Sleep posture recognition method based on frequency domain feature extraction

Abstract

The application discloses a sleep posture recognition method based on frequency domain feature extraction and relates to the field of intelligent sleep monitoring, and comprises the following steps: multi-source synchronous acquisition and physiological partition mapping, multi-dimensional frequency domain feature extraction and vector construction, cross-modal fusion decision, frequency domain trajectory stability determination, and cycle iterative update and physiological backstepping. A pillow surface array is divided into multiple physiological partitions, a two-dimensional Fourier transform is performed on each partition pressure sub-matrix to extract spatial features, a one-dimensional FFT is performed on key points and IMU time series to extract time domain and frequency domain features, and the features are fused into a standardized high-dimensional frequency domain feature vector, so that fine scene distinguishing requirements can be correspondingly met, pressure, posture and device state weights are dynamically divided, and when multiple sources are consistent, a result is directly output, and when conflicts occur, a secondary check is performed by using a historical stable template, interference signals are weakened, and the reliability of fusion confidence is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent sleep monitoring, and in particular to a sleeping posture recognition method based on frequency domain feature extraction. Background Technology

[0002] Sleep posture recognition, as a core component for achieving personalized health management and adaptive pillow adjustment, typically relies on pressure sensors and inertial measurement units for determination.

[0003] Existing technologies only perform simple time-domain statistics or global one-dimensional Fourier transforms on the overall signal. They have never divided the flexible pressure array into five distinct regions according to the physiological structure of the human head and neck: the central region of the posterior brain, the left and right neck and shoulder regions, the anterior neck support region, and the posterior occipital buffer zone. Furthermore, they have not performed two-dimensional Fourier transforms on the pressure submatrices of each region to extract spatial spectral features. As a result, the algorithm is unable to characterize the spatial morphology of the pressure distribution and the frequency rhythm of physiological micro-movements, and the recognition results cannot meet the requirements of fine-scale scenarios.

[0004] Existing technologies attempt to fuse pressure and inertial measurement unit data, but they use fixed weights to stitch together the original features without considering the physical consistency between pressure frequency domain features and attitude frequency domain features. When conflicts occur between pressure pre-classification, attitude determination and equipment indication, instantaneous interference is easily amplified, resulting in insufficient reliability of multi-source fusion results.

[0005] Existing technologies mostly rely on single-frame classification or simple moving average filtering or delayed confirmation, which cannot distinguish between the three trajectory patterns in real sleep position switching, instantaneous noise and posture adjustment. Under disturbance, the output is updated incorrectly, resulting in errors in sleep monitoring data.

[0006] The general recognition model is fixed once it leaves the factory. It does not set different cycles based on the labels of teenagers, pregnant women, and general populations, nor does it use the frequency domain baseline to back-calculate shoulder width growth, pregnancy progress, or weight fluctuations to achieve differentiated adjustment. As a result, the recognition accuracy of the device continues to deteriorate when the user grows and develops, changes in body shape during pregnancy, or weight fluctuations in ordinary adults.

[0007] Therefore, we propose a sleeping posture recognition method based on frequency domain feature extraction to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a sleeping posture recognition method based on frequency domain feature extraction, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a sleeping posture recognition method based on frequency domain feature extraction, comprising the following steps:

[0010] S1 Multi-source Synchronous Acquisition and Physiological Partition Mapping: Synchronously acquire two-dimensional pressure data of the pillow surface, head posture motion data and related device status data, and divide the pressure data into pressure sub-matrices of each partition according to the preset human head and neck physiological partitions.

[0011] S2 Multi-dimensional Frequency Domain Feature Extraction and Vector Construction: Two-dimensional Fourier transform is performed on the pressure submatrix of each region to extract the spatial spectrum feature set. One-dimensional Fourier transform is performed on the pressure time series and attitude time series of the selected location to extract the time and attitude spectrum features. The above features are fused to generate a high-dimensional frequency domain feature vector.

[0012] S3 cross-modal fusion decision: Combines pre-classification results and confidence, frequency domain features of motion data and device status information to perform dynamic weight allocation and multi-source consistency judgment, calculates the total fusion confidence based on weights, and outputs candidate sleeping postures and fusion confidence;

[0013] S4 Frequency Domain Trajectory Stability Determination: Obtain high-dimensional frequency domain feature vectors corresponding to candidate sleeping postures in multiple consecutive frames, calculate the cosine similarity of frequency domain feature vectors of adjacent frames frame by frame, and fit to form a frequency domain feature evolution trajectory curve; by calculating the monotonicity coefficient and fluctuation variance of the trajectory, update the stable sleeping posture label and output it only when it is determined to be a real sleeping posture change.

[0014] Preferably, the multiple physiological zones based on step S1 specifically include:

[0015] The central area of ​​the back of the head corresponds to the central calibration area of ​​the sensor array, which corresponds to the middle part of the back of the human head.

[0016] The left neck and shoulder area corresponds to the shoulder and neck calibration area on the left side of the sensor array, which corresponds to the left side of the human neck and shoulder.

[0017] The right neck and shoulder area corresponds to the neck and shoulder calibration area on the right side of the sensor array, which corresponds to the right side of the neck and shoulder of the human body.

[0018] The anterior neck support area corresponds to the cervical spine calibration area at the front edge of the sensor array, which corresponds to the front part of the human neck.

[0019] The occipital buffer zone corresponds to the cervical fossa calibration area at the rear edge of the sensor array, which corresponds to the area below the back of the head near the cervical fossa.

[0020] The coordinate range of each partition is pre-calibrated based on the physiological contact distribution of the human head and neck and is fixed in the MCU firmware. After the data acquisition is completed, the data of each partition is directly cut according to the calibrated coordinates.

[0021] Preferably, based on step S2, the basic features of the spatial spectrum include: spectral centroid, spectral entropy, spectral kurtosis, energy proportion of low spatial frequency bands, energy proportion of high spatial frequency bands, and spectral peak amplitude.

[0022] Among them, the centroid of the spectrum is the weighted average spatial frequency of the power spectrum, which is used to reflect the spatial scale of pressure distribution;

[0023] Spectral entropy is the information entropy of the normalized power spectrum, used to reflect the regularity of the pressure distribution pattern;

[0024] Spectral kurtosis is the fourth moment of the power spectrum, used to reflect the degree of peaks in the spectrum;

[0025] The energy proportion in the low spatial frequency band is the ratio of energy within a preset radial frequency range to the total energy, corresponding to a large-area, slowly varying pressure distribution;

[0026] The high spatial frequency band energy ratio is the ratio of energy in the preset high-frequency radial range to the total energy, corresponding to disturbance and noise signals;

[0027] The peak amplitude of the spectrum is the maximum value of the power spectrum after excluding the DC component, and is used to reflect the intensity of the main space pressure mode.

