Sleep quality evaluation method based on millimeter wave radar and AI monitoring sensor

Abstract

The invention discloses a sleep quality evaluation method based on a millimeter wave radar and an AI monitoring sensor, and belongs to the technical field of sleep quality evaluation, and the method comprises the steps: collecting annotation data through the millimeter wave radar and the AI monitoring sensor module; performing sample generation based on a conditional generative adversarial network of locality preserving projection to realize data expansion; inputting the expanded data into a feature extraction model to train the feature extraction model; inputting the data after feature extraction into a feature dimension reduction model to train the feature dimension reduction model; inputting the data subjected to feature dimension reduction into a classifier to train a classifier model; and processing new sleep monitoring data by utilizing the trained feature extraction model, the data dimension reduction model and the classifier model. According to the method, the newly input multi-source sleep data can be effectively processed, real-time classification and evaluation are carried out, the classification accuracy and efficiency are improved, a scientific sleep quality basis is provided for a user, and the sleep quality can be managed and improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of sleep quality assessment, and specifically relates to a sleep quality assessment method based on millimeter wave radar and AI monitoring sensors. Background Art

[0002] As the pace of society accelerates, sleep problems have gradually become an important factor affecting people's health. Traditional sleep monitoring methods mostly rely on physical contact sensors, such as mattress pressure sensors or body-attached devices, which often limit the continuity and accuracy of monitoring due to inconvenience and reduced comfort for users. In addition, these traditional technologies often fail to provide a comprehensive assessment of sleep quality because they cannot comprehensively consider the impact of environmental factors and individual physiological changes.

[0003] In the prior art, the Chinese invention patent with application number CN202310334857.6 discloses a sleep quality assessment method, device and computer-readable storage medium, the method comprising: obtaining multiple sleep indicators associated with the subject's sleep process, and determining the subject's first sleep assessment result based on the multiple sleep indicators; obtaining a sleep signal corresponding to the sleep process, and determining the subject's second sleep assessment result based on the sleep signal; and determining the sleep quality of the sleep process based on the first sleep assessment result and the second sleep assessment result. The Chinese invention patent with application number CN201811156995.5 discloses a sleep quality quantification method based on machine learning, including the steps of: obtaining sleep feature values ​​involved in training to obtain a training sample set; using the training sample set to train a sleep quality quantification model to obtain a trained sleep quality quantification model; obtaining a sleep quality quantification value based on the trained sleep quality quantification model. The Chinese invention patent with application number CN202110224051.2 discloses a sleep quality assessment method and device based on multi-dimensional feature parameters, including collecting brain wave signals at two leads on the forehead of a user when he or she is sleeping; extracting the feature parameters of the brain wave signals in the time domain, frequency domain and nonlinear dimensions to generate multi-dimensional feature parameters; inputting the multi-dimensional feature parameters into a preset sleep analysis model so that the sleep analysis model generates a sleep analysis result for the user based on the multi-dimensional feature parameters; and evaluating the sleep quality of the user based on the sleep analysis result.

[0004] The existing technology has the following problems: 1. Data collection in the existing technology usually relies on a single source or only involves limited monitoring dimensions, and cannot comprehensively capture the various factors that affect sleep quality; 2. The existing data expansion technology does not fully utilize the local structural information of the generated data, resulting in deviations in the statistical characteristics of the generated samples and the original data. At the same time, feature extraction and dimensionality reduction technology fail to effectively solve the problem of feature dependency and redundancy, affecting the predictive performance of the model; 3. The existing classifier training method does not consider the long-range dependency and topological structure of the data, resulting in insufficient accuracy and stability of the model when processing complex data. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a sleep quality assessment method based on millimeter wave radar and AI monitoring sensor.

[0006] The technical solution adopted to solve the above technical problems is: the sleep quality assessment method based on millimeter wave radar and AI monitoring sensor includes the following steps:

[0007] S1. Data collection and annotation

[0008] The millimeter-wave radar and AI monitoring sensor modules collect and manually annotate data, and the annotated data types include high-quality sleep, general sleep, and poor-quality sleep to monitor the user's sleep status;

[0009] S2. Data expansion

[0010] Sample generation based on conditional generative adversarial networks based on local preservation projections to achieve data expansion;

[0011] The training process of the conditional generative adversarial network based on local preservation projection is as follows:

[0012] S201, set initial parameters for the generator and discriminator of the conditional generative adversarial network, the parameters include the weight and bias of the network, and set the generator to be G c , the discriminator is D c , the weight of the generator is The weight of the discriminator is The bias of the generator is The bias of the discriminator is Initialize the bias of the generator and discriminator to 0, and initialize the weights of the generator and discriminator as follows:

[0013]

[0014] In formula (1), is the initial weight, n c is the number of nodes in the input layer of the generator, n cis 10, is a normal distribution with a mean of 0 and a standard deviation of 1;

[0015] S202, setting corresponding conditional variables according to the data type to be generated, and taking the conditional variables and the random noise vector as inputs of the generator, expressed as:

[0016]

[0017] y c =f c (Cla) (3)

[0018] In formula (2), z c is the random noise generated by a random process, σ c is the standard deviation of the noise, σ c is 1.0,∈ c is unit Gaussian noise;

[0019] In formula (3), y c is the conditional variable mapped according to the sleep data type, Cla is the monitoring indicator corresponding to different data categories, and f c (Cla) is a mapping function of monitoring indicators corresponding to different data categories; the mapping function according to the data category is a single hidden layer neural network, which performs nonlinear mapping according to samples of different categories;

[0020] S203. During the training process of the generator and the discriminator, the local preservation projection of each batch of data is calculated to ensure that the generated data maintains the local structural features of the original data. The calculation formula of the local preservation projection is:

[0021]

[0022] M c =IW c (5)

[0023]

[0024]

[0025]

[0026] In formula (4), L c For local preservation of projection, X c is the input data matrix, L c () is the local preservation projection calculation function, For X c The transpose of M c is the locality preserving matrix;

[0027] In formula (5), I is the unit matrix, W c is the weight matrix;

[0028] In formulas (6), (7), and (8), W c (i,j) is the element in the i-th row and j-th column of the weight matrix,

