Non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing

Abstract

The application belongs to the technical field of sleep regulation, and specifically relates to a non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing, which comprises the following steps: an edge node collects physiological signals through localized sleep stage cooperative identification, extracts time sequence features to obtain a physiological state vector, performs high-order spectrum analysis and heart-lung coupling calculation, combines body movement and posture information to generate a high-precision sleep state digital matrix, solves head posture deviation based on the sleep matrix, optimizes the phase and amplitude of the coil, generates individualized electromagnetic field waveforms through sub-threshold chaotic modulation, performs adaptive compensation in combination with environmental and metal interference, corrects the driving current, and outputs sleep abnormality risk control data, fuses multiple information iteration regulation strategies, performs intervention, collects physiological feedback to optimize parameters in real time, and forms a full-link edge autonomous double-layer closed-loop sleep regulation. In the application, multiple physiological signals are processed locally through edge computing, sleep is accurately identified, individualized electromagnetic fields are generated, and double-layer adaptive regulation is achieved.

Description

Technical Field

[0001] This invention belongs to the field of sleep regulation technology, specifically a non-head-mounted pulsed electromagnetic field intelligent sleep regulation system based on edge computing. Background Technology

[0002] The intelligent sleep regulation system adopts a non-head-mounted design, which collects sleep physiological data imperceptibly through multimodal sensors, and analyzes and generates adaptive pulse electromagnetic field parameters in real time at the edge.

[0003] Traditional control systems rely on cloud computing and lack local real-time processing capabilities. All physiological signals must be uploaded to the cloud for processing; sleep staging, strategy generation, and parameter adjustment are all completed remotely, with the local system only responsible for data collection and execution. During use, network latency, packet loss, or congestion can cause changes in sleep state but delays in control commands, and electromagnetic field parameters cannot match deep sleep, light sleep, and posture changes in real time. Interference cannot be compensated for locally in real time, waveform distortion persists, intervention is inaccurate, it easily triggers micro-awakening, and it is entirely passive in response, unable to adapt autonomously, seriously affecting the sleep improvement effect and the stability of use. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention proposes a non-head-mounted pulsed electromagnetic field intelligent sleep regulation system based on edge computing. This invention primarily addresses issues such as cloud latency, interference distortion, and ineffective intervention.

[0005] The non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing provided by the present invention includes: a signal acquisition module, used to acquire raw physiological signals through localized scene sleep stage collaborative identification based on edge nodes, extract temporal features from the raw physiological signals, and output a continuous physiological state vector.

[0006] The sleep modeling module is used to perform high-order spectral analysis on continuous physiological state vectors, calculate the cardiopulmonary coupling phase relationship and energy distribution characteristics, combine body movement and head posture information to obtain high-precision sleep parameters, and output a real-time sleep state digital matrix.

[0007] The waveform generation module is used to calculate the head's spatial position and posture offset based on the real-time sleep state digital matrix, dynamically optimize the phase and amplitude of multiple excitation coils, and generate a non-repeating individualized electromagnetic field excitation waveform using subthreshold random chaotic modulation.

[0008] The interference compensation module is used to perform adaptive compensation calculations based on the non-repeating individualized electromagnetic field excitation waveform, environmental magnetic field distortion, and metal interference data. It calculates the interference cancellation coefficient to obtain the corrected coil drive current, performs multi-source data processing and sleep assessment, and outputs sleep abnormality risk control data.

[0009] The closed-loop control module is used to integrate sleep abnormality risk control data with sleep difference status, equipment status and power grid load information, iterate pulse electromagnetic field control strategies and issue execution interventions, collect feedback physiological signals to optimize parameters, and output the full-link edge autonomous two-layer closed-loop sleep control process.

[0010] The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system provided by the present invention includes a signal acquisition module comprising: The multi-source acquisition unit is used to acquire environmental information, millimeter-wave radar and piezoelectric sensor raw physiological signals at the edge node, forming multi-source time-series raw data.

[0011] The sleep recognition unit is used to identify localized sleep stages by using multi-source raw data through time-series correlation analysis and output the current sleep stage identifier.

[0012] The adaptive sampling unit is used to dynamically adjust the sampling frequency, signal gain and sampling weight according to the sleep stage identifier, and adaptively sample the original physiological signal to obtain a regular signal sequence with optimized signal-to-noise ratio.

[0013] The signal purification unit is used to perform adaptive noise reduction, bandpass filtering and artifact removal on regular signal sequences to separate clean signal components such as heart rate, respiration, body movement and posture.

[0014] The feature fusion unit is used to extract temporal features from pure signal components, normalize and fuse them to generate continuous high-dimensional temporal features, and construct and output physiological state vectors.

[0015] The sleep modeling module of the non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing provided by the present invention includes: Higher-order spectral units are used to estimate the third-order cumulants of continuous physiological state vectors, remove Gaussian noise and linear interference, and obtain the higher-order spectral density matrix.

[0016] The cardiopulmonary coupling unit is used to extract the bispectral phase and amplitude features of the cardiopulmonary signal based on the high-order spectral density matrix, calculate the phase synchronization index and cross-spectral energy ratio, and form a cardiopulmonary coupling feature vector.

[0017] The feature fusion unit is used to perform feature alignment and normalization fusion with the cardiopulmonary coupling feature vector and body motion and head posture temporal data to construct a multi-dimensional physiological feature tensor.

[0018] The sleep reconstruction unit is used to perform windowed weighting and nonlinear mapping on multi-dimensional physiological feature tensors, fit the sleep stage discrimination function, obtain the sleep probability distribution sequence, reconstruct the sleep probability distribution sequence into an equal-dimensional numerical array by time slices, and output the real-time sleep state digital matrix.

[0019] According to the edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent regulation system provided by the present invention, the specific steps for forming the cardiopulmonary coupling feature vector in the cardiopulmonary coupling unit are as follows: By solving the bispectral phase angle and amplitude spectrum value point by point according to the high-order spectral density matrix, the cardiopulmonary main frequency coupling region is located, and non-coupled frequency interference is eliminated to obtain the bispectral phase-amplitude distribution sequence.

[0020] The phase difference and phase coherence value of adjacent frequency points are calculated based on the bispectral phase-amplitude distribution sequence. The phase synchronization duration and consistency degree are statistically analyzed to generate the cardiopulmonary phase synchronization index sequence.

[0021] Based on the phase synchronization index sequence and bispectral amplitude distribution, the energy proportion and energy concentration of each coupling frequency point are calculated, and the cardiopulmonary energy distribution characteristics are obtained after normalization.

[0022] The phase synchronization index sequence and cardiopulmonary energy distribution characteristics are concatenated by dimension and standardized to remove dimensional differences, forming a cardiopulmonary coupling feature vector.

[0023] The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system provided by the present invention includes a waveform generation module comprising: The attitude calculation unit is used to perform temporal sliding window statistics and spatial feature calculation on the real-time sleep state digital matrix, extract the attitude-sensitive weights corresponding to the sleep stage, and obtain the sleep-related attitude constraint coefficients.

[0024] The posture modeling unit is used to calculate the three-dimensional spatial coordinates and angular offset of the head using sleep-related posture constraint coefficients and real-time posture sampling data, and to generate a head posture offset feature vector.