[0028] Preferably, based on step S2, the higher-order features of spectral difference and symmetry include:

[0029] Symmetry of the centroid of the left and right neck and shoulder regions is the normalized result of the difference between the centroids of the left and right neck and shoulder regions, used to distinguish between supine and lateral positions.

[0030] The energy symmetry of the left and right neck and shoulder areas is the ratio of the proportion of energy in the low spatial frequency band of the left and right neck and shoulder areas, which is used to quantify the degree of symmetry of the pressure distribution on the left and right sides.

[0031] The spectral energy difference between the front and rear regions is the difference in spectral energy between the anterior neck support area and the posterior occipital buffer zone, which is used to identify the forward or backward head posture.

[0032] The central and peripheral spectral attenuation slope is the difference in spectral centroid between the central region of the posterior brain and the occipital buffer zone divided by the regional calibration interval, used to reflect the smoothness of the pressure distribution transition.

[0033] The radial odd-even ring energy ratio is the energy ratio value of the odd-even ring bands divided according to the spatial frequency radius;

[0034] The energy percentage of even-order radial rings is the ratio of the energy of even-numbered radius rings to the total energy.

[0035] The global spectral variance is the variance of the centroids of the spectral frequencies of all physiological regions.

[0036] Global spectral skewness is the skewness of the centroids of the spectrum across all physiological regions.

[0037] Preferably, based on step S2, the spectral features in the time dimension include: the energy proportion of the breathing frequency band, the energy proportion of the micro-motion frequency band, the energy proportion of the action frequency band, and the time domain spectral entropy;

[0038] Among them, the energy proportion of the respiratory frequency band reflects the pressure fluctuations caused by breathing, the energy proportion of the micro-motion frequency band reflects the heartbeat and muscle micro-movements, the energy proportion of the movement frequency band reflects the vibration of turning over, and the time domain spectral entropy is used to distinguish between the stable sleep state and the disturbed sleep state.

[0039] The spectral characteristics of the posture dimension include: the energy proportion of the action band and the peak frequency of the spectrum;

[0040] Among them, the energy proportion of the action frequency band reflects the frequency domain energy of the head turning action, and the peak frequency of the spectrum is the frequency corresponding to the maximum amplitude of the power spectrum, which is used to determine the dominant head movement rhythm.

[0041] Preferably, the dual-track mechanism in step S2 is specifically as follows:

[0042] Fast matching channel: Pre-store various sleeping position frequency domain templates exclusive to users, calculate the cosine similarity between the current high-dimensional frequency domain feature vector and each template one by one, and if the highest similarity meets the preset matching threshold, directly output the corresponding sleeping position label and confidence score, skipping model inference to reduce computing power consumption;

[0043] Model inference channel: When the highest similarity does not reach the preset matching threshold, the lightweight optimized classification model is called to perform inference, output the probability distribution of various sleeping positions, take the highest probability category as the pre-classification result, and the corresponding probability value as the confidence level.

[0044] Preferably, based on step S4, the coherence coefficient is the cosine similarity between the pressure frequency domain feature vector and the attitude frequency domain feature vector; the weight ratio of pressure features, attitude features and device status is dynamically set according to the interval of the coherence coefficient; when the associated electronic device is on and the screen angle is within the preset interval, the fusion confidence of the specific sleeping posture type is weighted and improved; if the multi-source results are consistent, the candidate sleeping posture is directly output; if they are inconsistent, the candidate sleeping posture is output after secondary verification combined with the historical stable template.

[0045] Preferably, based on step S4, the trajectory determination rule is as follows:

[0046] When the monotonicity coefficient meets the preset transition threshold, the similarity sequence has obvious valleys and the fluctuation variance is lower than the preset smoothing threshold, and the candidate labels of multiple consecutive frames remain consistent, it is determined to be a real sleeping posture switch, the stable sleeping posture label is updated and output;

[0047] When the fluctuation variance is higher than the preset noise threshold and the monotonicity coefficient does not reach the transition threshold, it is judged as instantaneous noise interference, and the output of the previous stable sleeping position is maintained.

[0048] When the fluctuation variance is in the intermediate transition range, the monotonicity coefficient has not reached the transition threshold, and the candidate labels are not unique for multiple consecutive frames, it is determined that the posture is being adjusted. The stable sleeping posture will not be updated for the time being, the original output will be maintained, and monitoring will continue.

[0049] Preferably, the system also includes step S5, periodic iterative update and physiological back-inference. After completing the stability determination and outputting the effective stable sleeping position, the system performs periodic triggering, extracts the frequency domain feature dataset of the stable sleeping position within the period, counts the baseline changes of the frequency domain features of various sleeping positions, and combines user group labels to back-infer the physiological state, providing input for subsequent adaptive iteration.

[0050] User demographic tags include teenager tags, pregnant women tags, and general adult tags, with different external cycle update intervals for each tag.

[0051] Preferably, based on step S5, it also includes an inner loop and an outer loop: the inner loop is triggered by the first cycle or the first number of high-confidence samples accumulated, and uses an unsupervised clustering algorithm to generate various frequency domain templates of sleeping positions exclusive to users and optimize the classification model; the outer loop is triggered by the second cycle according to the user's group label, and reverses the physiological state and adaptively adjusts the template matching range, recognition sensitivity or classification threshold.

[0052] The technical effects and advantages of this invention are as follows:

[0053] 1. Using data from a flexible pressure array, a 6-axis gyroscope, and Bluetooth device status, the occipital surface array is divided into five physiological zones: the central posterior brain region, the left and right neck and shoulder regions, the anterior neck support region, and the posterior occipital buffer zone. Two-dimensional Fourier transform is performed on the pressure submatrix of each zone to extract spatial features, and one-dimensional FFT is performed on key points and IMU timing to extract time-domain and frequency-domain features. These are then fused into a standardized high-dimensional frequency-domain feature vector, which facilitates the differentiation of fine-grained scenarios.

[0054] 2. The cosine similarity of the time-domain and frequency-domain features of pressure and attitude in the same dimension is defined as the coherence coefficient. Based on this, the weights of pressure, attitude, and equipment state are dynamically divided. If multiple sources are consistent, the results are directly output. In case of conflict, a second verification is performed using historical stable templates, which weakens the interference signal and significantly improves the reliability of the fusion confidence.

