CRNN layer. Since different antenna pairs of the radar sensor observe the same sleep process but are affected by different noises, the data set is shuffled and randomly sampled in units of antenna pairs in the first step of pre-training as a data enhancement so that the feature encoder can better resist observation noise. At the same time, since the category distribution during sleep is unbalanced (for example, REM sleep is less than 10% of the annotations), a category-sensitive loss function is used in the first step of pre-training. And the encoder parameters are modified by stochastic gradient descent (SGD) based on mini-batch gradient Optimize, the learning rate used in training is In the second stage, the overall model is trained end-to-end. The parameters are initialized according to the pre-training results. During training, the training data set is shuffled and randomly sampled in units of radar units, and the loss function is used. And SGD is used to optimize the encoder parameters Use a smaller learning rate Fine-tune the fusion recognizer parameters Use a larger learning rate .

[0088] In addition, this embodiment also uses two techniques to prevent overfitting. One is L2 regularization in the loss function to prevent the parameter value from exploding gradients when it is too large. The weight decay parameter is applied to the CRNN and NCRF parameter training process, and the value is 0.001. The other is Dropout, which randomly ignores a small number of hidden nodes in each training batch, uses Dropout with a discard probability of 0.2 for SRU features, and uses Dropout with a discard probability of 0.4 for CRNN joint features, but Dropout is only used in the training process and does not work in the verification and testing process.

[0089] In order to illustrate the effect of the present invention, this embodiment uses two different frequency bands of radio frequency data for verification. In this embodiment, 8 healthy adults were recruited as experimental subjects for a full-night sleep monitoring experiment. The experimental subjects had no sleep disorders and other related diseases and had regular work and rest. During the experiment, the radar sensor was fixed to the bedside, and the sensor angle was pointed obliquely downward to the middle chest area of ​​the bed. It emitted FMCW waves in the 60 GHz frequency band and received and recorded echo signals. The sampling frequency was 25 Hz, and all data were stored on the computer through an acquisition card. At the same time, the experimental subjects wore standard PSG equipment (SOMNOscreen plus, SOMNOmedics GmbH, Germany), and recorded 12-channel EEG / EOG, 1-channel ECG, respiratory flow, blood oxygen saturation and other data during the experimental subjects' sleep. According to the AASM manual, clinicians marked each 30-second sleep segment as one of the four sleep stages (wakefulness, REM, light sleep and deep sleep), which was used as the gold standard for sleep staging. Before each experiment, the PSG device and the host of the radar sensor were synchronized through the server, and the synchronization error was less than 1 second. In addition, this example also uses the RF-Sleep dataset provided by the Massachusetts Institute of Technology (MIT), using a 6 GHz radar sensor with a sampling frequency of 25 Hz to collect sleep data from 25 healthy adults. In each set of data, 30-second segments are also labeled for sleep stages based on EEG. In the end, there are 120 sets of radio frequency-based full-night sleep data, totaling about 100,000 30-second sleep segments.

[0090] This example evaluates the performance of the proposed model by training and model migration between devices in different frequency bands, and performs leave-one-out cross validation (LOOCV) on different experimental objects to calculate the recognition accuracy and The coefficients are calculated and the errors of the extracted sleep structure parameters are verified, and the feasibility of the proposed method for sleep monitoring applications is analyzed. For data sets of different frequency bands, due to the large sample size in the MIT RF-Sleep 6 GHz sensor data set, model training and cross-validation were performed separately for its data set; and the number of experimental subjects of the 60 GHz radar used in this embodiment is relatively small, so the model trained with the 6 GHz frequency band data set is used for parameter migration for initialization, and the model training process is improved through the learned transferable knowledge, and further model fine training and cross-validation are performed. For LOOCV, one experimental subject is selected each time, and its data is used as the test set, and the model training and verification are performed on the sleep data of other experimental subjects.

