Sleeping posture recognition method based on dynamic air pressure of air bag mattress

Abstract

The invention belongs to the field of mattresses, and particularly relates to a sleeping posture recognition method based on dynamic air pressure of an air bag mattress. In order to overcome the defect of low accuracy of the existing sleeping posture recognition method of the air bag mattress, the invention adopts the following technical scheme: the sleeping posture recognition method based on the dynamic air pressure of the air bag mattress comprises the following steps of: collecting air pressure values of air pressure sensors in various subareas in various action processes of a subject under a preset sampling frequency, and recording related data; calculating a time characteristic value and a frequency characteristic value of the air pressure dynamic change; the feature values are grouped according to partitions, all the feature values in the same group are reserved only when coefficients of all the feature values in the same group are not equal to zero, otherwise all the feature values in the same group are removed, and region features of group strong association are reserved in each type of action; and calculating a mean value and a standard deviation of the feature values in the reserved feature group. According to the sleeping posture recognition method based on the dynamic air pressure of the air bag mattress, a unique feature screening method is adopted, and the recognition accuracy can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of mattress, and particularly relates to a sleep posture recognition method based on dynamic air pressure of an air bag mattress. BACKGROUND

[0002] Sleep posture recognition is the premise of adaptive support adjustment of intelligent mattress. The current sleep posture recognition technology in the field of intelligent mattress mainly focuses on pressure sensing, static feature analysis and multi-sensor fusion, and the specific technical path and limitations are as follows:

[0003] 1. Pressure sensor array technology

[0004] 1.1 Principle: densely arrange pressure sensors (such as thin film pressure sensors) inside the mattress, acquire the pressure distribution image of the human body contacting the mattress, and match and recognize the sleep posture by combining the pre-set sleep posture pressure template (such as high hip and waist pressure when lying flat, and high unilateral shoulder and hip pressure when lying on one side).

[0005] 1.2 Disadvantages:

[0006] High cost: dozens or even hundreds of sensors are required, and the hardware and wiring costs increase significantly, making it difficult to mass-produce;

[0007] Poor reliability: the sensors are easily affected by mattress deformation and sweat corrosion, and have a high failure rate during long-term use;

[0008] Accuracy is limited by the comfort layer: thick comfort layers (such as memory foam and latex) will buffer the pressure, resulting in blurred pressure distribution images and low sleep posture recognition accuracy for medium-weight users (60-80 kg);

[0009] 2. Static air pressure analysis technology

[0010] 2.1 Principle: use the air pressure sensor of the existing zoned air bag of the mattress to acquire the static air pressure value of each zone when the user is static, and determine the sleep posture by the air pressure size relationship (such as the left shoulder and hip zone air pressure being more than 20% higher than the right side when lying on one side).

[0011] 2.2 Disadvantages:

[0012] Low dynamic sensitivity: the static air pressure changes little before and after turning over, which is much lower than the sensor noise threshold, resulting in missed turning events;

[0013] Poor individual adaptability: the static air pressure baseline values of users with different weights (such as 40 kg and 100 kg) differ greatly, and the template needs to be recalibrated for each user, resulting in poor user experience;

[0014] Cannot distinguish between body movement and turning over: small body movements (such as hand and foot movements) may cause static air pressure fluctuations.

[0015] 3. Simple dynamic threshold technology

[0016] 3.1 Principle: Set a fixed air pressure change threshold (e.g., an air pressure change exceeding 0.5 kPa within 5 seconds is considered turning over), and trigger the turning over event recognition through the threshold.

[0017] 3.2 Shortcomings:

[0018] Poor threshold universality: The reasonable threshold varies greatly among users with different initial airbag pressures (such as after adjusting the firmness) and different weights, making it impossible to set a unified threshold.

[0019] High false positive rate: Large movements (such as sitting up and then lying down) will trigger the threshold and be falsely identified as turning over. Summary of the Invention

[0020] This invention addresses the shortcomings of existing sleeping posture recognition methods for airbag mattresses, which have low accuracy, by providing a sleeping posture recognition method based on the dynamic air pressure of the airbag mattress, thereby improving the recognition accuracy.

[0021] To achieve the above objectives, the present invention adopts the following technical solution: a sleeping posture recognition method based on dynamic air pressure of an airbag mattress. The airbag mattress has multiple zones, each zone being equipped with an air pressure sensor. The sleeping posture recognition method based on dynamic air pressure of the airbag mattress includes:

[0022] Step S1: Collect the air pressure values ​​of the air pressure sensor in each zone during the subject's various actions at a preset sampling frequency, record the time point of the maximum air pressure value, and the total duration of each type of action is random.

[0023] Step S2: Calculate the time characteristic value and frequency characteristic value of the dynamic change of air pressure for each subject in each type of action;

[0024] Step S3: Group the feature values ​​according to the partition. Only when the coefficients of all features in the same group are not equal to 0, are all feature values ​​in the same group retained; otherwise, all feature values ​​in the same group are removed. q is retained for each action category. c Strongly correlated regional characteristics;

[0025] Step S4, in q c Calculate the mean μ_q of p feature values ​​in a group feature. c and standard deviation σ_q c ;

[0026] Step S5: After the user lies down, monitor the air pressure values ​​of each zone within the mattress in real time. Calculate the air pressure characteristic value of each zone within the window time t, and determine whether the air pressure characteristic value is within [μ_q]. c -θ*σ_q c Within the range of θ, if so, it is determined that a corresponding rolling or body movement event occurred within the window time, where θ is the percentile of the normal distribution.