[0055] 3. Input high-dimensional frequency domain vectors of multiple consecutive frames, calculate the cosine similarity of adjacent frames to form an evolution trajectory, and judge it with two indicators: monotonicity coefficient and fluctuation variance. A real switch must meet the conditions of monotonically decreasing, troughing, monotonically rising and having minimal variance, while noise has extremely large variance and is not monotonic. During the adjustment, the system maintains the previous stable label by default and only updates it when a real switch is confirmed, which ensures the accuracy of the adjustment.

[0056] 4. The internal circulation refreshes the user's sleeping posture-specific frequency domain template regularly with high-confidence samples through unsupervised clustering, automatically adapting to the user's unique sleeping posture habits and body shape characteristics, thus solving the problem of poor individual adaptability of the general model; the external circulation differentiates the cycle according to the adolescent / pregnant woman / general label, and back-infers the shoulder width growth, gestational age progress or weight fluctuation from the frequency domain baseline to achieve differentiated adjustment. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a flowchart of the multi-dimensional frequency domain feature extraction process of the present invention;

[0059] Figure 3 This is a flowchart illustrating the construction process of the multi-dimensional frequency domain feature vector in this invention.

[0060] Figure 4 This is a flowchart of the frequency domain feature classification process of the present invention;

[0061] Figure 5 This is a flowchart of the cross-modal fusion decision process of the present invention;

[0062] Figure 6 This is a flowchart of the frequency domain trajectory stability determination process of the present invention;

[0063] Figure 7 This is a flowchart of the periodic iterative update and physiological back-inference process of this invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] This invention provides, for example Figures 1-7 The sleeping posture recognition method based on frequency domain feature extraction shown includes the following steps:

[0066] S1 Multi-source Synchronous Acquisition and Physiological Partition Mapping: Synchronously acquire two-dimensional pressure data of the pillow surface, head posture motion data and related device status data, and divide the pressure data into pressure sub-matrices of each partition according to the preset human head and neck physiological partitions.

[0067] S2 Multi-dimensional Frequency Domain Feature Extraction and Vector Construction: Two-dimensional Fourier transform is performed on the pressure submatrix of each region to extract the spatial spectrum feature set. One-dimensional Fourier transform is performed on the pressure time series and attitude time series of the selected location to extract the time and attitude spectrum features. The above features are fused to generate a high-dimensional frequency domain feature vector.

[0068] S3 cross-modal fusion decision: Combines pre-classification results and confidence, frequency domain features of motion data and device status information to perform dynamic weight allocation and multi-source consistency judgment, calculates the total fusion confidence based on weights, and outputs candidate sleeping postures and fusion confidence;

[0069] S4 Frequency Domain Trajectory Stability Determination: Obtain high-dimensional frequency domain feature vectors corresponding to candidate sleeping positions in multiple consecutive frames, calculate the cosine similarity of frequency domain feature vectors of adjacent frames frame by frame, and fit to form a frequency domain feature evolution trajectory curve; by calculating the monotonicity coefficient and fluctuation variance of the trajectory, update the stable sleeping position label and output only when it is determined to be a real sleeping position change.

[0070] S5 Periodic Iterative Updates and Physiological Back-Inference: The system performs periodic triggering, extracts the frequency domain feature dataset of stable sleeping positions within the period, statistically analyzes the baseline changes of frequency domain features for various sleeping positions, and uses user group labels to back-infer physiological state.

[0071] Specifically, step S1 involves multi-source synchronous acquisition and physiological zone mapping.

[0072] The specific implementation process is as follows:

[0073] The flexible pressure sensor array uses thin-film sensors and is laid under the pillow support layer. The flexible pressure sensor array detects the pressure on the pillow and outputs the detected pressure data as a two-dimensional pressure matrix.

[0074] A 6-axis gyroscope is embedded inside the pillow to synchronously collect three-axis angular velocity and acceleration data of the head. Its sampling frequency is consistent with the sampling frequency of the flexible pressure sensor array. Preferably, it is uniformly set to 10Hz to match the time granularity of subsequent feature extraction and stability determination.

[0075] The screen angle and on / off status of the associated electronic device are read using a low-power Bluetooth module;

[0076] Using the MCU's local unified clock as a reference and the pressure sampling trigger signal as a synchronization trigger point, the 6-axis gyroscope data is linearly interpolated and aligned. The time error of the aligned data is ≤10ms, ensuring the temporal consistency of multi-source data. If the Bluetooth module data is lost or disconnected, the previous valid frame state is used to fill the gap. If the continuous disconnection exceeds 3 seconds, the device state weight is reset to 0 to eliminate the impact of that data on the fusion decision.

[0077] The original pressure matrix is ​​first subjected to global DC component removal to eliminate the interference of sensor substrate zero drift on subsequent frequency domain calculations.

[0078] Physiological partition space mapping

[0079] By dividing the flexible pressure sensor array into physiologically corresponding zones, the sensing points are mapped to the head and neck area of ​​the human body, specifically:

[0080] The central region of the back of the head is the central area of ​​the flexible pressure sensor array on the occipital surface, corresponding to the center of the back of the human head.

[0081] The left neck and shoulder area corresponds to the left shoulder and neck position of the flexible pressure sensor array on the pillow surface, which corresponds to the left side of a person's neck and shoulder.

[0082] The right neck and shoulder area corresponds to the right shoulder and neck position of the flexible pressure sensor array on the pillow surface, which corresponds to the right side of a person's neck and shoulder.

[0083] The anterior neck support area corresponds to the cervical spine at the anterior edge of the flexible pressure sensor array on the pillow surface, which corresponds to the front of the human neck.

[0084] The occipital buffer zone corresponds to the rear edge of the flexible pressure sensor array on the pillow surface, which is located in the area below the back of the head near the neck.

[0085] The partition coordinates are embedded in the MCU firmware. After data acquisition, each partition data can be directly cropped according to the corresponding coordinates.

[0086] Preferably, the flexible pressure sensor array uses a 32×32FSR thin-film sensor, and the flexible pressure sensor array is divided into physiologically corresponding partitions, specifically:

[0087] The central region of the back of the head is the central 8×8 area of ​​the 32×32 flexible pressure sensor array on the occipital surface, with rows 12-19 and columns 12-19.