[0029] For data points and The square of the Euclidean distance between c is the adjustment parameter, t c is 0.5, For point The local density of For point The local density of is the neighbor set of the ith data point, For data points and The square of the Euclidean distance between c is the standard deviation parameter;

[0030] S204, perform adversarial training iterations, alternately train the generator and the discriminator, the generator generates data close to the real data to deceive the discriminator, and the discriminator tries to distinguish the real data from the generated data. The loss function calculation method of adversarial training is expressed as:

[0031]

[0032]

[0033] In formula (9), is the loss function of the generator, For the expected operation, D c () is the discriminator function, G c () is the generator function, λ c To control the weight of the local preservation projection in the loss function, η c is the regularization coefficient, η c is 0.001, Weight The square of the L2 norm of ;

[0034] In formula (10), is the loss function of the discriminator;

[0035] S205, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed;

[0036] S3. Training feature extraction model

[0037] Based on the whale optimization algorithm, the search capability of the whale optimization algorithm is used to optimize the weights and biases of the neural network by simulating the quantum phase change process, so as to achieve efficient feature extraction and improve the convergence speed of the feature extraction model.

[0038] S301, initialize the parameters of the neural network, including the weight and bias of the neural network, and the initialization method is expressed as:

[0039]

[0040]

[0041] In formula (12), W p is the weight of the neural network, n p is the number of nodes in the previous layer of neural network,

[0042] is a normal distribution with a mean of 0 and a standard deviation of 1;

[0043] In formula (13), b p is bias;

[0044] S302. In the whale optimization algorithm, the position of the individual whale represents a set of parameters of the neural network. Each individual whale updates its position according to the quantum phase change theory, simulating the behavior of quantum particles in the potential energy field to explore the solution space. The parameter update method is expressed as:

[0045]

[0046]

[0047] In formulas (14) and (15), and are the weights and biases of the updated neural network, α p is the search step length, α p is 0.5, and are the weights and biases of the neural network with the best performance in the current search, λ p is the attenuation coefficient, λ p is 0.95, int(t) is the current iteration number, ΔW p and Δb p are the changes in weight and bias respectively;

[0048] S303, perform fitness evaluation of individual whales, use the current position of individual whales to evaluate the performance of the neural network model in the feature extraction task, and the fitness evaluation is calculated by the loss function, and the calculation method is expressed as:

[0049]

[0050] In formula (18), F p represents the fitness value, MSE() represents the mean square error, Y is the true label, is the predicted label, n dp is the number of neurons in the output layer, ω pi is the weight of the i-th neuron in the output layer, Y i and are the true value and predicted value of the i-th output respectively;

[0051] S304: Select and crossover whale individuals. According to the fitness results, select the best performing whale individuals and generate a new generation of whale individuals through quantum crossover operations to enhance the diversity and exploration ability of the model. In the process of updating weights and biases, the method of selecting whale individuals is expressed as follows:

[0052]

[0053]

[0054] In formula (19), and are two sets of weights randomly selected from the current population,

[0055] Generates a [0, ],

[0056] Is the selection weight and The probability of performing a crossover;

[0057] In formula (20), and is the corresponding bias;

[0058] S305, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed;

[0059] S4. Training feature dimensionality reduction model

[0060] The data after feature extraction is input into the feature dimensionality reduction model for training. The variational autoencoder based on nonlinear hashing is used for feature dimensionality reduction. Nonlinear hashing technology and deep feature decoupling strategy are used to map high-dimensional data to low-dimensional hash space through hash functions. Deep feature decoupling optimizes the independence between various features through the variational autoencoder framework, and enhances the model's ability to learn the intrinsic structure of features.

[0061] The training process of the variational autoencoder algorithm based on nonlinear hashing is as follows:

[0062] S401, initialize the network parameters of the variational autoencoder, and the initialization method is expressed as:

[0063]

[0064]

[0065] In formula (22), W r is the weight of the variational autoencoder, n r is the number of nodes in the input layer of the variational autoencoder encoder, It means the mean is 0 and the variance is The normal distribution of is the variance of the initialized Gaussian distribution, is 0.01;

[0066] In formula (23), b r is the bias of the variational autoencoder;

[0067] S402, the data after feature extraction is forward propagated through the encoder layer, where each layer is processed by a nonlinear activation function and finally reaches the hash layer, which maps the data to a low-dimensional space. The calculation method of the mapping function is expressed as:

[0068] z r =tanh(W r ·x r +b r ) (twenty four)

[0069] H r =sign(z r ) (25)

[0070] In formula (24), z r is the latent variable, is the output of the tanh activation function, tanh() is the tanh activation function, x r is the data input to the hash layer;

[0071] In formula (25), H r To map features to low-dimensional space, sign() is a sign function that converts continuous activation values ​​to {-1, 1};

[0072] S403. In the decoding stage, the difference between the reconstructed data and the original data is quantified by the reconstruction loss. In order to maintain the diversity of low-dimensional representation, KL divergence is used as a regularization term to measure the difference between the encoded potential distribution and the prior distribution. The calculation method of the loss function of the variational autoencoder is expressed as:

[0073]

[0074] In formula (26), L r is the loss function of the variational autoencoder, is the output reconstructed by the decoder, x r is the data input to the variational self-encoder, is the data after feature extraction, and β r is the regularization coefficient, β r is 0.3, D KL () is the KL divergence calculation function, which is used to quantize the potential distribution q(z r |x r ) and the prior distribution p(z r )

[0075] S404, using the gradient descent method to update the network parameters of the variational autoencoder, the parameter update method is expressed as:

[0076]

[0077]

[0078] In formulas (28) and (29), and are the updated weights and biases of the variational autoencoder, η r is the learning rate of the variational autoencoder, η r is 0.01, and is the loss function L r Gradients with respect to weights and biases;

[0079] S405, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed;

[0080] S5. Training classifier model

[0081] The data after feature dimension reduction is input into the classifier to train the classifier model. The fractional-order neural network based on heuristic convergence is used as the classification algorithm. The fractional-order calculus theory is used to adjust the propagation and update rules in the neural network to better capture the long-range dependence and intrinsic dynamic characteristics in the data.