[0091] Among them, the recognition accuracy is defined as the percentage of sleep segments whose sleep stage recognition results are consistent with PSG in the total number of sleep segments. The coefficient is the Kappa coefficient, defined as To identify the stage The standard value for the stage is The number of installments, To identify the stage The number of installments, The standard value for the period is The number of installments, is the total number of sleep stages in the task, The coefficient is defined as

[0092]

[0093] In the formula, , , The coefficient takes into account the chance of correct identification due to luck and the imbalance of categories, so it has higher robustness and effectiveness. The value is usually between 0 and 1, 0< <0.2 indicates low consistency, 0.2< <0.4 represents general consistency, 0.4< <0.6 represents moderate consistency, 0.6< <0.8 indicates high consistency. >0.8 means almost identical.

[0094] For the sleep structure parameter estimation, the parameter values ​​in the sleep report issued by the doctor based on PSG were used as standard values, and its performance was evaluated by calculating the mean error (ME), mean absolute error (MAE) and mean accuracy (MACC):

[0095]

[0096]

[0097]

[0098] In the formula, For the estimated sleep structure parameters of the night sleep sequence, is the standard value measured by PSG, is the total number of sleep sequences throughout the night in the dataset.

[0099] After experimental verification, the overall accuracy of the non-contact sleep staging method proposed in the present invention is 79.2%. The coefficient is 0.679, which generally proves that the model has the ability to extract characteristic information of physical sign changes from chest physiological movement signals and accurately identify sleep stages. Figure 4 The figure shows an example of sleep staging. Its recognition accuracy is at an average level. The overall trend of the sleep staging results shown in the example is basically consistent with the standard value.

[0100] Figure 5 The figure shows the sleep stage results of different experimental subjects by CRNN model and the model of the present invention. coefficient. The standard deviation of the coefficient is 0.053, and the maximum and minimum values ​​are 0.736 and 0.552 respectively. The standard deviation of the recognition accuracy is 3.1%, and the highest and lowest accuracy are 83.2% and 72.8% respectively. The results of cross-validation show that the method of the present invention has good adaptability to different experimental subjects. At the same time, the performance of the CRNN model and the CRNN+NCRF model are also compared in the figure. The addition of the NCRF layer improves the accuracy by about 0.036. The coefficient and an accuracy of about 2.6% demonstrate the effectiveness of NCRF in making joint optimal decisions from the output layer.

[0101] Figure 6 (a) shows the sleep stage results of different experimental subjects in the cross-validation of the 60 GHz dataset. The average values ​​of coefficient and recognition accuracy are 0.630 and 77.9% respectively.

[0102] Since the sample size of 60 GHz frequency band data is significantly smaller than that of 6 GHz, its recognition task is challenging. When training and learning are performed only by randomly initializing model parameters, there are significant bottlenecks in the training speed and final convergence accuracy of the model, which is prone to overfitting. The final model often cannot accurately identify sleep stages. Figure 6(b) shows the comparison of the loss function values ​​during the training process after the CRNN parameters are initialized using the model parameters trained with the 6 GHz dataset based on the concept of transfer learning. After the pre-training parameters are initialized, the convergence trend of the model in the early stage of training is basically the same, but it is not prone to overfitting, and the final recognition accuracy is significantly improved. The model migration results also show that although there are differences in the measurement of vital signs and the impact of noise on devices in different frequency bands, the extracted high-level features and sleep transition structures are still consistent, indicating that the proposed model only needs a small amount of training and fine-tuning to adapt to the needs of new devices, proving the effectiveness and adaptability of the model.

[0103] A non-contact sleep staging identification and structural analysis method model proposed in the present invention can conveniently and accurately perform sleep staging identification and structural analysis on the target's sleep. The present invention is effective and robust in the sleep staging task, has convenient and reliable non-contact measurement and basically equivalent recognition performance, and provides a practical solution for sleep cycle assessment and sleep event detection.