[0025] The coil optimization unit is used to establish a coil excitation response model based on the head attitude offset feature vector, calculate the optimal phase compensation and amplitude gain of each excitation coil, and obtain multi-coil phase amplitude optimization parameters.

[0026] The waveform generation unit is used to substitute the phase amplitude optimization parameters of the multi-coil into the subthreshold chaotic mapping, perform random phase perturbation and non-periodic iterative modulation, and generate a non-repeating individualized electromagnetic field excitation waveform.

[0027] According to the edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system provided by the present invention, the specific steps for obtaining the multi-coil phase amplitude optimization parameters in the coil optimization unit are as follows: The three-dimensional spatial coordinates and angular offset information of the head are extracted from the head posture offset feature vector. Based on the inherent parameters and installation position of the multi-excitation coil, the correlation mapping between coil excitation and spatial magnetic field distribution is constructed, and a basic coil excitation response model is established.

[0028] By substituting the position and angle data in the head attitude offset feature vector into the basic coil excitation response model, the coil excitation effect under different phases and amplitudes is simulated, the magnetic field distribution error and coupling efficiency are calculated, and the optimization objective function is determined.

[0029] The phase and amplitude of each excitation coil are iteratively solved according to the optimization objective function. The phase compensation amount and amplitude gain are corrected to obtain the preliminary optimal parameters for each coil.

[0030] The initial optimal parameters are processed according to the coil number, and abnormal parameters are eliminated by combining the confidence weight of the head attitude offset feature vector to obtain the multi-coil phase amplitude optimization parameters.

[0031] According to the edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system provided by the present invention, the specific steps for generating a non-repeating individualized electromagnetic field excitation waveform in the waveform generation unit are as follows: Based on the multi-coil phase amplitude optimization parameters, the optimal phase compensation amount and amplitude gain coefficient corresponding to each coil are extracted, the parameters that meet the subthreshold requirements are selected, and the abnormal parameters that exceed the threshold range are eliminated to obtain the basic parameters for subthreshold coil excitation.

[0032] Using the basic parameters of the subthreshold coil excitation as the initial input, the control parameters and initial values ​​of the chaotic mapping are set, and the initial iteration is initialized using the chaotic mapping to obtain the initial sequence of chaotic iteration.

[0033] The phase compensation quantity is randomly perturbed based on the initial sequence of chaotic iteration, and the perturbation intensity is controlled within a preset range to generate a perturbed phase iteration sequence.

[0034] The phase iteration sequence and amplitude gain coefficient are input into the chaotic mapping operation process for non-periodic iterative modulation. The number of iterations is controlled to ensure non-periodicity, and a chaotic excitation parameter sequence is obtained.

[0035] The chaotic excitation parameter sequence is converted into a time-domain waveform signal, and waveform synthesis is performed by combining it with the coil excitation frequency. After smoothing, a unique electromagnetic field excitation waveform without repetition is generated.

[0036] The non-head-mounted pulsed electromagnetic field sleep intelligent control system based on edge computing provided by the present invention includes an interference compensation module comprising: The data preprocessing unit is used to perform synchronous time-series alignment and noise preprocessing on the three types of data, based on the non-repeating individualized electromagnetic field excitation waveform, the collected environmental magnetic field distortion data, and the metal interference data, to eliminate data bias and construct the basic dataset for interference analysis.

[0037] The interference modeling unit is used to extract the characteristic correlation between the non-repeating individualized electromagnetic field excitation waveform and various interference data based on the basic dataset of interference analysis, establish an adaptive compensation calculation model, and calculate the interference cancellation coefficient by comparing and analyzing the amplitude and phase differences between the excitation waveform and the interference signal.

[0038] The current correction unit is used to substitute the interference cancellation coefficient into the initial calculation model of the coil drive current, correct the drive current corresponding to the non-repeating individualized electromagnetic field excitation waveform, and obtain the coil drive current.

[0039] The risk control output unit is used to perform multi-source data fusion processing on coil drive current, non-repetitive individualized electromagnetic field excitation waveform and interference data, correlate with real-time sleep physiological characteristics to carry out sleep assessment, screen sleep abnormality indicators, and output sleep abnormality risk control data.

[0040] According to the edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system provided by the present invention, the specific steps for calculating the interference cancellation coefficient in the interference modeling unit are as follows: The non-repeating individualized electromagnetic field excitation waveform, environmental magnetic field distortion data, and metal interference data in the basic dataset for interference analysis are separated. The frequency domain amplitude, phase characteristics, and time series variation patterns of the three types of data are extracted respectively to determine the characteristic correspondence between the excitation waveform and the interference data.

[0041] An adaptive compensation computation model is constructed based on the feature correspondence to adapt to the non-repeating chaotic excitation waveform.

[0042] The characteristic parameters of the excitation waveform and various interference data are input into the adaptive compensation calculation model to analyze the amplitude difference and phase shift of the excitation waveform and various interference data at different frequency points, and to quantify the degree of influence of interference on the excitation waveform.

[0043] Based on the mapping relationship between the degree of interference and the adaptive compensation calculation model, the interference cancellation coefficient is obtained by iteratively calculating the correction parameters.

[0044] The non-head-mounted pulsed electromagnetic field sleep intelligent control system based on edge computing provided by the present invention includes a closed-loop control module comprising: The data fusion unit is used by edge nodes to perform multi-source data normalization and feature fusion based on sleep abnormality risk control data, sleep difference state quantification data, equipment operating parameters and grid load digital information, and output a standardized control decision input matrix.

[0045] The strategy optimization unit is used to generate a pulse electromagnetic field regulation strategy by iteratively optimizing the frequency, intensity, and timing parameters of the pulse electromagnetic field based on the standardized regulation decision input matrix and historical regulation parameters and sleep intervention effect data through edge AI.

[0046] The intervention feedback unit is used to conduct sleep intervention using electromagnetic fields according to the pulse electromagnetic field modulation strategy, collect feedback physiological signals during the intervention process, perform noise reduction and feature extraction on the signals, and output feedback physiological characteristic parameters.

[0047] The closed-loop correction unit is used to compare the feedback physiological characteristic parameters with the preset sleep regulation target parameters, calculate the deviation value, correct the pulse electromagnetic field regulation parameters through the edge iterative algorithm, update the regulation strategy, and form a full-link edge autonomous double-layer closed-loop sleep regulation data flow process.

[0048] The non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing provided by this invention accurately identifies sleep and generates individualized electromagnetic fields by locally processing multi-source physiological signals through edge computing, thereby achieving dual-layer closed-loop adaptive regulation.

[0049] The beneficial effects of this invention are as follows: 1. This invention deploys all computing units locally at edge nodes to perform real-time acquisition, time-series calibration, adaptive sampling, and signal purification of multi-source time-series data from millimeter-wave radar, piezoelectric sensors, and environmental sensors, without relying on cloud transmission and computation. This enables localized and rapid identification of sleep stages and construction of physiological state vectors. Through local high-order spectral analysis, cardiopulmonary coupling feature extraction, and sleep state modeling, it eliminates network latency, data congestion, and external dependency risks, ensuring millisecond-level response throughout physiological signal processing, sleep stage calculation, and posture shift estimation. By employing local data closed-loop management and a strategy of minimizing information uploads, it avoids the leakage of personal physiological privacy data while reducing device power consumption and hardware costs. Ultimately, it achieves edge autonomy, offline availability, low latency, high privacy, and high stability, providing a secure and reliable underlying support for real-time sleep regulation.