[0082] The training process of the fractional-order neural network algorithm based on heuristic convergence is as follows:

[0083] S501, initializing the parameters of the fractional-order neural network, the initialization method is expressed as:

[0084]

[0085] b u,init =0 (31)

[0086] In formula (30), W u,init is the initialization value of the weight of the fractional-order neural network, n u-1 and n u are the number of nodes in the previous layer and the current layer respectively, ∈ u is a random number drawn from a standard normal distribution;

[0087] In formula (31), b u,init is the initialization value of the bias of the fractional-order neural network;

[0088] S502, input the feature vector after feature dimension reduction into the fractional-order neural network, process it through the fractional-order derivative and the nonlinear activation function, and pass it layer by layer until the output layer of the fractional-order neural network. The number of neurons in the output layer of the fractional-order neural network is equal to the number of sleep quality categories, and each neuron corresponds to one category. For each layer of the fractional-order neural network except the output layer, the calculation method of its output is expressed as:

[0089]

[0090] In formula (32), y u is the layer output, Sig() is the Sigmoid activation function, W ui is the weight of the i-th neuron in the current layer, x ui is the input from the previous layer, b u is the bias of the current layer;

[0091] S503, calculating the loss function of the fractional-order neural network, using a loss function based on a manifold regularization term, the manifold regularization term is used to keep data points of the same type close to each other in the mapped low-dimensional space, and to separate data points of different types as much as possible to ensure the smoothness of the classification boundary. The calculation method of the loss function of the fractional-order neural network is expressed as:

[0092]

[0093] In formula (33), L u is the loss function of the fractional neural network, Mr is the number of samples input in the current batch, and t ui is the target label, y ui is the network prediction output, λ u is the regularization parameter, A jk is the similarity between the jth data point and the kth data point;

[0094] S504, using the fractional calculus strategy to back-propagate the loss function, calculate the gradient of each parameter, and when updating the parameters, use a dynamically adjusted learning rate to adapt to the changes in data during training and the convergence speed of the algorithm. The calculation method of the update amount of the fractional parameter is expressed as:

[0095]

[0096]

[0097] In formulas (36) and (37), ΔW ui and Δb u are the update amounts of the weights and biases of the fractional-order neural network in the current iteration, η gdf is the learning rate of the fractional-order neural network, is the fractional derivative of the α-order loss function with respect to the weight parameter, γ gdf is the parameter used to adjust the update amount of the previous iteration, is the update amount of the weight of the fractional-order neural network in the previous iteration;

[0098] S505, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed;

[0099] S6. Sleep quality monitoring

[0100] The new sleep monitoring data is processed using the trained feature extraction model, data dimensionality reduction model and classifier model to achieve real-time monitoring and evaluation of sleep quality. The collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing, and the processed features are input into the classifier model for classifier training to obtain the classification results.

[0101] Furthermore, the AI ​​monitoring sensor module in step S1 includes a pressure sensor, a temperature and humidity sensor, and a sound sensor;

[0102] The pressure sensor is used to monitor the body weight distribution (R a4 ) and body movement in bed (R a5 )Condition;

[0103] The temperature and humidity sensor is used to monitor the temperature of the sleeping environment (R a6 ) and humidity (R a7 );

[0104] The sound sensor is used to detect the sound level during sleep (R a8 ).

[0105] Furthermore, in step S204, the weight of controlling the local retention projection in the loss function is set by dynamic adjustment, and the calculation method is expressed as:

[0106]

[0107] In formula (11), is the weight of the control local projection in the loss function after this iteration update, α c is the learning rate, α c is 0.01, tanh() is the hyperbolic tangent function, acc c is the discrimination accuracy of the discriminator of the current iteration, γ c is the target accuracy, γ c is 0.85, indicating that the discriminator is expected to have at least 85% accuracy in distinguishing between real and generated data, β c is the coefficient that adjusts the sensitivity of the reaction.

[0108] Furthermore, the calculation method of the change in weight and bias in step S302 is expressed as:

[0109]

[0110]

[0111] In formulas (16) and (17), θ p is the quantum phase angle, which controls the balance between randomness and determinism in the search direction, θ p is π / 4, It means the mean is 0 and the variance is The normal distribution of p is a parameter that controls the intensity of random perturbations.

[0112] Furthermore, the probability calculation method of the crossover in step S304 is expressed as:

[0113]

[0114] In formula (21), To select weight and The probability of performing a crossover;

[0115] for and The Euclidean distance between p The parameter selected for control, γ p The control selection parameter is 0.1, which can prompt the algorithm to choose individuals with closer distances for crossover to maintain the diversity of the population.

[0116] Furthermore, in step S403, the KL divergence is calculated by distribution comparison, which is expressed as:

[0117]

[0118] In formula (27), σ rj and μ rj are the encoded latent variables z r The logarithm of the standard deviation and the square of the mean of the j-th dimension, Mg is the dimension of the latent variable after encoding.

[0119] Furthermore, the similarity A in step S503 jk The calculation method is expressed as:

[0120]

[0121] In formula (34), x j is the jth data point, x k is the kth data point;

[0122] Substituting into the manifold regularization term, the calculation method is expressed as:

[0123]

[0124] In formula (35), y uj and uk x j and x k In the output after network processing, Lg is the dimension of the output layer, Lg is 0.3, y ujl and ukl They are x j and x k After the output of the lth neuron in the output layer, σ gs is the Gaussian kernel width parameter used to adjust the similarity perception range.

[0125] Furthermore, the calculation method of the fractional derivative of the α-order loss function with respect to the weight parameter in step S504 is expressed as:

[0126]

[0127] In formula (38), α is the order of the fractional derivative, h is the step size, is the generalized binomial coefficient, int(k) is the kth integer;

[0128] The learning rate dynamically adjusted in step S504 is expressed as:

[0129]

[0130] In formula (39), η t is the learning rate of the fractional-order neural network in the th iteration, η0 is the initial learning rate of the fractional-order neural network, η0 is 0.1, int(t) represents the current number of iterations, ∈ hf is a very small number, ∈ hf is 0.0001.