[0104] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1Process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0107] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0108] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0109] The above specific implementation modes are used to explain the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A non-contact sleep stage identification and structure analysis method, characterized in that: The following steps are involved: Acquire a chest physiological motion signal to be staged, wherein the chest physiological motion signal to be staged is a plurality of sleep segments, including a chest breathing motion signal and a heartbeat motion signal of a human body; Inputting the chest physiological movement signal to be staged into a pre-trained neural network model to obtain sleep stage recognition and structure analysis results output by the neural network model; The neural network model includes a CRNN layer and an NCRF layer, the CRNN layer includes a CNN network and an RNN network, the CNN network is used to capture the multi-scale short-time features in the sleep segment, and the RNN network is used to extract sleep analysis sequence knowledge based on the multi-scale short-time features; the NCRF layer is used to obtain the node probability distribution at the sleep segment and the edge probability distribution of the adjacent associated segments, calculate the joint conditional probability distribution of the chest physiological motion signal to be staged, and make a global optimal decision on the sleep stage identification and structure analysis results; the CNN network input is the sleep segment of the chest physiological motion signal to be staged, and the output is the wakefulness label and the non-awakening label; the RNN network input is the non-awakening sleep segment, and the output is the rapid eye movement label, the light sleep label and the deep sleep label; the NCRF layer input is the entire sleep segment sequence with sleep labels, the sleep labels include the wakefulness label and the non-awakening label, and the non-awakening label includes the rapid eye movement label, the light sleep label and the deep sleep label; the output is the optimized sleep stage and structure label sequence.

2. A non-contact sleep stage identification and structure analysis method according to claim 1, characterized in that: The method for obtaining the chest physiological motion signal is as follows: The FMCW wave is emitted into the environment through the radar system and the reflected signal of the human body is received, and the chest breathing movement signal and the heartbeat movement signal in the reflected signal are extracted to obtain the chest physiological movement signal.

3. The non-contact sleep stage identification and structure analysis method according to claim 1, characterized in that: The training method of the neural network model is: Acquire chest physiological motion signals throughout the night through a multi-transmit and multi-receive radar system, wherein the multi-transmit and multi-receive radar system includes a plurality of radar units, and the radar units include a plurality of antenna pairs; Preprocessing the whole-night chest physiological motion signal to obtain an original sample set; The neural network model is trained using the original sample set through a two-stage training method to obtain a trained neural network model.

4. The non-contact sleep stage identification and structure analysis method according to claim 3, characterized in that: The preprocessing of the whole night chest physiological motion signal is specifically as follows: The whole-night chest physiological movement signal is divided into a number of sleep segments, the sleep segments are normalized, and a sleep stage label is assigned to each of the sleep segments. The sleep segments and the corresponding sleep stage labels are used as original samples, and an original sample set is generated based on multiple original samples.

5. The non-contact sleep stage identification and structure analysis method according to claim 3, characterized in that: The two-stage training method includes: The first stage is: shuffling and randomly sampling the original sample set in units of antenna pairs to obtain a pre-training sample set, and using the pre-training sample set to pre-train the CRNN layer; The second stage is: shuffling and randomly sampling the original sample set in units of radar units to obtain an overall training sample set, and using the overall training sample set to perform end-to-end overall training on the neural network model.

6. A non-contact sleep stage identification and structure analysis method according to claim 5, characterized in that: The pre-training of the CRNN layer uses an L1 loss function and a stochastic gradient descent method based on mini-batch gradient to optimize parameters; The end-to-end overall training of the neural network model uses an L2 loss function and a stochastic gradient descent method based on mini-batch gradient to optimize parameters.

7. The non-contact sleep stage identification and structure analysis method according to claim 5, characterized in that: During the end-to-end overall training of the neural network model, a smaller learning rate is used for the CRNN layer parameters, and a larger learning rate is used for the NCRF layer parameters.

8. The non-contact sleep stage identification and structure analysis method according to claim 1, characterized in that: The RNN network adopts a bidirectional long short-term memory network.

9. A non-contact sleep stage identification and structure analysis system, characterized in that: The non-contact sleep stage identification and structure analysis system comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the non-contact sleep stage identification and structure analysis method described in any one of claims 1 to 8.

Images