[0050] 2. This invention integrates cardiopulmonary coupling characteristics, body motion data, three-dimensional head posture, and a real-time sleep state matrix to construct a high-dimensional physiological feature tensor and achieve precise sleep staging, significantly improving the reliability of sleep state assessment. Through head posture constraint calculation and dynamic optimization of the phase amplitude of multiple excitation coils, coherent magnetic focusing and stable superposition of the target magnetic field are achieved, ensuring the electromagnetic field accurately acts on the target area. Subthreshold random chaotic modulation generates non-repetitive individualized excitation waveforms, avoiding the human adaptation effect and discomfort caused by fixed waveforms, while ensuring the magnetic field remains within a safe subthreshold range. Adaptive compensation for environmental magnetic field distortion and metallic interference is used to correct the coil drive current in real time, eliminating magnetic field distortion. Ultimately, this results in a highly individualized, comfortable, precise, and non-perceptible sleep intervention effect, significantly improving the proportion of deep sleep and overall sleep quality.

[0051] 3. This invention integrates sleep abnormality risk control, individual differences, equipment status, and power grid load information to construct a standardized decision matrix. It then uses edge AI to iteratively optimize the frequency, intensity, and timing of the electromagnetic field, generating a dynamically adaptable control strategy. By executing interventions and collecting real-time feedback physiological signals such as ECG, respiration, and blood oxygenation, physiological characteristics are extracted and the deviation is calculated against the control target. The control parameters are then corrected in real-time using an edge iterative algorithm. Through an inner-layer intervention, feedback, and correction mechanism, along with an outer-layer data and strategy update dual-closed-loop structure, the entire process achieves self-sensing, self-decision-making, self-execution, and self-optimization. This results in a fully automatic intelligent control effect that requires no manual intervention, adapts to individual changes, resists environmental interference, and maintains stable long-term effects, ensuring continuous optimal sleep intervention. Attached Figure Description

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a block diagram of the non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing provided in an embodiment of the present invention; Figure 2 This is a flowchart of the non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the steps for forming a cardiopulmonary coupling feature vector according to an embodiment of the present invention. Detailed Implementation

[0054] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below according to specific embodiments.

[0055] like Figures 1 to 3 As shown in the embodiment of the present invention, the non-head-mounted pulsed electromagnetic field sleep intelligent control system based on edge computing includes: The signal acquisition module is used to synchronously acquire environmental information and raw physiological signals from millimeter-wave radar and piezoelectric sensors based on the localized scene-sleep stage collaborative recognition of edge nodes. It dynamically and adaptively adjusts the acquisition frequency, gain and sampling weight according to the recognition results, and completes signal denoising, filtering and temporal feature extraction at the edge end to obtain high-dimensional temporal features such as heart rate, respiration, body movement and posture deviation, and constructs and outputs a continuous physiological state vector.

[0056] The multi-source acquisition unit is used by edge nodes to collect environmental information, raw physiological signals from millimeter-wave radar and piezoelectric sensors, forming multi-source time-series raw data. Environmental information includes parameters such as temperature, humidity, ambient noise, and light intensity, which are acquired in real-time by environmental sensors and converted into digital signals. The raw physiological signals from the millimeter-wave radar are the raw time-domain sampled values ​​obtained by analog-to-digital conversion of the echo signals received from the human body after the radar emits millimeter waves, reflecting physiological dynamics such as chest cavity fluctuations and body movements. The raw physiological signals from the piezoelectric sensors are the raw data obtained by sampling the voltage signals obtained after the sensors sense pressure changes during human respiration and body movement. The multi-source time-series raw data is generated by aligning the three types of signals according to timestamps to form a synchronously acquired time-series dataset.

[0057] The sleep recognition unit utilizes multi-source time-series raw data to perform localized sleep stage identification through time-series correlation analysis, outputting the current sleep stage identifier. In technical terms, time-series correlation analysis involves mining the correlation between different signals over time to capture the correspondence between physiological signals and sleep stages. Localized sleep stage identification refers to completing the identification locally at the edge node, without uploading to the cloud, reducing latency. The identified sleep stage identifiers mainly include three discrete categories: light sleep, deep sleep, and REM sleep. The digital processing procedure is as follows: First, the multi-source time-series raw data is time-synchronized and calibrated to eliminate sampling delays from different sensors. Then, the basic time-domain features of each signal are extracted. Using a time-series correlation algorithm, the extracted features are matched with locally pre-stored sleep stage feature templates to calculate similarity. Based on the similarity threshold, the current sleep stage is determined, and finally, the corresponding sleep stage identifier is output.

[0058] The adaptive sampling unit dynamically adjusts the sampling frequency, signal gain, and sampling weights based on sleep stage identifiers to adaptively sample the raw physiological signals, resulting in a well-ordered signal sequence with optimized signal-to-noise ratio (SNR). In technical terms, sampling frequency refers to the number of times the sensor samples signals per second; signal gain is a parameter used to amplify weak physiological signals; sampling weights are weighting coefficients assigned to different sensor signals; SNR is the ratio of the effective signal component to the noise component; and a well-ordered signal sequence refers to a signal sequence with uniform time intervals and a reasonable amplitude range after the sampling parameters are optimized. The digital processing involves: preset sampling parameter thresholds corresponding to different sleep stages; calling the corresponding parameter thresholds based on the output sleep stage identifiers; dynamically adjusting the sampling frequency and signal gain of the millimeter-wave radar and piezoelectric sensor; and adjusting the sampling weights based on environmental information. The raw physiological signals are then resampled according to the adjusted parameters, and outliers in the sampling process are removed to obtain a well-ordered signal sequence with optimized SNR and time alignment.

[0059] The signal purification unit performs adaptive noise reduction, bandpass filtering, and artifact removal on the regularized signal sequence to separate the pure signal components of heart rate, respiration, body movement, and posture. Technical terms include adaptive noise reduction, bandpass filtering, artifacts, and pure signal components. The digital processing procedure is as follows: An adaptive noise reduction algorithm is used to analyze the noise characteristics of the regularized signal sequence, dynamically adjust the filter kernel parameters, and filter out random noise. For the respective frequency ranges of heart rate, respiration, body movement, and posture deviation, corresponding bandpass filters are designed and applied to the regularized signal sequence, retaining the frequency components corresponding to each physiological signal. Through threshold judgment, artifacts in the signal are removed, separating the pure signal components corresponding to heart rate, respiration, body movement, and posture.

[0060] The sleep modeling module is used to perform high-order spectral analysis on continuous physiological state vectors, calculate the cardiopulmonary coupling phase relationship and energy distribution characteristics, and combine body movement and head posture information to obtain high-precision sleep stage probability, micro-arousal index and sleep state transition prediction value, forming and outputting a real-time sleep state digital matrix.