[0088] The left neck and shoulder area corresponds to the left shoulder and neck position in the occipital surface array, rows 10-21, columns 0-7;

[0089] The right neck and shoulder area corresponds to the right shoulder and neck position in the occipital array, rows 10-21, columns 24-31;

[0090] The anterior cervical support area corresponds to the cervical vertebrae at the anterior edge of the occipital array, rows 0-7, columns 12-19;

[0091] The posterior occipital buffer is the corresponding position at the rear edge of the occipital array, row 24-31, column 12-19;

[0092] The data from the flexible pressure sensor array, the 6-axis gyroscope, and the Bluetooth module are time-aligned. Data exceeding the sensor's range is identified as abnormal data and removed. The abnormal values ​​are replaced with the mean of the neighborhood within the partition where the point is located. The gaps at the partition edges are filled in, and the frequency domain information of the original data is fully preserved.

[0093] Specifically, the MCU sorts the three data streams by timestamp, the pressure matrix is ​​stored directly, the gyroscope data is aligned using the aforementioned interpolation method, outliers are discarded after threshold comparison, and edge defects are filled by averaging the surrounding valid values.

[0094] The preset pillow surface partitions are mapped within a spatial range, the physical pillow position is transformed into the aforementioned partition sub-matrix, and the standardized mapped data is sent to the frequency domain feature extraction module.

[0095] Preferably, standardized data is only sent to the frequency domain feature extraction module when the proportion of effective data in the whole frame is ≥90%; if the effective proportion is insufficient, it is judged as a low-quality frame and discarded directly, and the recognition result of the previous effective frame is used.

[0096] Specifically, the standardized partitioned pressure matrix, IMU time series and device status data are output and pushed to the frequency domain feature extraction module. The 32×32 matrix is ​​divided into 5 sub-matrices according to the fixed coordinates, the gyroscope sequence is stored as an array, and the Bluetooth status is converted into a flag bit. All of these are then input to the frequency domain feature extraction module.

[0097] Specifically, step S2 involves multi-dimensional frequency domain feature extraction and vector construction.

[0098] The corresponding physiological partitions of the flexible pressure sensor array are input with a standardized two-dimensional pressure matrix, time series of key points in the neck and shoulder area, and time series of angular velocity from a 6-axis gyroscope.

[0099] For each physiological region, spatial spectral features are extracted. First, a two-dimensional Fourier transform is performed on the pressure submatrix of each region to obtain the power spectrum. Then, calculate the following indicators from the spectrum, specifically including:

[0100] Spectral centroid: The weighted average spatial frequency of the power spectrum of this region is calculated using the following formula:

[0101]

[0102] Where C is the spectral centroid, Two-dimensional frequency coordinates The power spectral value at that location, The radial spatial frequency r reflects the spatial variation scale of pressure distribution; it is lower when lying supine with a gradual change and higher when lying on one's side with a steep change.

[0103] Spectral entropy: Calculated after normalizing the power spectrum, reflecting the regularity of the pressure pattern, indicating stable sleep with low pressure and easy to roll over with high pressure.

[0104] Spectral kurtosis: The fourth moment of the power spectrum is calculated, reflecting the degree of peaks in the spectrum; it is higher in lateral narrow-band high-pressure systems.

[0105] Extremely low frequency energy ratio: The ratio of spatial frequency radius band energy to total energy, corresponding to large areas of slowly varying pressure;

[0106] High-frequency energy percentage: The energy percentage of the spatial frequency radius ring band, corresponding to noise or jitter;

[0107] Peak amplitude of the spectrum: The maximum value of the power spectrum after excluding DC, reflecting the intensity of the main space mode;

[0108] By fusing the spatial features of multiple physiological regions, a spatial spectral feature set is extracted, and the spectral difference and symmetry features of the regions are calculated.

[0109] Based on the spectral characteristics of each partition, calculate the higher-order difference features:

[0110] Symmetry of the centroid of the left and right neck and shoulder regions: The absolute value of the ratio or difference of the centroid of the left and right neck and shoulder regions is normalized to distinguish between supine and lateral lying positions.

[0111] Energy symmetry in the left and right neck and shoulder areas: the ratio of extremely low-frequency energy in the left and right areas, quantifying the degree of symmetry;

[0112] Energy difference in the anterior and posterior regions: The energy difference between the anterior neck support area and the posterior occipital buffer zone is used to identify forward tilting or backward tilting.

[0113] Central and peripheral spectral attenuation slope: The difference between the centroid of the central region of the posterior brain and the centroid of the occipital buffer zone, divided by the distance, reflects the smoothness of pressure transition;

[0114] Radial odd-even ring energy ratio: The energy ratio of odd-even ring bands based on spatial frequency radius, replacing the traditional harmonic concept;

[0115] Even-order radial ring energy ratio: the ratio of energy of even-numbered rings with even spatial frequency radii to total energy, replacing the traditional harmonic concept;

[0116] Global spectral variance: The variance of the centroids of the spectrum across all partitions;

[0117] Global spectral skewness: The skewness of the centroids of the spectrum across all partitions;

[0118] The aforementioned differential characteristics quantify the symmetry and smoothness of pressure distribution, which are the core features for distinguishing subtle sleeping postures.

[0119] One-dimensional FFT transformation in the time domain:

[0120] Key sampling points were selected in the left and right neck and shoulder areas. The four points with the highest average pressure in each area were selected as key sampling points. Their average time series was used as input to reduce the influence of single-point noise. The most recent continuous time series was selected and zeros were padded to 128 points at the end to adapt to the radix-2 FFT algorithm. The computer fast Fourier transform requires the data length to be an integer power of 2. Zero padding improves frequency resolution but does not increase real information.

[0121] After superimposing a one-dimensional Hanning window onto a weighting function that reduces spectral leakage, a 128-point one-dimensional FFT is performed to obtain the power spectrum of 64 effective frequency points with a frequency resolution of 0.1 Hz.

[0122] Extracting 4D time-spectral features:

[0123] Energy percentage of respiratory frequency band: reflects pressure fluctuations caused by breathing. Preferably, the respiratory frequency band is set to 0.15-0.6Hz.

[0124] Micro-motion frequency band energy ratio: reflects the heartbeat or muscle micro-movement. Preferably, the micro-motion frequency band is set to 0.6-3Hz.

[0125] Energy percentage of the movement frequency band: This reflects the vibration during turning over. Preferably, the movement frequency band is set to 3-5Hz.

[0126] Time-domain spectral entropy: Based on the information entropy after power spectrum normalization, it distinguishes between stability and disturbance;

[0127] Attitude frequency domain feature extraction:

[0128] Perform the same one-dimensional FFT transformation on the continuous time sequence of the IMU's three-axis angular velocity summation to extract attitude spectrum features:

[0129] Energy proportion of the movement frequency band: reflects the frequency domain energy of the head turning movement;

[0130] Peak frequency of the power spectrum: The frequency corresponding to the maximum amplitude of the power spectrum, used to determine the dominant motion rhythm;

[0131] The energy proportions of the respiratory and micro-motion frequency bands in the attitude spectrum are extracted simultaneously and cross-validated with the pressure-time characteristics;

[0132] Based on the matching of pose spectrum features with the preset head pose template, the pre-classification result of head pose and the corresponding confidence score are output as one input for cross-modal fusion.