[0131] The beneficial effects of the present invention are as follows: (1) The present invention adopts a conditional generative adversarial network based on local preserving projection, and improves the quality and relevance of data generation by dynamically adjusting the loss function and considering the preservation of the local structure of the data. (2) The present invention adopts a neural network based on the quantum phase change whale optimization algorithm for feature extraction, and uses the quantum phase change theory to guide the optimization process to improve the search efficiency and accuracy. (3) The present invention adopts a variational autoencoder based on nonlinear hashing for feature dimensionality reduction, optimizes the independence between features through a deep feature decoupling strategy, and enhances the model's learning ability for feature structure. (4) The present invention adopts a fractional-order neural network based on heuristic convergence, combined with fractional-order calculus theory and manifold regularization terms, to effectively improve the accuracy and generalization ability of the classification model. (5) The sleep quality monitoring method of the present invention can effectively process newly input multi-source sleep data, and perform real-time classification and evaluation, which not only improves the accuracy and efficiency of classification, but also provides users with a scientific basis for sleep quality, which is conducive to personalized sleep quality management and improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0132] Figure 1 It is a principle block diagram of an embodiment of the sleep quality assessment method based on millimeter wave radar and AI monitoring sensor of the present invention. DETAILED DESCRIPTION

[0133] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0134] like Figure 1 As shown, the sleep quality assessment method based on millimeter wave radar and AI monitoring sensor of this embodiment includes the following steps:

[0135] S1. Data collection and annotation

[0136] Data is collected and manually labeled through millimeter-wave radar and AI monitoring sensor modules. The labeled data types include high-quality sleep, general sleep, and poor-quality sleep to monitor the user's sleep status.

[0137] Millimeter wave radar is used to detect the user's physiological parameters, including heart rate (R a1 ), respiratory rate (R a2 ) and body movement (R a3 )wait.

[0138] The AI ​​monitoring sensor module includes a pressure sensor, a temperature and humidity sensor, and a sound sensor. The pressure sensor is used to monitor weight distribution (R a4 ) and body movement in bed (R a5 ) situation, the temperature and humidity sensor is used to monitor the temperature of the sleeping environment (R a6 ) and humidity (R a7 ), monitoring the temperature and humidity of the sleeping environment, which may affect the quality of sleep, and the sound sensor is used to detect the sound level during sleep (R a8 ), such as snoring and other noises that may disrupt sleep.

[0139] S2. Data expansion

[0140] The collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model.

[0141] This embodiment uses a conditional generative adversarial network based on local preserving projection to generate samples, thereby achieving data expansion. Based on the traditional conditional generative adversarial network, the present invention uses a dynamically adjusted loss function to measure the difference between real data and generated data, while considering the preservation of the local structure of the data to achieve more refined data generation control.

[0142] The training process of the conditional generative adversarial network based on local preservation projection is as follows:

[0143] S201, set initial parameters for the generator and discriminator of the conditional generative adversarial network, the parameters include the weight and bias of the network, and set the generator to G c , the discriminator is D c , the weight of the generator is The weight of the discriminator is The bias of the generator is The bias of the discriminator is Initialize the bias of the generator and discriminator to 0, and initialize the weights of the generator and discriminator as follows:

[0144]

[0145] In formula (1), is the initial weight, n c is the number of nodes in the input layer of the generator, n c is 10, is a normal distribution with a mean of 0 and a standard deviation of 1.

[0146] S202, setting corresponding conditional variables according to the data type to be generated, the data types include high-quality sleep, general sleep, and poor-quality sleep, and taking the conditional variables and the random noise vector as inputs of the generator, expressed as:

[0147]

[0148] y c =f c (Cla) (3)

[0149] In formula (2), z c is the random noise generated by a random process, σ c is the standard deviation of the noise, σ c is 1.0,∈ c is unit Gaussian noise.

[0150] In formula (3), y c is the conditional variable mapped according to the sleep data type, Cla is the monitoring indicator corresponding to different data categories, and f c (Cla) is a mapping function of monitoring indicators corresponding to different data categories. The mapping function according to the data category is a single hidden layer neural network, which performs nonlinear mapping according to samples of different categories.

[0151] S203. During the training process of the generator and the discriminator, the local preservation projection of each batch of data is calculated to ensure that the generated data maintains the local structural features of the original data. The calculation formula of the local preservation projection is:

[0152]

[0153] M c =IW c (5)

[0154]

[0155]

[0156]

[0157] In formula (4), L c For local preservation of projection, X c is the input data matrix, L c () is the local preservation projection calculation function, For X c The transpose of M c is the locality preserving matrix.

[0158] In formula (5), I is the unit matrix, W c is the weight matrix.

[0159] In formulas (6), (7), and (8), W c (i,j) is the element in the i-th row and j-th column of the weight matrix,

[0160] For data points and The square of the Euclidean distance between c is the adjustment parameter, t c is 0.5, For point The local density of For point The local density of is the neighbor set of the ith data point, For data points and The square of the Euclidean distance between c is the standard deviation parameter.

[0161] S204, perform adversarial training iterations, alternately train the generator and the discriminator, the generator generates data close to the real data to deceive the discriminator, and the discriminator tries to distinguish the real data from the generated data. The loss function calculation method of adversarial training is expressed as:

[0162]

[0163]

[0164] In formula (9), is the loss function of the generator, For the expected operation, D c () is the discriminator function, G c () is the generator function, λ c To control the weight of the local preservation projection in the loss function, η c is the regularization coefficient, η c is 0.001, Weight The square of the L2 norm of .

[0165] In formula (10), is the loss function of the discriminator.

[0166] Among them, the weight controlling the local preservation projection in the loss function is set by dynamic adjustment, and the calculation method is expressed as:

[0167]

[0168] In formula (11), is the weight of the control local projection in the loss function after this iteration update, α c is the learning rate, α c is 0.01, tanh() is the hyperbolic tangent function, acc c is the discrimination accuracy of the discriminator of the current iteration, γ c is the target accuracy, γ c is 0.85, indicating that the discriminator is expected to have at least 85% accuracy in distinguishing between real and generated data, β c is the coefficient that adjusts the sensitivity of the reaction.