[0061] The higher-order spectral unit is used to perform third-order cumulant estimation on the continuous physiological state vector, removing Gaussian noise and linear interference to obtain the higher-order spectral density matrix. The continuous physiological state vector is a one-dimensional time-series vector containing multiple physiological parameters such as electrocardiogram (ECG), respiration (RESP), and blood oxygenation, acquired in real time by sensors. The sampling frequency is fixed, and each sampling point corresponds to a combination of multiple physiological parameters at a given time moment. Third-order cumulant estimation is a nonlinear signal processing method that performs third-order joint statistical operations on the time-series vector. Specifically, it iterates through all time-series sampling points of the vector, calculates the average of the products of any three samples at different times, and subtracts the product of the averages of the samples at each time moment, preserving the nonlinear correlation structure and non-Gaussian characteristics of the signal. Since the third-order cumulant of Gaussian noise is always zero, and the third-order cumulant of linear interference signals has no effective nonlinear information, this operation can directly suppress Gaussian white noise and linear interference, filtering out meaningless signal components. By utilizing the Fourier transform relationship between cumulants and spectral densities, the third-order cumulants calculated in the time domain are transformed into a bispectral distribution in the frequency domain. Then, through matrix transformation, a higher-order spectral density matrix is ​​obtained. The rows and columns of the higher-order spectral density matrix correspond to signal frequency points, and the matrix element values ​​represent the nonlinear coupling strength between the corresponding two frequency points. It only contains nonlinear coupling information of core physiological signals such as cardiopulmonary function, completely eliminating Gaussian noise and linear interference, and serves as the sole input for the next step of processing.

[0062] The cardiopulmonary coupling unit is used to extract bispectral phase and amplitude features of cardiopulmonary signals based on a high-order spectral density matrix, calculate the phase synchronization index and cross-spectral energy ratio, and form a cardiopulmonary coupling feature vector. The core value of the high-order spectral density matrix is ​​that it reflects the nonlinear coupling relationship of cardiopulmonary signals in the frequency domain. The regions with non-zero values ​​in the matrix correspond to the dominant frequency coupling positions of the cardiopulmonary signals, i.e., the frequency range where the dominant ECG and respiratory frequencies interact. The bispectral phase and amplitude features are the core information extracted from this matrix: the bispectral amplitude refers to the element value at the corresponding frequency point in the matrix, directly reflecting the energy intensity of the cardiopulmonary signal at that coupling frequency point; the larger the amplitude, the stronger the cardiopulmonary coupling at that frequency point. The bispectral phase refers to the phase angle corresponding to the matrix element, reflecting the phase difference distribution between the ECG and respiratory signals at the corresponding coupling frequency point; the more stable the phase difference, the better the synchronization of the cardiopulmonary signals. The phase synchronization index is a quantitative indicator based on bispectral phase calculation. Specifically, it involves statistically analyzing the phase difference across all coupling frequencies, calculating the consistency and stability of these phase differences, and obtaining a value between 0 and 1. The closer the value is to 1, the higher the degree of phase synchronization of the cardiopulmonary signal. The cross-spectral energy ratio is calculated based on bispectral amplitude. Specifically, it calculates the proportion of the bispectral amplitude of each coupling frequency to the total amplitude of all coupling frequencies, reflecting the concentration of cardiopulmonary coupling energy distribution. Frequency points with higher ratios better reflect the core characteristics of cardiopulmonary coupling. Finally, the phase synchronization index and corresponding cross-spectral energy ratio of all coupling frequencies are combined dimensionally according to frequency order. After normalization to eliminate differences in numerical ranges, a one-dimensional structured cardiopulmonary coupling feature vector is formed.

[0063] The feature fusion unit aligns and normalizes the cardiopulmonary coupling feature vector with body motion and head posture temporal data to construct a multi-dimensional physiological feature tensor. Body motion temporal data is a one-dimensional temporal signal collected by a body motion sensor, reflecting the intensity of body activity during sleep; the value at each moment corresponds to the amplitude of body motion at that moment. Head posture temporal data is temporal data collected by gyroscopes and accelerometers, reflecting the three-dimensional spatial posture of the head, including temporal changes in pitch, roll, and yaw angles; each moment corresponds to a set of posture parameters. Feature alignment is the primary processing step, unifying the sampling rate of all signals and synchronizing the timestamps of all signals to ensure a one-to-one correspondence between the cardiopulmonary coupling features, body motion data, and head posture data at each moment. Simultaneously, a sliding window of the same length is used to segment the three types of signals to ensure dimensional consistency in subsequent fusion. Normalization fusion is the core step, specifically involving the normalization of all feature components of the three types of signals (using min-max normalization to map all values ​​to between 0 and 1) to eliminate interference from differences in physical dimensions (cardiopulmonary coupling features are dimensionless, body movement data are amplitude values, and head posture data are angle values) and amplitude ranges. Then, the normalized cardiopulmonary coupling feature vectors, body movement temporal data, and head posture temporal data are superimposed and combined along the feature dimension to construct a multi-dimensional physiological feature tensor. This tensor contains three core dimensions: a time dimension (corresponding to the temporal changes of the signal), a feature dimension (corresponding to the three types of features: cardiopulmonary coupling, body movement, and head posture), and a spatial dimension (corresponding to the three-dimensional information of head posture), thus fully integrating physiological, motor, and spatial posture information during sleep.

[0064] The sleep reconstruction unit performs windowing weighting and nonlinear mapping on multi-dimensional physiological feature tensors, fits a sleep stage discrimination function, obtains a sleep probability distribution sequence, and reconstructs the sleep probability distribution sequence into an equal-dimensional numerical array by time slices, outputting a real-time sleep state digital matrix. The core of windowing weighting is to enhance effective signals and weaken redundant information. Specifically, it uses a fixed-length sliding time window (the window length is adaptively adjusted according to the sleep cycle, and the window step size is half the window length) to traverse the time dimension of the multi-dimensional physiological feature tensor. Weighting operations are performed on the feature tensor elements within each time window, assigning higher weights to features from recent times and lower weights to features from distant times, thereby highlighting the influence of real-time physiological state. Nonlinear mapping transforms high-dimensional features into low-dimensional features suitable for sleep stage classification. Through a nonlinear mapping function, the multi-dimensional physiological feature tensor is projected from the high-dimensional space to the low-dimensional discrimination space, filtering out redundant features and retaining core features highly correlated with sleep stage. Based on the mapped low-dimensional features, a sleep stage discrimination function is fitted. This function takes the low-dimensional features as input and outputs the probabilities of different sleep stages (light sleep, deep sleep, REM sleep, wakefulness). The function parameters are determined through training with a large number of samples. This discrimination function calculates the probability value corresponding to each sleep stage based on the low-dimensional features at each time step. The probability values ​​from all time steps are combined to form a continuous sleep probability distribution sequence. Each element in this sequence is a multi-dimensional probability vector, corresponding to the probability distribution of the sleep state at a given time step. The sleep probability distribution sequence is divided into fixed time slices. The probability vectors within each slice are arranged in a regular row or column pattern, reconstructing a fixed-dimensional numerical array. This numerical array is a real-time sleep state digital matrix, where rows correspond to time slices, columns correspond to sleep stage probabilities, and each element represents the probability of a specific sleep stage within the corresponding time slice.