[0133] The above-mentioned spatial features, difference features, temporal features, and attitude features are concatenated into the original feature vector; the Z-score normalization method is used to normalize each dimension to eliminate the difference in dimensions, and finally output a high-dimensional frequency domain feature vector.

[0134] Output high-dimensional frequency domain feature vectors and periodically iteratively updated model parameters to quickly predict confidence and calculate whether the cosine similarity between the current feature and the template is greater than a threshold.

[0135] When a new user uses the service for the first time, a general population sleeping posture frequency domain template is loaded as the initial template. After the first iteration of the inner loop is completed, it is replaced with a user-specific template.

[0136] When the similarity is greater than the threshold, the matching sleeping posture label and its high confidence are directly output and pushed to the cross-modal fusion decision module;

[0137] When the similarity is less than the threshold, the probability of various sleeping positions is calculated by periodically iteratively updating the model parameters and model inference, the probability distribution of sleeping position type is output, the category with the highest probability is taken as the pre-classification result, the label and confidence are output, and pushed to the cross-modal fusion decision module.

[0138] Typical example of frequency domain fingerprint of sleeping posture:

[0139] In the supine position, low-frequency energy is concentrated in the central area of ​​the posterior brain, and the symmetry between the left and right sides is >0.9;

[0140] In the supine and backward posture, the energy proportion of the posterior head buffer increases, the energy difference between the front and back shifts negatively, and the low frequency in the central region still dominates, but the center of mass shifts slightly backward.

[0141] The peak spectral value in the left neck and shoulder area was significantly increased in the standard left lateral decubitus position, with a left-right symmetry of <0.3;

[0142] The peak spectral value of the right neck and shoulder area was significantly increased in the standard right lateral decubitus position, with a left-right symmetry of <0.3;

[0143] When lying on your side with your head tilted forward to look at your phone, the proportion of high-frequency energy in the anterior neck support area increases, and the energy difference between the front and back shifts positively.

[0144] In the lateral recumbent cubital position, the spectral peak in the unilateral neck and shoulder area is significant and the energy in the anterior neck support area is moderate, showing an overall cubital characteristic with a left-right symmetry of <0.4.

[0145] In a semi-reclining posture, the energy distribution of each zone is balanced, the energy difference between front and back approaches zero, the centroid of the spectrum is centered, and the global spectrum variance is small.

[0146] In the prone position, the proportion of high-frequency energy in the anterior neck support area increases and there is also contact in the central area of ​​the back of the head, resulting in a negative shift in the energy difference between the front and back.

[0147] Prone tendency is defined as a semi-prone posture with the chest and above lying face down, which is different from a fully prone posture; various sleeping posture labels are set with mutually exclusive judgment rules, and only one label with the highest probability is output per frame.

[0148] The frequency domain fingerprints of different sleeping positions show significant differences and have extremely high distinguishability.

[0149] Preliminary sleeping posture classification is achieved based on frequency domain fingerprinting. A dual-track mechanism of fast template matching and fine-grained model inference is adopted to balance accuracy and computational efficiency. The specific implementation process is as follows:

[0150] Fast matching channel: Pre-store user-specific sleeping posture frequency domain templates, and calculate the cosine similarity between the current frequency domain fingerprint and the template one by one;

[0151] If the highest similarity is ≥0.9, the corresponding sleeping posture label and confidence score are directly output, skipping model inference and reducing computing power consumption;

[0152] When the similarity is less than 0.9, the classification model is called for inference; preferably, the pruned and optimized LightGBM classification model is called for inference. The model has 15 decision trees, a maximum depth of 6, ≤16 leaf nodes per tree, a model size of <50KB, and a single inference time of <8ms when deployed on an STM32L4 microcontroller.

[0153] The model takes a frequency domain fingerprint vector as input and outputs a probability distribution of sleeping posture. The category with the highest probability is taken as the pre-classification result, and the corresponding probability value is taken as the confidence level.

[0154] The sleep posture system specifically includes: standard supine, supine with back extension, standard left lateral decubitus, standard right lateral decubitus, side-lying with forward tilt while looking at a phone, side-lying with arms bent, semi-reclining with support, and prone tilt, covering all sleep postures.

[0155] The pre-classified labels and corresponding confidence scores are pushed to the cross-modal fusion decision module.

[0156] Specifically, step S3 involves cross-modal fusion decision.

[0157] Input the pre-classification results and confidence scores of sleeping postures, the frequency domain feature judgment results of postures, and the status of associated electronic devices. Perform dynamic weight allocation, perform weighted calculation and fusion of total confidence scores, and judge whether the multi-source results are consistent.

[0158] The coherence coefficient is defined as the cosine similarity between the pressure frequency domain feature vector and the attitude frequency domain feature vector, which quantifies the consistency between the pressure and attitude frequency domains.

[0159] When calculating the coherence coefficient, only the time-domain spectral features of the same dimension in pressure and attitude features are used for cosine similarity calculation to ensure that the dimensional matching is consistent with the physical meaning.

[0160] When the screen of the associated electronic device is on and the screen angle is between 30° and 90°, it is determined to be in use. The confidence of the two sleeping postures of lying on the side and leaning forward to look at the phone and half-lying and leaning is increased by 15%; no additional weighting is applied when the screen is off.

[0161] When the multi-source results are consistent, the candidate sleeping posture and fusion confidence are output. The candidate sleeping posture is equal to the pre-classification result, and the confidence is equal to the weighted value. The result is then pushed to the stability determination module.

[0162] When the results from multiple sources are inconsistent, the frequency domain similarity between the multi-source data and the historical stable template is calculated, and the result with higher similarity is taken as the candidate sleeping position. Finally, the pre-fusion confidence of the candidate sleeping position is output and pushed to the stability judgment module.

[0163] If the results from multiple sources conflict pairwise, the one with higher confidence between the pressure pre-classification and the attitude pre-classification is selected, and the candidate result is output after secondary verification by combining it with the historical template.