[0169] S205, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0170] S3. Training feature extraction model

[0171] The expanded data is input into the feature extraction model for training the feature extraction model. This embodiment adopts a neural network algorithm based on quantum phase change whale optimization as the feature extraction model. In the traditional whale optimization algorithm, the search process relies on random or empirical strategies.

[0172] Based on the whale optimization algorithm, this embodiment uses the search capability of the whale optimization algorithm to optimize the weights and biases of the neural network by simulating the quantum phase change process, thereby achieving efficient feature extraction and improving the convergence speed of the feature extraction model.

[0173] S301, initialize the parameters of the neural network, including the weight and bias of the neural network, and the initialization method is expressed as:

[0174]

[0175]

[0176] In formula (12), W p is the weight of the neural network, n p is the number of nodes in the previous layer of neural network,

[0177] is a normal distribution with a mean of 0 and a standard deviation of 1.

[0178] In formula (13), b p For bias.

[0179] S302. In the whale optimization algorithm, the position of the individual whale represents a set of parameters of the neural network. Each individual whale updates its position according to the quantum phase change theory, simulating the behavior of quantum particles in the potential energy field to explore the solution space. The parameter update method is expressed as:

[0180]

[0181]

[0182] In formulas (14) and (15), and are the weights and biases of the updated neural network, α p is the search step length, α p is 0.5, and are the weights and biases of the neural network with the best performance in the current search, λ p is the attenuation coefficient, λ p is 0.95, int(t) is the current iteration number, ΔW p and Δb p are the changes in weight and bias respectively.

[0183] The calculation method of the change in weight and bias is expressed as:

[0184]

[0185]

[0186] In formulas (16) and (17), θ p is the quantum phase angle, which controls the balance between randomness and determinism in the search direction, θ p is π / 4, It means the mean is 0 and the variance is The normal distribution of p is a parameter that controls the intensity of random perturbations.

[0187] S303, perform fitness evaluation of individual whales, use the current position of individual whales to evaluate the performance of the neural network model in the feature extraction task, and the fitness evaluation is calculated by the loss function, and the calculation method is expressed as:

[0188]

[0189] In formula (18), F p represents the fitness value, MSE() represents the mean square error, Y is the true label, is the predicted label, n dp is the number of neurons in the output layer, ω piis the weight of the i-th neuron in the output layer, Y i and are the true value and predicted value of the i-th output respectively.

[0190] S304: Select and crossover whale individuals. According to the fitness results, select the best performing whale individuals and generate a new generation of whale individuals through quantum crossover operations to enhance the diversity and exploration ability of the model. In the process of updating weights and biases, the method of selecting whale individuals is expressed as follows:

[0191]

[0192]

[0193] In formula (19), and are two sets of weights randomly selected from the current population,

[0194] Generates a [0, ],

[0195] Is the selection weight and The probability of performing a crossover.

[0196] In formula (20), and is the corresponding bias.

[0197] The probability model of crossover operation is used to accurately control the crossover behavior in the genetic algorithm. The probability calculation method of crossover is expressed as:

[0198]

[0199] In formula (21), To select weight and The probability of performing a crossover.

[0200] for and The Euclidean distance between p The parameter selected for control, γ p The control selection parameter is 0.1, which can prompt the algorithm to choose individuals with closer distances for crossover to maintain the diversity of the population.

[0201] S305, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1200 times.

[0202] S4. Training feature dimensionality reduction model

[0203] The data after feature extraction is input into the feature dimensionality reduction model for training. The variational autoencoder based on nonlinear hashing is used for feature dimensionality reduction. Nonlinear hashing technology and deep feature decoupling strategy are used to map high-dimensional data to low-dimensional hash space through hash functions. Deep feature decoupling optimizes the independence of each feature through the framework of variational autoencoder, and enhances the model's ability to learn the intrinsic structure of features.

[0204] The training process of the variational autoencoder algorithm based on nonlinear hashing is as follows:

[0205] S401, initialize the network parameters of the variational autoencoder, and the initialization method is expressed as:

[0206]

[0207]

[0208] In formula (22), W r is the weight of the variational autoencoder, n r is the number of nodes in the input layer of the variational autoencoder encoder, It means the mean is 0 and the variance is The normal distribution of is the variance of the initialized Gaussian distribution, is 0.01.

[0209] In formula (23), b r is the bias of the variational autoencoder.

[0210] S402, the data after feature extraction is forward propagated through the encoder layer, where each layer is processed by a nonlinear activation function and finally reaches the hash layer, which maps the data to a low-dimensional space. The calculation method of the mapping function is expressed as:

[0211] z r =tanh(W r ·x r +b r ) (twenty four)

[0212] H r =sign(z r ) (25)

[0213] In formula (24), z r is the latent variable, is the output of the tanh activation function, tanh() is the tanh activation function, x r is the data input to the hash layer.

[0214] In formula (25), H r To map features to low-dimensional space, sign() is a sign function that converts continuous activation values ​​to {-1, 1}.

[0215] S403. In the decoding stage, the difference between the reconstructed data and the original data is quantified by the reconstruction loss. In order to maintain the diversity of low-dimensional representation, KL divergence is used as a regularization term to measure the difference between the encoded potential distribution and the prior distribution. The calculation method of the loss function of the variational autoencoder is expressed as:

[0216]

[0217] In formula (26), L r is the loss function of the variational autoencoder, is the output reconstructed by the decoder, x r is the data input to the variational self-encoder, is the data after feature extraction, and β r is the regularization coefficient, β r is 0.3, D KL () is the KL divergence calculation function, which is used to quantize the potential distribution q(z r |x r ) and the prior distribution p(z r ) between the two.

[0218] KL divergence is calculated by distribution comparison, which is expressed as:

[0219]

[0220] In formula (27), σ rj and μ rj are the encoded latent variables z r The logarithm of the standard deviation and the square of the mean of the j-th dimension, Mg is the dimension of the latent variable after encoding.

[0221] S404, using the gradient descent method to update the network parameters of the variational autoencoder, the parameter update method is expressed as:

[0222]

[0223]

[0224] In formulas (28) and (29), and are the updated weights and biases of the variational autoencoder, η r is the learning rate of the variational autoencoder, η r is 0.01, and is the loss function L r Gradients with respect to weights and biases.