[0065] The waveform generation module is used to calculate the head spatial position and posture offset from the real-time sleep state digital matrix, dynamically optimize the phase and amplitude of multiple excitation coils, achieve coherent magnetic focusing to ensure stable superposition of the target area magnetic field, and at the same time use subthreshold random chaotic modulation to generate and output a non-repeating individualized electromagnetic field excitation waveform.

[0066] The attitude calculation unit processes the real-time sleep state digital matrix. First, a fixed-length sliding time window is determined, with the window length adaptively adjusted based on the sampling frequency. The window step size is set to half the window length. Temporal segments of the matrix are extracted window by window, and the mean, variance, skewness, kurtosis, and spatiotemporal correlation of all elements within each window are calculated to extract temporal statistical features. Singular value decomposition is used to perform spatial feature calculation on the calculated covariance matrix, separating the core sleep state components and the spatial attitude-related components. Based on the differences in sensitivity to spatial perturbations among different sleep stages (light sleep, deep sleep, REM sleep), linear weighting combined with normalization is used to fit sleep-related attitude constraint coefficients with values ​​between 0 and 1, and with constraint range and confidence attributes.

[0067] The attitude modeling unit uses sleep-related attitude constraint coefficients as weighting factors, multiplying them point-by-point with real-time discrete attitude sampling data acquired by triaxial accelerometers and gyroscopes to suppress noise interference. Outliers in the sampling sequence are removed using the 3σ criterion, and missing points resulting from outlier removal are filled in using linear interpolation. The X, Y, and Z axis position components of the head in a three-dimensional Cartesian coordinate system are calculated through direction cosine matrix and quaternion update operations, along with the offsets of pitch, roll, and yaw Euler angles. The position components and angular offsets are standardized to eliminate dimensional differences, and then concatenated dimensionally to generate a head attitude offset feature vector containing position offset, angular offset, and confidence weights.

[0068] The coil optimization unit is used to construct a linear response model between the spatial distribution of the electromagnetic field of multiple excitation coils and the head posture deviation, taking the head posture deviation feature vector as input. It clarifies the correspondence between posture changes and the excitation phase and amplitude of each coil, and establishes the transfer function matrix. Using the uniformity of the spatial electromagnetic field and the coupling efficiency between the coils and the human body as joint optimization objectives, reasonable weighting coefficients are set to construct the overall objective function. The conjugate gradient method is used for iterative solution, determining the initial and termination conditions, iterating until the objective function reaches its optimum. This yields the optimal phase compensation and amplitude gain coefficient for each excitation coil. The relevant parameters of all coils are combined sequentially to form a multi-coil phase and amplitude optimization parameter set.

[0069] The waveform generation unit extracts the optimal phase compensation and amplitude gain coefficients for each coil based on the multi-coil phase and amplitude optimization parameters. It selects a Logistic chaotic mapping as the modulation model, filters parameters that meet subthreshold requirements, and removes outliers. The amplitude gain coefficients are controlled within the subthreshold range to avoid physiological stimulation, thus obtaining the basic excitation parameters for the subthreshold coils. Using these basic excitation parameters as initial input, the optimal phase compensation and amplitude gain coefficients for each coil are used as the initial and control parameters for the chaotic mapping, respectively. The control parameters and initial iteration values ​​for the chaotic mapping are set, and the Logistic chaotic mapping is used for initial iteration initialization to obtain the initial chaotic iteration sequence. Based on the initial chaotic iteration sequence, the phase compensation is randomly perturbed, controlling the perturbation intensity within a preset reasonable range to avoid excessive perturbation affecting stability, generating a perturbed phase iteration sequence. The perturbed phase iteration sequence and amplitude gain coefficients are input into the Logistic chaotic mapping operation flow for aperiodic iterative modulation. A sufficient number of iterations are set, and the non-periodic nature of the chaotic sequence is utilized to obtain a non-repeating chaotic excitation parameter sequence. The chaotic excitation parameter sequence is converted into a time-domain waveform signal, and waveform synthesis is performed by combining it with the coil excitation frequency. After smoothing, the iteration results of each coil are spliced ​​together according to the time sequence to generate an individualized electromagnetic field excitation waveform that dynamically matches the real-time sleep state and head posture, and has no repetition and meets the subthreshold requirements.

[0070] The interference compensation module is used to perform adaptive compensation calculations on the non-repeating individualized electromagnetic field excitation waveform combined with environmental magnetic field distortion and metal interference data. It calculates the interference cancellation coefficient in real time, obtains and outputs the corrected coil drive current to eliminate magnetic field distortion, and simultaneously performs multimodal fusion, temporal alignment and standardization processing on multi-source physiological, environmental and behavioral data. It dynamically updates monitoring thresholds and early warning rules, completes sleep state assessment and abnormal physiological-behavioral correlation tracing, and obtains and outputs sleep abnormality risk control data.

[0071] The data preprocessing unit performs synchronous temporal alignment and noise preprocessing on three types of data—non-repeating individualized electromagnetic field excitation waveforms, acquired environmental magnetic field distortion data, and metal interference data—to eliminate data bias and construct a basic dataset for interference analysis. The non-repeating individualized electromagnetic field excitation waveform is a time-domain excitation signal that dynamically changes with the sleep state and has no periodic repetition, generated through subthreshold chaotic modulation. The environmental magnetic field distortion data is interference signals generated by external magnetic fields in the sleep environment, collected by a magnetic field sensor. The metal interference data is the scattering and shielding interference signal generated by metal objects in the sleep scene on the excitation magnetic field. Temporal alignment is the core preprocessing step, specifically unifying the sampling rate and timestamps of the three types of data to ensure a one-to-one correspondence between the excitation waveform, environmental magnetic field distortion, and metal interference data at each moment, eliminating biases in the time dimension. Noise preprocessing employs an adaptive filtering algorithm to filter out irrelevant noise such as Gaussian white noise and power frequency interference from the three types of data, retaining only valid signal components. Finally, the time-aligned and noise-preprocessed data are integrated along the time dimension to form a structured dataset containing the excitation signal and the two types of interference signals—the basic dataset for interference analysis.

[0072] The interference modeling unit extracts feature correlations between the non-repeating individualized electromagnetic field excitation waveform and various types of interference data from the interference analysis dataset, establishes an adaptive compensation calculation model, and calculates the interference cancellation coefficient by comparing and analyzing the amplitude and phase differences between the excitation waveform and the interference signal. The interference analysis dataset includes three types of signals: preprocessed excitation waveform, environmental magnetic field distortion, and metallic interference. Feature correlation extraction refers to extracting the frequency domain amplitude and phase characteristics and time-series variation patterns of the three types of signals from the dataset, analyzing the correspondence between the non-repeating individualized electromagnetic field excitation waveform and the two types of interference data at different frequency points, and determining the influence of the interference signal on the amplitude attenuation and phase shift of the excitation waveform. The adaptive compensation calculation model is a dedicated interference compensation model adapted to non-repeating chaotic excitation waveforms. Unlike traditional fixed waveform compensation models, its core is to construct a dynamic mapping relationship between excitation waveform features and interference features, setting model adaptation parameters to ensure real-time tracking of the dynamic changes of the non-repeating excitation waveform. The comparative analysis process involves substituting the amplitude and phase characteristics of the extracted excitation waveform and the interference signal into the model, calculating the amplitude difference and phase shift between the two at each frequency point, quantifying the degree of interference of the interference signal on the excitation waveform, and then calculating the interference cancellation coefficient that can accurately cancel the environmental magnetic field distortion and metal interference through linear fitting and iterative correction. This coefficient is updated in real time with the dynamic changes of the excitation waveform and the interference signal.