[0164] Preferably, the multi-source weights are dynamically set based on the coherence coefficient:

[0165] Coherence coefficient ≥ 0.7: Pressure frequency domain feature weight set to 75%, attitude feature weight set to 15%, and equipment status weight set to 10%;

[0166] Coherence coefficient 0.4-0.7: Maintain default weights, pressure characteristics 60%, attitude characteristics 30%, equipment status 10%;

[0167] Coherence coefficient < 0.4: Increase attitude feature weight to 60%, decrease pressure feature weight to 30%, and equipment status weight to 10% to weaken the influence of interference signals.

[0168] Then, the multi-source results are judged: if the pressure pre-classification result, posture determination result, and device status indication are consistent, the candidate sleeping posture is directly output, and the confidence level is the value calculated according to the above weights, and pushed to the stability determination module.

[0169] Specifically, step S4 involves determining the stability of the frequency domain trajectory.

[0170] If there are conflicts among the results from multiple sources, the frequency domain template of the user's historical stable sleeping posture is called, the cosine similarity between the current feature and each template is calculated, and the result with the higher similarity is adopted as the candidate sleeping posture to avoid recognition jumps caused by instantaneous interference.

[0171] Output: Output candidate sleeping posture labels and fusion confidence scores, and push them to the stability determination module.

[0172] Stability determination is performed in the frequency domain, and the specific implementation process is as follows:

[0173] Input the frequency domain features corresponding to subsequent sleeping postures for continuous duration, calculate the cosine similarity between adjacent frame domain features, and fit to form the frequency domain feature evolution trajectory curve;

[0174] Trajectory construction: Take the frequency domain fingerprints corresponding to the candidate sleeping postures of 5 consecutive frames, calculate the cosine similarity between adjacent frames, and fit to form the frequency domain feature evolution trajectory curve;

[0175] Preferably, a 5-frame sliding window is used, corresponding to a duration of 0.5 seconds, to cover the shortest transition cycle of a normal human turning over, balancing recognition response speed and anti-interference ability.

[0176] Trajectory Pattern Quantification: Calculating Two Core Indicators:

[0177] Monotonicity coefficient: Characterizes the directional consistency of trajectory changes. The actual sleeping posture switching shows a transitional pattern of monotonically decreasing similarity, reaching a trough, and monotonically rebounding. It is obtained by calculating the first-order difference of the similarity sequence with the same sign.

[0178] Fluctuation variance: Characterizes the degree of random jitter of the trajectory. Instantaneous noise interference manifests as random jumps in similarity. The variance of the similarity sequence is calculated, and if the fluctuation variance is significantly higher than the threshold, it is judged as noise.

[0179] Specifically, based on the 44-dimensional frequency domain fingerprint vector output from the upstream source that has been normalized by Z-score, the specific implementation process is as follows:

[0180] First, input the frequency domain fingerprint data corresponding to 5 consecutive frames of candidate sleeping postures, corresponding to 0.5 seconds at a system sampling rate of 10Hz, denoted as . , , , , , where t is the frame index at the current time; This represents the 44-dimensional frequency domain fingerprint vector corresponding to the t-th frame. The frequency domain fingerprint vector is obtained by concatenating and standardizing spatial spectral features, difference features, temporal spectral features, and attitude spectral features.

[0181] The cosine similarity between adjacent frames is calculated pairwise, using the following formula:

[0182]

[0183]

[0184]

[0185]

[0186] in, The cosine similarity operator is calculated by dividing the dot product of two vectors by their magnitudes, and is used to measure the similarity of the directions of frequency domain fingerprints.

[0187] (i=0, 1, 2, 3) represents the i-th similarity value obtained, with a value range of [−1, 1], and approaches 1 under stable sleeping posture;

[0188] The similarity sequence is obtained from the above calculation. The sequence is fitted sequentially to form the frequency domain feature evolution trajectory curve.

[0189] Next, trajectory morphology quantification is performed, and the following two core indicators are calculated:

[0190] Monotonicity coefficient:

[0191] Calculate the first difference of sequence S:

[0192] (i=0, 1, 2)

[0193] in Represents the similarity difference between the i-th and i-th adjacent pairs. Perform symbol segmentation: segment adjacent symbols with the same symbol. If a segment is grouped into one segment, and the signs are reversed, a new segment is created. Let the total number of segments be k, and the monotonicity coefficient be defined as:

[0194]

[0195] in:

[0196] k is The number of symbolic segments;

[0197] The denominator 3 is The normalized benchmark with the maximum possible number of segments has only 3 differences, so the maximum k=3.

[0198] M is used to characterize the directional consistency of trajectory changes.

[0199] The actual sleeping posture transitions exhibit a two-stage concave shape: a monotonous decrease in similarity, reaching a trough, and then a monotonous rebound. At this point, k=2 and M=0.75.

[0200] Set the judgment threshold to M≥0.7, and require that there be a valley value in S. <0.88, here For any element in sequence S, 0.88 is the statistical median level of the switching trough.

[0201] Volatility variance:

[0202] The overall variance of the similarity sequence S is calculated using the following formula:

[0203]

[0204] in: The variance represents the degree of random jitter in the trajectory; Let be the i-th element of sequence S; mean(S) is the arithmetic mean of the four elements of sequence S, i.e. The coefficient 1 / 4 is because the sequence contains all 4 samples and adopts the population variance caliber; ∑ represents the summation of the corresponding terms from i=0 to 3.

[0205] The aforementioned variance threshold was determined through offline calibration before shipment: Actual measured data was collected and divided into three sample groups—noise pile, adjustment pile, and smooth switching pile—based on accelerometer data and manual annotation. Data from each pile was then collected. Using the 95th quantile as the boundary, we obtain: noise pile ≥0.006, adjust the heap size to 0.003≤ <0.006 and M<0.7, smooth heap switching <0.003 and M≥0.7.

[0206] Users can reassess this threshold every 30 days based on their historical high-confidence samples during the medium-term period, with each adjustment not exceeding ±10% of the initial value;

[0207] When M≥0.7 and 0.003≤σ²<0.006, it is judged as a low-confidence sleep position switch, and the observation needs to be extended to 8 consecutive frames. The monotonicity condition needs to be met continuously before the update is confirmed.

[0208] Added bed-off detection: When the mean of the global pressure matrix is ​​lower than the sensor noise floor for more than 2 seconds, it is determined that the user has left the bed, and sleeping posture recognition and template updates are stopped until the pressure signal recovers to above the effective threshold.

[0209] Based on the above indicators, the trajectory type is determined as follows:

[0210] A smooth, monotonous transition reflects realistic sleeping posture changes: If the continuous similarity is stable and the expected sleeping posture change pattern is observed, the sleeping posture is confirmed as valid. The stable sleeping posture label is updated, output to the downstream system, and simultaneously stored in the historical feature database. Preferably, the judgment condition is M ≥ 0.7 and the presence of... <0.88 and <0.003, and if the candidate labels are consistent for 3 consecutive frames, the sleeping posture is confirmed to be valid. The stable sleeping posture label is updated, output to the downstream system, and stored in the historical feature database simultaneously.