[0225] S405, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0226] S5. Training classifier model

[0227] The data after feature dimensionality reduction is input into the classifier to train the classifier model. A fractional-order neural network based on heuristic convergence is adopted as the classification algorithm. The fractional-order calculus theory is used to adjust the propagation and update rules in the neural network, so as to better capture the long-range dependence and intrinsic dynamic characteristics in the data. The present invention adopts a manifold regularization term so that the learned features not only reflect the distribution of the data, but also retain the topological structure of the data itself, thereby improving the classification accuracy.

[0228] The training process of the fractional-order neural network algorithm based on heuristic convergence is as follows:

[0229] S501, initializing the parameters of the fractional-order neural network, the initialization method is expressed as:

[0230]

[0231] b u,init =0 (31)

[0232] In formula (30), W u,init is the initialization value of the weight of the fractional-order neural network, n u-1 and n u are the number of nodes in the previous layer and the current layer respectively, ∈ u are random numbers drawn from a standard normal distribution.

[0233] In formula (31), b u,init is the initialization value of the bias of the fractional-order neural network.

[0234] S502, input the feature vector after feature dimensionality reduction into the fractional-order neural network, process it through fractional-order derivatives and nonlinear activation functions, and pass it layer by layer until the output layer of the fractional-order neural network. The number of neurons in the output layer of the fractional-order neural network is equal to the number of sleep quality categories, and each neuron corresponds to a category. In this embodiment, the categories include high-quality sleep, general sleep, and poor-quality sleep, and the number of neurons in the output layer is 3.

[0235] For each layer of the fractional-order neural network except the output layer, the calculation method of its output is expressed as:

[0236]

[0237] In formula (32), y u is the layer output, Sig() is the Sigmoid activation function, W ui is the weight of the i-th neuron in the current layer, x ui is the input from the previous layer, b u is the bias of the current layer.

[0238] S503, calculating the loss function of the fractional-order neural network, using a loss function based on a manifold regularization term, the manifold regularization term is used to keep data points of the same type close to each other in the mapped low-dimensional space, and to separate data points of different types as much as possible to ensure the smoothness of the classification boundary. The calculation method of the loss function of the fractional-order neural network is expressed as:

[0239]

[0240] In formula (33), L u is the loss function of the fractional neural network, Mr is the number of samples input in the current batch, and t ui is the target label, y ui is the network prediction output, λ u is the regularization parameter, A jk is the similarity between the j-th data point and the k-th data point.

[0241] Similarity A jk The calculation method is expressed as:

[0242]

[0243] In formula (34), x j is the jth data point, x k is the kth data point.

[0244] Substituting formula (34) into the manifold regularization term, the calculation method is expressed as:

[0245]

[0246] In formula (35), y uj and uk x j and x k In the output after network processing, Lg is the dimension of the output layer, Lg is 0.3, y ujl and ukl They are x j and x k After the output of the lth neuron in the output layer, σ gs is the Gaussian kernel width parameter used to adjust the similarity perception range.

[0247] S504, using the fractional calculus strategy to back-propagate the loss function, calculate the gradient of each parameter, and when updating the parameters, use a dynamically adjusted learning rate to adapt to the changes in data during training and the convergence speed of the algorithm. The calculation method of the update amount of the fractional parameter is expressed as:

[0248]

[0249]

[0250] In formulas (36) and (37), ΔW ui and Δb u are the update amounts of the weights and biases of the fractional-order neural network in the current iteration, η gdf is the learning rate of the fractional-order neural network, is the fractional derivative of the α-order loss function with respect to the weight parameter, γ gdf is the parameter used to adjust the update amount of the previous iteration, is the update amount of the weight of the fractional-order neural network in the previous iteration.

[0251] The calculation method of the fractional derivative of the α-order loss function with respect to the weight parameter is expressed as:

[0252]

[0253] In formula (38), α is the order of the fractional derivative, h is the step size, is the generalized binomial coefficient and int(k) is the kth integer.

[0254] The dynamically adjusted learning rate is expressed as:

[0255]

[0256] In formula (39), η tis the learning rate of the fractional-order neural network in the th iteration, η0 is the initial learning rate of the fractional-order neural network, η0 is 0.1, int(t) represents the current number of iterations, ∈ hf is a very small number, ∈ hf is 0.0001.

[0257] S505, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0258] S6. Sleep quality monitoring

[0259] The new sleep monitoring data is processed using the trained feature extraction model, data dimensionality reduction model and classifier model to achieve real-time monitoring and evaluation of sleep quality. The collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing, and the processed features are input into the classifier model for classifier training to obtain the classification results.

[0260] The sleep quality monitoring method of the present invention can effectively process newly input multi-source sleep data and perform real-time classification and evaluation, which not only improves the accuracy and efficiency of classification, but also provides users with a scientific basis for sleep quality, which is conducive to personalized sleep quality management and improvement.