[0073] The current correction unit is used to substitute the interference cancellation coefficient into the initial calculation model of the coil drive current, correcting the drive current corresponding to the non-repeating individualized electromagnetic field excitation waveform to obtain the coil drive current. The initial calculation model of the coil drive current is a computational model constructed based on the electromagnetic field excitation principle, used to convert the excitation waveform into the coil drive current. The inputs are the amplitude and phase parameters of the non-repeating individualized electromagnetic field excitation waveform, and the output is the initial coil drive current. The interference cancellation coefficient is a correction parameter calculated in the previous step, used to cancel environmental and metallic interference. After substituting the correction parameter into the initial calculation model, the amplitude and phase of the initial drive current are corrected in real time through multiplicative weighted operations. This compensates for the drive current deviation caused by interference signals, eliminates the influence of interference on the excitation magnetic field, and obtains a coil drive current that can accurately output the target excitation magnetic field.

[0074] The risk control output unit is used to perform multi-source data fusion processing on coil drive current, non-repetitive individualized electromagnetic field excitation waveforms, and interference data. This data is then correlated with real-time sleep physiological characteristics to conduct sleep assessments, screen for sleep abnormality indicators, and output sleep abnormality risk control data. Multi-source data fusion processing normalizes four types of data: coil drive current, non-repetitive individualized electromagnetic field excitation waveforms, environmental magnetic field distortion data, and metallic interference data, eliminating dimensional differences. Through feature splicing and weight allocation, it integrates the effective information from various data types to obtain fused data that comprehensively reflects the excitation effect and interference state. Real-time sleep physiological characteristics refer to synchronously collected physiological parameters such as heart rate, respiration, and blood oxygenation. Correlation processing matches the fused data with physiological characteristics in a time sequence, analyzes the impact of excitation waveform correction on sleep physiological state, and conducts sleep assessments. The sleep assessment process compares the data with preset normal sleep physiological thresholds, screens out sleep abnormality indicators such as abnormal heart rate, respiratory disturbances, and low blood oxygenation, quantifies and classifies these abnormal indicators, and finally integrates information such as abnormal indicators, interference state, and excitation effect to output sleep abnormality risk control data for sleep risk management.

[0075] The closed-loop control module is used to integrate sleep abnormality risk control data with sleep difference status, equipment status and power grid load information, iterate the pulse electromagnetic field multi-parameter control strategy in real time and issue it to execute intervention, collect intervention feedback physiological signals, and reverse optimize the collection weight, magnetic focusing parameters and excitation waveform to form and output a full-link edge autonomous two-layer closed-loop digital control process. The entire process is performed locally, and only the core features are uploaded to the cloud in a lightweight manner. In extreme scenarios, cloud-assisted decision-making is initiated.

[0076] The data fusion unit is used by edge nodes to perform multi-source data normalization and feature fusion based on sleep anomaly risk control data, sleep difference state quantification data, equipment operating parameters, and digitized grid load information, outputting a standardized control decision input matrix. Edge nodes are terminal computing units deployed locally within the sleep scenario, possessing data processing and decision-making capabilities. They do not rely on the cloud and can achieve localized real-time processing. Sleep anomaly risk control data is structured data output from previous steps, containing sleep anomaly indicators, disturbance states, and stimulus effects, reflecting the degree of sleep risk. Sleep difference state quantification data is data obtained by quantifying differences in different sleep stages and individual sleep habits; quantification indicators include sleep cycle duration and physiological parameter fluctuation range. Equipment operating parameters refer to real-time operating data from devices such as multi-excitation coils and sensors, including coil operating voltage, current, and sensor sampling frequency. Grid load digitization information is data obtained by digitally collecting and quantifying the grid load status in the area where the sleep scenario is located, reflecting the real-time grid load situation. Edge computing refers to data processing operations performed locally at edge nodes, avoiding data transmission delays. The specific digital processing procedure is as follows: First, the four types of multi-source data are normalized using a min-max normalization method, mapping the values ​​of each type of data to the same interval to eliminate interference caused by differences in physical dimensions and amplitude ranges. Next, feature fusion is performed to extract the core features of each data type. These features are then weighted and concatenated dimensionally to form a structured and standardized matrix—the standardized control decision input matrix. This matrix fully contains all the multi-source data information required for control decisions.

[0077] The strategy optimization unit is used to generate a pulse electromagnetic field (PEM) control strategy by iteratively optimizing the frequency, intensity, and timing parameters of the pulse electromagnetic field based on a standardized control decision input matrix, historical control parameters, and sleep intervention effect data through edge AI. Historical control parameters refer to the pulse electromagnetic field parameters used in past sleep interventions, including frequency, intensity, and timing intervals. Sleep intervention effect data refers to data collected under the corresponding historical control parameters, such as changes in sleep physiological characteristics and improvements in sleep quality, reflecting the effectiveness of the control parameters. Edge AI is a lightweight artificial intelligence algorithm model deployed on edge nodes, requiring no cloud computing power and enabling local real-time iterative optimization. The digital processing involves using the standardized control decision input matrix as the core input to the edge AI model, along with historical control parameters and sleep intervention effect data. The AI ​​model trains and learns from the historical data to uncover the mapping relationship between the standardized control decision input and the optimal control parameters. Iterative optimization focuses on the core parameters of the pulse electromagnetic field (frequency, intensity, and timing). Each iteration combines current multi-source data with historical effect data to correct parameter values, ensuring that the parameters are adapted to the current sleep state, equipment operating status, and power grid load. After iterating until the predicted intervention effect corresponding to the parameters reaches the optimal level, the iteration stops. The optimized frequency, intensity, and timing parameters are logically combined to generate a pulse electromagnetic field modulation strategy that can be directly executed. This strategy is individualized and real-time, and can adapt to all dynamic changes in the current sleep scenario.