[0211] Random sudden changes or transient noise: these are considered noise, and the previous stable sleeping position is maintained, along with the output data for the original stable sleeping position; preferably, the determination condition is... If the value is ≥0.006 and M<0.7, it is determined to be noise interference, and the previous stable sleeping position and its output data are maintained.

[0212] During the reciprocating fluctuation and posture adjustment process: When the device is in the posture adjustment phase, it temporarily stops updating output data, performs continuous monitoring, and maintains the original stable sleeping posture while outputting data; preferably, the judgment condition is 0.003 ≤ If the value is less than 0.006 and M is less than 0.7, and the candidate labels are not unique for 5 consecutive frames, the output data will not be updated for the time being. The monitoring will continue, and the original stable sleeping posture output data will be maintained.

[0213] Specifically disclosed, the monotonicity coefficient M and the volatility variance mentioned above... The calculations are only for preferred embodiments.

[0214] The system outputs the stable sleeping posture label of the previous frame by default, and only updates the output when the valid switching conditions are met, so as to avoid frequent changes in the results.

[0215] Specifically, step S5 involves iterative updates and physiological back-calculation.

[0216] After completing the stability determination and outputting the effective stable sleeping position, the system will periodically trigger to extract the frequency domain feature dataset of the stable sleeping position within the period, count the baseline changes of the frequency domain features of various sleeping positions, and combine the user group labels to infer the physiological state, providing input for subsequent adaptive iteration;

[0217] Before using the pillow, user demographics are identified:

[0218] Adolescent tag: By analyzing the growth trend of the spectral width of the side-sleeping zone and the changes in the spacing of pressure peaks, the growth of shoulder width and changes in cervical curvature are inferred, and the matching range of the side-sleeping spectral template is expanded simultaneously, and the feature boundaries of supine support are updated to avoid the decline in recognition accuracy during the growth and development period.

[0219] Pregnancy tag: By analyzing the changes in the proportion of side-lying and curled-up sleeping positions and the shift in pressure spectrum, the degree of abdominal protrusion and gestational age progression can be inferred, automatically improving the recognition sensitivity and feature weight of this type of sleeping position;

[0220] General labeling: By inferring weight and body shape fluctuations through changes in global spectral energy baseline, the classification threshold is optimized;

[0221] A two-layer closed loop is constructed, consisting of internal circulation for individual adaptation and external circulation for physiological adaptation. The specific implementation process is as follows:

[0222] Triggering mechanism: The inner loop is triggered once every 7 days or after accumulating 1000 high-confidence samples; the outer loop has a different cycle set according to user group tags: 30 days for teenager mode, 14 days for pregnant women mode, and 90 days for general adult mode.

[0223] Inner loop: Unsupervised frequency domain self-learning

[0224] High-confidence stable sleeping posture samples with a fusion confidence score ≥ 0.85 are extracted within the period. The K-means unsupervised clustering algorithm is used to cluster and update the frequency domain features of each sleeping posture to generate user-specific frequency domain templates for 8 sleeping postures. At the same time, the feature weights and classification thresholds of the classification model are optimized based on the sample distribution.

[0225] The internal loop requires no manual annotation and automatically adapts to the user's unique sleeping posture and body shape, solving the problem of poor individual adaptability of the general model.

[0226] External circulation: physiological evolution and adaptation

[0227] The frequency domain feature baselines of various sleeping positions within the period are extracted, the long-term evolution trend is analyzed, and the physiological state is inferred by combining user group tags.

[0228] A sleeping posture recognition system based on frequency domain feature extraction includes:

[0229] The sensing module is configured to collect pillow surface pressure data, head movement data, and status data of associated electronic devices;

[0230] The mapping module is configured to perform spatial mapping and standardization of pressure data according to physiologically corresponding partitions.

[0231] The feature extraction module is configured to perform two-dimensional and one-dimensional Fourier transforms to extract spatial and temporal frequency domain features, fuse and standardize them into a high-dimensional frequency domain feature vector, and output the sleeping posture pre-classification results.

[0232] The fusion decision module is configured to dynamically allocate weights and output candidate sleeping positions and fusion confidence scores by fusing multi-source data.

[0233] The stability determination module is configured to determine stable sleeping posture labels based on the evolution trajectory of frequency domain features in continuous frames.

[0234] The update module is configured to periodically update user-specific templates and inverse physiological states.

[0235] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A sleeping posture recognition method based on frequency domain feature extraction, characterized in that, Includes the following steps: S1 Multi-source Synchronous Acquisition and Physiological Partition Mapping: Synchronously acquire two-dimensional pressure data of the pillow surface, head posture motion data and related device status data, and divide the pressure data into pressure sub-matrices of each partition according to the preset human head and neck physiological partitions. S2 Multi-dimensional Frequency Domain Feature Extraction and Vector Construction: Two-dimensional Fourier transform is performed on the pressure submatrix of each region to extract the spatial spectrum feature set. One-dimensional Fourier transform is performed on the pressure time series and attitude time series of the selected location to extract the time and attitude spectrum features. The above features are fused to generate a high-dimensional frequency domain feature vector. S3 cross-modal fusion decision: Combines pre-classification results and confidence, frequency domain features of motion data and device status information to perform dynamic weight allocation and multi-source consistency judgment, calculates the total fusion confidence based on weights, and outputs candidate sleeping postures and fusion confidence; S4 Frequency Domain Trajectory Stability Determination: Obtain high-dimensional frequency domain feature vectors corresponding to candidate sleeping postures in multiple consecutive frames, calculate the cosine similarity of frequency domain feature vectors of adjacent frames frame by frame, and fit to form a frequency domain feature evolution trajectory curve; by calculating the monotonicity coefficient and fluctuation variance of the trajectory, update the stable sleeping posture label and output it only when it is determined to be a real sleeping posture change.

2. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S1, multiple physiological regions specifically include: The central area of ​​the back of the head corresponds to the central calibration area of ​​the sensor array, which corresponds to the middle part of the back of the human head. The left neck and shoulder area corresponds to the shoulder and neck calibration area on the left side of the sensor array, which corresponds to the left side of the human neck and shoulder. The right neck and shoulder area corresponds to the neck and shoulder calibration area on the right side of the sensor array, which corresponds to the right side of the neck and shoulder of the human body. The anterior neck support area corresponds to the cervical spine calibration area at the front edge of the sensor array, which corresponds to the front part of the human neck. The occipital buffer zone corresponds to the cervical fossa calibration area at the rear edge of the sensor array, which corresponds to the area below the back of the head near the cervical fossa. The coordinate range of each partition is pre-calibrated based on the physiological contact distribution of the human head and neck and is fixed in the MCU firmware. After the data acquisition is completed, the data of each partition is directly cut according to the calibrated coordinates.

3. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S2, the basic characteristics of the spatial spectrum include: spectral centroid, spectral entropy, spectral kurtosis, energy proportion of low spatial frequency bands, energy proportion of high spatial frequency bands, and spectral peak amplitude. Among them, the centroid of the spectrum is the weighted average spatial frequency of the power spectrum, which is used to reflect the spatial scale of pressure distribution; Spectral entropy is the information entropy of the normalized power spectrum, used to reflect the regularity of the pressure distribution pattern; Spectral kurtosis is the fourth moment of the power spectrum, used to reflect the degree of peaks in the spectrum; The energy proportion in the low spatial frequency band is the ratio of energy within a preset radial frequency range to the total energy, corresponding to a large-area, slowly varying pressure distribution; The high spatial frequency band energy ratio is the ratio of energy in the preset high-frequency radial range to the total energy, corresponding to disturbance and noise signals; The peak amplitude of the spectrum is the maximum value of the power spectrum after excluding the DC component, and is used to reflect the intensity of the main space pressure mode.

4. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S2, the higher-order features of spectral difference and symmetry include: Symmetry of the centroid of the left and right neck and shoulder regions is the normalized result of the difference between the centroids of the left and right neck and shoulder regions, used to distinguish between supine and lateral positions. The energy symmetry of the left and right neck and shoulder areas is the ratio of the proportion of energy in the low spatial frequency band of the left and right neck and shoulder areas, which is used to quantify the degree of symmetry of the pressure distribution on the left and right sides. The spectral energy difference between the front and rear regions is the difference in spectral energy between the anterior neck support area and the posterior occipital buffer zone, which is used to identify the forward or backward head posture. The central and peripheral spectral attenuation slope is the difference in spectral centroid between the central region of the posterior brain and the occipital buffer zone divided by the regional calibration interval, used to reflect the smoothness of the pressure distribution transition. The radial odd-even ring energy ratio is the energy ratio value of the odd-even ring bands divided according to the spatial frequency radius; The energy percentage of even-order radial rings is the ratio of the energy of even-numbered radius rings to the total energy. The global spectral variance is the variance of the centroids of the spectral frequencies of all physiological regions. Global spectral skewness is the skewness of the centroids of the spectrum across all physiological regions.

5. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S2, the spectral features in the time dimension include: the energy proportion of the breathing frequency band, the energy proportion of the micro-motion frequency band, the energy proportion of the action frequency band, and the time domain spectral entropy; Among them, the energy proportion of the respiratory frequency band reflects the pressure fluctuations caused by breathing, the energy proportion of the micro-motion frequency band reflects the heartbeat and muscle micro-movements, the energy proportion of the movement frequency band reflects the vibration of turning over, and the time domain spectral entropy is used to distinguish between the stable sleep state and the disturbed sleep state. The spectral characteristics of the posture dimension include: the energy proportion of the action band and the peak frequency of the spectrum; Among them, the energy proportion of the action frequency band reflects the frequency domain energy of the head turning action, and the peak frequency of the spectrum is the frequency corresponding to the maximum amplitude of the power spectrum, which is used to determine the dominant head movement rhythm.

6. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, The dual-track mechanism in step S2 is specifically as follows: Fast matching channel: Pre-store various sleeping position frequency domain templates exclusive to users, calculate the cosine similarity between the current high-dimensional frequency domain feature vector and each template one by one, and if the highest similarity meets the preset matching threshold, directly output the corresponding sleeping position label and confidence score, skipping model inference to reduce computing power consumption; Model inference channel: When the highest similarity does not reach the preset matching threshold, the lightweight optimized classification model is called to perform inference, output the probability distribution of various sleeping positions, take the highest probability category as the pre-classification result, and the corresponding probability value as the confidence level.

7. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S4, the coherence coefficient is the cosine similarity between the pressure frequency domain feature vector and the attitude frequency domain feature vector; the weight ratio of pressure features, attitude features and device status is dynamically set according to the interval of the coherence coefficient; when the associated electronic device is on and the screen angle is within the preset interval, the fusion confidence of the specific sleeping posture type is weighted and improved; if the multi-source results are consistent, the candidate sleeping posture is directly output; if they are inconsistent, the candidate sleeping posture is output after secondary verification combined with the historical stable template.

8. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, Based on step S4, the trajectory determination rules are as follows: When the monotonicity coefficient meets the preset transition threshold, the similarity sequence has obvious valleys and the fluctuation variance is lower than the preset smoothing threshold, and the candidate labels of multiple consecutive frames remain consistent, it is determined to be a real sleeping posture switch, the stable sleeping posture label is updated and output; When the fluctuation variance is higher than the preset noise threshold and the monotonicity coefficient does not reach the transition threshold, it is judged as instantaneous noise interference, and the output of the previous stable sleeping position is maintained. When the fluctuation variance is in the intermediate transition range, the monotonicity coefficient has not reached the transition threshold, and the candidate labels are not unique for multiple consecutive frames, it is determined that the posture is being adjusted. The stable sleeping posture will not be updated for the time being, the original output will be maintained, and monitoring will continue.

9. The sleeping posture recognition method based on frequency domain feature extraction according to claim 1, characterized in that, It also includes step S5, periodic iterative update and physiological back-inference. After completing the stability determination and outputting the effective stable sleeping position, the system will trigger periodically, extract the frequency domain feature dataset of the stable sleeping position within the period, count the baseline changes of the frequency domain features of various sleeping positions, and combine the user group labels to back-infer the physiological state, providing input for subsequent adaptive iteration. User demographic tags include teenager tags, pregnant women tags, and general adult tags, with different external cycle update intervals for each tag.

10. The sleeping posture recognition method based on frequency domain feature extraction according to claim 9, characterized in that, Based on step S5, it also includes an inner loop and an outer loop: the inner loop is triggered by the first cycle or the first number of high-confidence samples accumulated, and uses an unsupervised clustering algorithm to generate various frequency domain templates of sleeping positions exclusive to users and optimize the classification model; the outer loop is triggered by the second cycle based on the user's group label, infers the physiological state and adaptively adjusts the template matching range, recognition sensitivity or classification threshold.

Images