[0261] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A sleep quality assessment method based on millimeter wave radar and AI monitoring sensor, characterized in that: The following steps are involved: S1. Data collection and annotation The millimeter-wave radar and AI monitoring sensor modules collect and manually annotate data, and the annotated data types include high-quality sleep, general sleep, and poor-quality sleep to monitor the user's sleep status; S2. Data expansion Sample generation based on conditional generative adversarial networks based on local preservation projections to achieve data expansion; The training process of the conditional generative adversarial network based on local preservation projection is as follows: S201, set initial parameters for the generator and discriminator of the conditional generative adversarial network, the parameters include the weight and bias of the network, and set the generator to be G c , the discriminator is D c , the weight of the generator is The weight of the discriminator is The bias of the generator is The bias of the discriminator is Initialize the bias of the generator and discriminator to 0, and initialize the weights of the generator and discriminator as follows: In formula (1), is the initial weight, n c is the number of nodes in the input layer of the generator, n c is 10, N(0,1) is a normal distribution with a mean of 0 and a standard deviation of 1; S202, setting corresponding conditional variables according to the data type to be generated, and taking the conditional variables and the random noise vector as inputs of the generator, expressed as: y c =f c (Cla) (3) In formula (2), z c is the random noise generated by a random process, σ c is the standard deviation of the noise, σ c is 1.0,∈ c is unit Gaussian noise; In formula (3), y c is the conditional variable mapped according to the sleep data type, Cla is the monitoring indicator corresponding to different data categories, and f c (Cla) is a mapping function of monitoring indicators corresponding to different data categories; the mapping function according to the data category is a single hidden layer neural network, which performs nonlinear mapping according to samples of different categories; S203. During the training process of the generator and the discriminator, the local preservation projection of each batch of data is calculated to ensure that the generated data maintains the local structural features of the original data. The calculation formula of the local preservation projection is: M c =I-W c (5) In formula (4), L c For local preservation of projection, X c is the input data matrix, L c () is the local preservation projection calculation function, For X c The transpose of M c is the locality preserving matrix; In formula (5), I is the unit matrix, W c is the weight matrix; In formulas (6), (7), (8), W c (i,j) is the element in the i-th row and j-th column of the weight matrix, For data points and The square of the Euclidean distance between c is the adjustment parameter, t c is 0.5, ρ ci For point The local density, ρ cj For point The local density of is the neighbor set of the ith data point, For data points and The square of the Euclidean distance between c is the standard deviation parameter; S204, perform adversarial training iterations, alternately train the generator and the discriminator, the generator generates data close to the real data to deceive the discriminator, and the discriminator tries to distinguish the real data from the generated data. The loss function calculation method of adversarial training is expressed as: In formula (9), is the loss function of the generator, For the expected operation, D c () is the discriminator function, G c () is the generator function, λ c To control the weight of the local preservation projection in the loss function, η c is the regularization coefficient, η c is 0.001, Weight The square of the L2 norm of ; In formula (10), is the loss function of the discriminator; S205, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed; S3. Training feature extraction model Based on the whale optimization algorithm, the search capability of the whale optimization algorithm is used to optimize the weights and biases of the neural network by simulating the quantum phase change process, so as to achieve efficient feature extraction and improve the convergence speed of the feature extraction model. S301, initialize the parameters of the neural network, including the weight and bias of the neural network, and the initialization method is expressed as: In formula (12), W p is the weight of the neural network, n p is the number of nodes in the previous layer of neural network, is a normal distribution with a mean of 0 and a standard deviation of 1; In formula (13), b p is bias; S302. In the whale optimization algorithm, the position of the individual whale represents a set of parameters of the neural network. Each individual whale updates its position according to the quantum phase change theory, simulating the behavior of quantum particles in the potential energy field to explore the solution space. The parameter update method is expressed as: In formulas (14) and (15), and are the weights and biases of the updated neural network, α p is the search step length, α p is 0.5, and are the weights and biases of the neural network with the best performance in the current search, λ p is the attenuation coefficient, λ p is 0.95, int(t) is the current iteration number, ΔW p and Δb p are the changes in weight and bias respectively; S303, perform fitness evaluation of individual whales, use the current position of individual whales to evaluate the performance of the neural network model in the feature extraction task, and the fitness evaluation is calculated by the loss function, and the calculation method is expressed as: In formula (18), F p represents the fitness value, MSE() represents the mean square error, Y is the true label, is the predicted label, n dp is the number of neurons in the output layer, ω pi is the weight of the i-th neuron in the output layer, Y i and are the true value and predicted value of the i-th output respectively; S304: Select and crossover whale individuals. According to the fitness results, select the best performing whale individuals and generate a new generation of whale individuals through quantum crossover operations to enhance the diversity and exploration ability of the model. In the process of updating weights and biases, the method of selecting whale individuals is expressed as follows: In formula (19), and are two sets of weights randomly selected from the current population, Generates a [0, ], Is the selection weight and The probability of performing a crossover; In formula (20), and is the corresponding bias; S305, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed; S4. Training feature dimensionality reduction model The data after feature extraction is input into the feature dimensionality reduction model for training. The variational autoencoder based on nonlinear hashing is used for feature dimensionality reduction. Nonlinear hashing technology and deep feature decoupling strategy are used to map high-dimensional data to low-dimensional hash space through hash functions. Deep feature decoupling optimizes the independence between various features through the variational autoencoder framework, and enhances the model's ability to learn the intrinsic structure of features. The training process of the variational autoencoder algorithm based on nonlinear hashing is as follows: S401, initialize the network parameters of the variational autoencoder, and the initialization method is expressed as: In formula (22), W r is the weight of the variational autoencoder, n r is the number of nodes in the input layer of the variational autoencoder encoder, It means the mean is 0 and the variance is The normal distribution of is the variance of the initialized Gaussian distribution, is 0.01; In formula (23), b r is the bias of the variational autoencoder; S402, the data after feature extraction is forward propagated through the encoder layer, where each layer is processed by a nonlinear activation function and finally reaches the hash layer, which maps the data to a low-dimensional space. The calculation method of the mapping function is expressed as: z r =tanh(W r ·x r +b r ) (24) H r =sign(z r ) (25) In formula (24), z r is the latent variable, is the output of the tanh activation function, tanh() is the tanh activation function, x r is the data input to the hash layer; In formula (25), H r To map features to low-dimensional space, sign() is a sign function that converts continuous activation values ​​to {-1, 1}; S403. In the decoding stage, the difference between the reconstructed data and the original data is quantified by the reconstruction loss. In order to maintain the diversity of low-dimensional representation, KL divergence is used as a regularization term to measure the difference between the encoded potential distribution and the prior distribution. The calculation method of the loss function of the variational autoencoder is expressed as: In formula (26), L r is the loss function of the variational autoencoder, is the output reconstructed by the decoder, x r is the data input to the variational self-encoder, is the data after feature extraction, and β r is the regularization coefficient, β r is 0.3, D KL () is the KL divergence calculation function, which is used to quantize the potential distribution q(z r |x r ) and the prior distribution p(z r ) S404, using the gradient descent method to update the network parameters of the variational autoencoder, the parameter update method is expressed as: In formulas (28) and (29), and are the updated weights and biases of the variational autoencoder, η r is the learning rate of the variational autoencoder, η r is 0.01, and is the loss function L r Gradients with respect to weights and biases; S405, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed; S5. Training classifier model The data after feature dimension reduction is input into the classifier to train the classifier model. The fractional-order neural network based on heuristic convergence is used as the classification algorithm. The fractional-order calculus theory is used to adjust the propagation and update rules in the neural network to better capture the long-range dependence and intrinsic dynamic characteristics in the data. The training process of the fractional-order neural network algorithm based on heuristic convergence is as follows: S501, initializing the parameters of the fractional-order neural network, the initialization method is expressed as: b u,init ＝0 (31) In formula (30), W u,init is the initialization value of the weight of the fractional-order neural network, n u-1 and n u are the number of nodes in the previous layer and the current layer respectively, ∈ u is a random number drawn from a standard normal distribution; In formula (31), b u,init is the initialization value of the bias of the fractional-order neural network; S502, input the feature vector after feature dimension reduction into the fractional-order neural network, process it through the fractional-order derivative and the nonlinear activation function, and pass it layer by layer until the output layer of the fractional-order neural network. The number of neurons in the output layer of the fractional-order neural network is equal to the number of sleep quality categories, and each neuron corresponds to one category. For each layer of the fractional-order neural network except the output layer, the calculation method of its output is expressed as: In formula (32), y u is the layer output, Sig() is the Sigmoid activation function, W ui is the weight of the i-th neuron in the current layer, x ui is the input from the previous layer, b u is the bias of the current layer; S503, calculating the loss function of the fractional-order neural network, using a loss function based on a manifold regularization term, the manifold regularization term is used to keep data points of the same type close to each other in the mapped low-dimensional space, and to separate data points of different types as much as possible to ensure the smoothness of the classification boundary. The calculation method of the loss function of the fractional-order neural network is expressed as: In formula (33), L u is the loss function of the fractional-order neural network, Mr is the number of samples input in the current batch, and t ui is the target label, y ui is the network prediction output, λ u is the regularization parameter, A jk is the similarity between the jth data point and the kth data point; S504, using the fractional calculus strategy to back-propagate the loss function, calculate the gradient of each parameter, and when updating the parameters, use a dynamically adjusted learning rate to adapt to the changes in data during training and the convergence speed of the algorithm. The calculation method of the update amount of the fractional parameter is expressed as: In formulas (36) and (37), ΔW ui and Δb u are the update amounts of the weights and biases of the fractional-order neural network in the current iteration, η gdf is the learning rate of the fractional-order neural network, is the fractional derivative of the α-order loss function with respect to the weight parameter, γ gdf is the parameter used to adjust the update amount of the previous iteration, is the update amount of the weight of the fractional-order neural network in the previous iteration; S505, repeating the above steps until the preset stop iteration condition is met, indicating that the model training is completed; S6. Sleep quality monitoring The new sleep monitoring data is processed using the trained feature extraction model, data dimensionality reduction model and classifier model to achieve real-time monitoring and evaluation of sleep quality. The collected original data is input into the trained feature extraction and feature dimensionality reduction models for feature processing, and the processed features are input into the classifier model for classifier training to obtain the classification results.

2. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1 is characterized in that: The AI ​​monitoring sensor module in step S1 includes a pressure sensor, a temperature and humidity sensor, and a sound sensor; The pressure sensor is used to monitor the body weight distribution (R a4 ) and body movement in bed (R a5 )Condition; The temperature and humidity sensor is used to monitor the temperature of the sleeping environment (R a6 ) and humidity (R a7 ); The sound sensor is used to detect the sound level during sleep (R a8 ).

3. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: In step S204, the weight of controlling the local retention projection in the loss function is set by dynamic adjustment, and the calculation method is expressed as: In formula (11), is the weight of the control local projection in the loss function after this iteration update, α c is the learning rate, α c is 0.01, tanh() is the hyperbolic tangent function, acc c is the discrimination accuracy of the discriminator of the current iteration, γ c is the target accuracy, γ c is 0.85, indicating that the discriminator is expected to have at least 85% accuracy in distinguishing between real and generated data, β c is the coefficient that adjusts the sensitivity of the reaction.

4. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: The calculation method of the change in weight and bias in step S302 is expressed as: In formulas (16) and (17), θ p is the quantum phase angle, which controls the balance between randomness and determinism in the search direction, θ p is π / 4, It means the mean is 0 and the variance is The normal distribution of p is a parameter that controls the intensity of random perturbations.

5. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: The calculation method of the crossover probability in step S304 is expressed as: In formula (21), To select weight and The probability of performing a crossover; for and The Euclidean distance between p The parameter selected for control, γ p The control selection parameter is 0.1, which can prompt the algorithm to choose individuals with closer distances for crossover to maintain the diversity of the population.

6. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: In step S403, the KL divergence is calculated by distribution comparison, which is expressed as: In formula (27), σ rj and μ rj are the encoded latent variables z r The logarithm of the standard deviation and the square of the mean of the j-th dimension, Mg is the dimension of the latent variable after encoding.

7. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: The similarity A in step S503 jk The calculation method is expressed as: In formula (34), x j is the jth data point, x k is the kth data point; Substituting into the manifold regularization term, the calculation method is expressed as: In formula (35), y uj and uk x j and x k In the output after network processing, Lg is the dimension of the output layer, Lg is 0.3, y ujl and ukl They are x j and x k After the output of the lth neuron in the output layer, σ gs is the Gaussian kernel width parameter used to adjust the similarity perception range.

8. The sleep quality assessment method based on millimeter wave radar and AI monitoring sensor according to claim 1, characterized in that: The calculation method of the fractional derivative of the α-order loss function with respect to the weight parameter in step S504 is expressed as: In formula (38), α is the order of the fractional derivative, h is the step size, is the generalized binomial coefficient, int(k) is the kth integer; The learning rate dynamically adjusted in step S504 is expressed as: In formula (39), η t is the learning rate of the fractional-order neural network in the th iteration, η0 is the initial learning rate of the fractional-order neural network, η0 is 0.1, int(t) represents the current number of iterations, ∈ hf is a very small number, ∈ hf is 0.0001.

Images