[0078] The intervention feedback unit is used to conduct sleep intervention using electromagnetic fields according to a pulsed electromagnetic field modulation strategy. It collects feedback physiological signals during the intervention process, performs noise reduction and feature extraction on the signals, and outputs feedback physiological characteristic parameters. The pulsed electromagnetic field modulation strategy is an execution plan generated in the previous step, containing frequency, intensity, and timing parameters. Sleep intervention refers to controlling multiple excitation coils to output pulsed electromagnetic fields according to this strategy to regulate the human sleep state, ensuring the electromagnetic field remains within the subthreshold range to avoid physiological stimulation. Feedback physiological signals are real-time physiological response signals collected by physiological sensors during the intervention process, including ECG, respiration, blood oxygenation, and EEG, used to reflect the intervention effect. Noise reduction processing uses an adaptive filtering algorithm to filter out irrelevant components such as Gaussian white noise, power frequency interference, and equipment operating noise from the collected feedback physiological signals, retaining only valid physiological signals. Feature extraction involves extracting core features related to sleep state and intervention effect from the noise-reduced feedback physiological signals, including heart rate variability, respiratory rate, blood oxygen saturation, and EEG rhythm characteristics. These features are then quantified to form structured feedback physiological feature parameters. These parameters directly reflect the actual impact of pulsed electromagnetic field intervention on human sleep and serve as the core basis for subsequent parameter correction.

[0079] The closed-loop correction unit compares the feedback physiological characteristic parameters with the preset sleep regulation target parameters, calculates the deviation value, and corrects the pulse electromagnetic field regulation parameters through an edge iterative algorithm, updating the regulation strategy and forming a full-link edge autonomous two-layer closed-loop sleep regulation data flow process. The preset sleep regulation target parameters are physiological characteristic standard values ​​set according to normal sleep standards and individual sleep needs, including normal heart rate range, respiratory rate range, blood oxygen saturation threshold, and the proportion of each sleep stage, serving as the benchmark for judging the intervention effect. The deviation value is calculated by comparing the feedback physiological characteristic parameters with the preset target parameters dimension by dimension, quantifying the gap between the intervention effect and the target. The larger the deviation value, the lower the fit between the current regulation parameters and the target. The edge iterative algorithm is a lightweight iterative optimization algorithm deployed on edge nodes, enabling local real-time parameter correction without relying on external computing power. The digital processing process is as follows: the calculated deviation value is input into the edge iterative algorithm. The algorithm, based on the magnitude and direction of the deviation value, specifically corrects the frequency, intensity, and timing parameters of the pulse electromagnetic field, with the correction direction being to reduce the deviation between the feedback parameters and the target parameters. After the correction is completed, the original pulse electromagnetic field modulation strategy is updated. The new modulation strategy will be used again for sleep intervention. At the same time, new feedback physiological signals are collected, and the above comparison and correction process is repeated to form a two-layer closed loop of full-link edge autonomy. The inner closed loop is intervention → feedback → correction, and the outer closed loop is multi-source data → strategy → intervention → feedback → strategy update. This realizes real-time adaptive optimization of sleep regulation without cloud intervention.

[0080] In summary, this embodiment provides a non-head-mounted pulsed electromagnetic field sleep intelligent regulation system based on edge computing. By integrating sleep abnormality risk control, individual differences, device status, and power grid load information, a standardized decision matrix is ​​constructed. Edge AI iteratively optimizes the frequency, intensity, and timing of the electromagnetic field to generate dynamically adaptable regulation strategies. Interventions are executed, and real-time feedback physiological signals such as ECG, respiration, and blood oxygenation are collected. Physiological features are extracted and the deviation from the regulation target is calculated. Edge iterative algorithms are used to correct the regulation parameters in real time. Through an inner-layer intervention, feedback, and correction, and an outer-layer data, strategy, and update dual-closed-loop structure, the system achieves full-process self-sensing, self-decision-making, self-execution, and self-optimization. This results in a fully automatic intelligent regulation effect that requires no manual intervention, adapts to individual changes, resists environmental interference, and maintains stable long-term effects, ensuring continuous optimal sleep intervention.

[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-head-mounted pulsed electromagnetic field sleep intelligent control system based on edge computing, characterized in that, include: The signal acquisition module is used to collect raw physiological signals based on edge nodes through localized scene sleep stage collaborative identification, extract temporal features from the raw physiological signals, and output a continuous physiological state vector. The sleep modeling module is used to perform high-order spectral analysis on the continuous physiological state vector, calculate the cardiopulmonary coupling phase relationship and energy distribution characteristics, combine body movement and head posture information to obtain high-precision sleep parameters, and output a real-time sleep state digital matrix. The waveform generation module is used to calculate the head spatial position and posture offset based on the real-time sleep state digital matrix, dynamically optimize the phase and amplitude of the multi-excitation coil, and generate a non-repeating individualized electromagnetic field excitation waveform using subthreshold random chaotic modulation. The interference compensation module is used to perform adaptive compensation calculations based on the non-repeating individualized electromagnetic field excitation waveform, environmental magnetic field distortion, and metal interference data, calculate the interference cancellation coefficient to obtain the corrected coil drive current, perform multi-source data processing and sleep assessment, and output sleep abnormality risk control data. The closed-loop control module is used to integrate the sleep abnormality risk control data with sleep difference status, equipment status and power grid load information, iterate the pulse electromagnetic field control strategy and issue it for execution intervention, collect feedback physiological signals to optimize parameters, and output the full-link edge autonomous dual-layer closed-loop sleep control process.

2. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 1, characterized in that: The signal acquisition module includes: The multi-source acquisition unit is used to acquire environmental information, millimeter-wave radar and piezoelectric sensor raw physiological signals from edge nodes to form multi-source time-series raw data. The sleep recognition unit is used to perform localized sleep stage recognition through time-series correlation analysis using the multi-source raw data and output the current sleep stage identifier. An adaptive sampling unit is used to dynamically adjust the sampling frequency, signal gain, and sampling weight according to the current sleep stage identifier, and adaptively sample the original physiological signal to obtain a regular signal sequence with optimized signal-to-noise ratio. The signal purification unit is used to perform adaptive noise reduction, bandpass filtering and artifact removal on regular signal sequences to separate clean signal components of heart rate, respiration, body movement and posture. The feature fusion unit is used to extract temporal features from the pure signal components and normalize and fuse them to generate continuous high-dimensional temporal features, and construct and output physiological state vectors.

3. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 1, characterized in that: The sleep modeling module includes: Higher-order spectral units are used to estimate the third-order cumulants of continuous physiological state vectors, remove Gaussian noise and linear interference, and obtain the higher-order spectral density matrix. The cardiopulmonary coupling unit is used to extract the bispectral phase and amplitude features of the cardiopulmonary signal based on the higher-order spectral density matrix, calculate the phase synchronization index and cross-spectral energy ratio, and form a cardiopulmonary coupling feature vector. The feature fusion unit is used to perform feature alignment and normalization fusion with the cardiopulmonary coupling feature vector and body motion and head posture time series data to construct a multi-dimensional physiological feature tensor. The sleep reconstruction unit is used to perform windowed weighting and nonlinear mapping on the multi-dimensional physiological feature tensor, fit the sleep stage discrimination function, obtain the sleep probability distribution sequence, reconstruct the sleep probability distribution sequence into an equal-dimensional numerical array by time slices, and output the real-time sleep state digital matrix.

4. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 3, characterized in that: The specific steps for forming the cardiopulmonary coupling feature vector in the cardiopulmonary coupling unit are as follows: Based on the higher-order spectral density matrix, the bispectral phase angle and amplitude spectrum value are solved point by point to locate the cardiopulmonary main frequency coupling region, and non-coupled frequency interference is eliminated to obtain the bispectral phase-amplitude distribution sequence. Calculate the phase difference and phase coherence value of adjacent frequency points based on the bispectral phase-amplitude distribution sequence, statistically analyze the phase synchronization duration and consistency degree, and generate a cardiopulmonary phase synchronization index sequence. Based on the phase synchronization index sequence and bispectral amplitude distribution, the energy proportion and energy concentration of each coupling frequency point are calculated, and the cardiopulmonary energy distribution characteristics are obtained after normalization. The phase synchronization index sequence and cardiopulmonary energy distribution characteristics are concatenated by dimension and standardized to remove dimensional differences, forming a cardiopulmonary coupling feature vector.

5. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 1, characterized in that: The waveform generation module includes: The attitude calculation unit is used to perform temporal sliding window statistics and spatial feature calculation on the real-time sleep state digital matrix, extract the attitude-sensitive weights corresponding to the sleep stage, and obtain sleep-related attitude constraint coefficients. The posture modeling unit is used to calculate the three-dimensional spatial coordinates and angular offset of the head using sleep-related posture constraint coefficients and real-time posture sampling data, and to generate a head posture offset feature vector. The coil optimization unit is used to establish a coil excitation response model based on the head posture offset feature vector, calculate the optimal phase compensation and amplitude gain of each excitation coil, and obtain multi-coil phase amplitude optimization parameters. The waveform generation unit is used to substitute the multi-coil phase amplitude optimization parameters into the subthreshold chaotic mapping, perform random phase perturbation and non-periodic iterative modulation, and generate a non-repeating individualized electromagnetic field excitation waveform.

6. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 5, characterized in that: In the coil optimization unit, the specific steps for obtaining the phase amplitude optimization parameters of multiple coils are as follows: The head's three-dimensional spatial coordinates and angle offset information are extracted from the head posture offset feature vector. Based on the inherent parameters and installation position of the multi-excitation coil, the correlation mapping between coil excitation and spatial magnetic field distribution is constructed, and a basic coil excitation response model is established. Substitute the position and angle data in the head posture offset feature vector into the basic coil excitation response model to simulate the coil excitation effect under different phases and amplitudes, calculate the magnetic field distribution error and coupling efficiency, and determine the optimization objective function. The phase and amplitude of each excitation coil are iteratively solved according to the optimization objective function, and the phase compensation amount and amplitude gain are corrected to obtain the preliminary optimal parameters corresponding to each coil. The preliminary optimal parameters are processed according to coil number, and abnormal parameters are eliminated by combining the confidence weight of the head attitude offset feature vector to obtain the multi-coil phase amplitude optimization parameters.

7. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 5, characterized in that: The specific steps for generating a unique, non-repeating electromagnetic field excitation waveform in the waveform generation unit are as follows: Based on the multi-coil phase amplitude optimization parameters, the optimal phase compensation amount and amplitude gain coefficient corresponding to each coil are extracted, parameters that meet the subthreshold requirements are selected, and abnormal parameters that exceed the threshold range are eliminated to obtain the basic parameters for subthreshold coil excitation. Using the basic parameters of the subthreshold coil excitation as the initial input, the control parameters and initial values ​​of the chaotic mapping are set, and the chaotic mapping is used to perform the initial iteration initialization to obtain the initial sequence of chaotic iteration. The phase compensation amount is randomly perturbed according to the initial chaotic iteration sequence, and the perturbation intensity is controlled within a preset range to generate a perturbed phase iteration sequence. The phase iteration sequence and amplitude gain coefficient are input into the chaotic mapping operation process for non-periodic iterative modulation. The number of iterations is controlled to ensure non-periodicity, and a chaotic excitation parameter sequence is obtained. The chaotic excitation parameter sequence is converted into a time-domain waveform signal, and waveform synthesis is performed in combination with the coil excitation frequency. After smoothing, a unique electromagnetic field excitation waveform without repetition is generated.

8. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 1, characterized in that: The interference compensation module includes: The data preprocessing unit is used to perform synchronous temporal alignment and noise preprocessing on the three types of data based on the non-repeating individualized electromagnetic field excitation waveform and the collected environmental magnetic field distortion data and metal interference data, to eliminate data bias and construct a basic dataset for interference analysis. The interference modeling unit is used to extract the feature correlation between the non-repeating individualized electromagnetic field excitation waveform and various interference data based on the interference analysis basic dataset, establish an adaptive compensation calculation model, and calculate the interference cancellation coefficient by comparing and analyzing the amplitude and phase differences between the excitation waveform and the interference signal. The current correction unit is used to substitute the interference cancellation coefficient into the initial calculation model of the coil drive current, correct the drive current corresponding to the non-repeating individualized electromagnetic field excitation waveform, and obtain the coil drive current. The risk control output unit is used to perform multi-source data fusion processing on the coil drive current, non-repetitive individualized electromagnetic field excitation waveform and interference data, correlate it with real-time sleep physiological characteristics to carry out sleep assessment, screen sleep abnormality indicators, and output sleep abnormality risk control data.

9. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 8, characterized in that: In the interference modeling unit, the specific steps for calculating the interference cancellation coefficient are as follows: Separate the non-repeating individualized electromagnetic field excitation waveform, environmental magnetic field distortion data and metal interference data from the basic dataset of interference analysis, extract the frequency domain amplitude, phase characteristics and time series variation law of the three types of data respectively, and determine the characteristic correspondence between the excitation waveform and the interference data. An adaptive compensation calculation model is constructed based on the aforementioned feature correspondence to adapt to non-repeating chaotic excitation waveforms. The characteristic parameters of the excitation waveform and various interference data are input into the adaptive compensation calculation model to analyze the amplitude difference and phase shift of the excitation waveform and various interference data at different frequency points, and to quantify the degree of influence of interference on the excitation waveform. Based on the mapping relationship between the degree of interference on the excitation waveform and the adaptive compensation operation model, the interference cancellation coefficient is obtained by iteratively calculating the correction parameters.

10. The edge computing-based non-head-mounted pulsed electromagnetic field sleep intelligent control system according to claim 1, characterized in that: The closed-loop control module includes: The data fusion unit is used by edge nodes to perform multi-source data normalization and feature fusion based on sleep abnormality risk control data, sleep difference state quantitative data, equipment operating parameters and grid load digital information, and output a standardized control decision input matrix. The strategy optimization unit is used to generate a pulse electromagnetic field regulation strategy by iteratively optimizing the frequency, intensity, and timing parameters of the pulse electromagnetic field through edge AI based on the standardized regulation decision input matrix and historical regulation parameters and sleep intervention effect data. The intervention feedback unit is used to perform sleep intervention with electromagnetic fields according to the pulse electromagnetic field modulation strategy, collect feedback physiological signals during the intervention process, perform noise reduction and feature extraction on the signals, and output feedback physiological feature parameters. The closed-loop correction unit is used to compare the feedback physiological characteristic parameters with the preset sleep regulation target parameters, calculate the deviation value, correct the pulse electromagnetic field regulation parameters through the edge iteration algorithm, update the regulation strategy, and form a full-link edge autonomous double-layer closed-loop sleep regulation data flow process.

Images