[0027] The present invention relates to a sleeping posture recognition method based on dynamic air pressure of an airbag mattress. This method collects air pressure data from each zone during various actual movements of the subject, calculates the time and frequency characteristic values ​​of dynamic air pressure changes, and groups these characteristic values ​​according to zone. Only when the coefficients of all characteristics within a group are not equal to 0 are all characteristic values ​​retained; otherwise, all characteristic values ​​within the same group are discarded. q is retained for each type of movement. c Strongly correlated regional features, in q c Calculate the mean μ_q of p feature values ​​in a group feature. c and standard deviation σ_q c This can effectively improve the accuracy of recognition.

[0028] As an improvement, in step S2, the following air pressure characteristic values ​​are calculated for each subject's action: maximum air pressure change rate, maximum air pressure change value, time difference of air pressure change start in each zone, mean and standard deviation of air pressure difference, inter-zone air pressure correlation coefficient, air pressure change frequency, and air pressure change energy.

[0029] As an improvement, in step S2, the maximum pressure change rate V i _C a-n The calculation formula is expressed as:

[0030]

[0031] In the formula, |P i _C a-n (k+1)-P i _C a-n (k)| represents the absolute value of the pressure difference between time k and time k+1 in the i-th partition of the n-th event in the a-th type of event, max k This indicates the maximum absolute value of the pressure difference within the window time, where ΔT is the sampling interval.

[0032] As an improvement, in step S2, the maximum pressure change value ΔP i _C a-n The calculation formula is expressed as:

[0033] ΔP i _C a-n =max k |P i _C a-n (k)|-min k |P i _C a-n (k)|

[0034] In the formula, max k |P i _C a-n(k)| represents the maximum air pressure value of the i-th partition in the n-th event of the a-th type of event, min k |P i _C a-n (k)| represents the minimum air pressure value of the i-th partition in the n-th event of the a-th type of event.

[0035] As an improvement, in step S2, the starting time difference ΔT of the zone pressure change is... ij _C a-n This represents the time difference (in seconds) between the start of a significant pressure change in the airbags of the i-th and j-th zones. i _C a-n (k)-P o _C a-n When (1)>δ, k i-start _C a-n Let ΔT be the starting index of the change in the i-th partition, and let ΔT be the starting time difference of the pressure change in the partition. ij _C a-n The calculation formula is expressed as:

[0036] ΔT ij _C a-n =|k i-start _C a-n -k j-start _C a-n |*ΔT

[0037] In the formula, δ is the preset threshold and ΔT is the duration of a single sampling interval.

[0038] As an improvement, in step S2, the dominant frequency of air pressure change f i,dom _C a-n The calculation formula process includes:

[0039] a. Generate a pressure difference sequence: ΔP i (k)_C a-n =P i (k)_C a-n -P i (k-1)_C a-n ;

[0040] b. Discrete Fourier Transform based on the air pressure difference sequence:

[0041]

[0042] In the formula, f i (m) represents the Fourier transform result of the pressure difference sequence in region i, where m represents the frequency index, with values ​​m = 1, 2, 3, ..., t. a-n j represents the imaginary unit, j 2 =-1;

[0043] c. Based on f i (m) Calculate the power spectral density and the dominant frequency f in each category of events. i,dom _C a-n , is represented as:

[0044]

[0045] f i,dom _C a-n =arg max fm∈[0,fs / 2] (PSD i (fm));

[0046] The energy of pressure change is the total energy of pressure change in each region during each event, expressed as:

[0047]

[0048] In the formula, f low-i This indicates the first frequency to the left of the dominant frequency that satisfies PSD. i The frequency (f) ≤ τ, f high-i This indicates the first frequency to the right of the dominant frequency that satisfies PSD. i (f)≤τ, where τ represents the power threshold, [f low-i f high-i ] represents the effective energy frequency range, Δf represents the frequency resolution, Δf = f s / (t a-n -1).

[0049] As an improvement, in step S3, based on the elastic network regression, the feature values ​​are grouped according to the body region to learn a stable “feature-event” mapping pattern.

[0050] As an improvement, in step S3, the calculation formula for feature grouping is expressed as follows:

[0051]

[0052]

[0053] In the formula, The purpose of this formula is to find a set of parameters β0, β g,p Minimize the subsequent expression;

[0054] M represents the total sample size, which is the total number of samples for the four types of events;

[0055] b represents the sample index, which ranges from 1 to M, representing the b-th sample;

[0056] y b This represents the event label for the b-th sample;

[0057] β0 represents a constant term, which is the baseline value for predicting the time label when all feature values ​​are 0;

[0058] G represents the total number of body regions;

[0059] g represents the feature group index, which is a group divided according to body regions, with a value range of 1 to G;

[0060] p represents the feature index, indicating the p-th feature;

[0061] p∈g represents the p specific features within the g-th group;

[0062] β g,p This represents the coefficient of the p-th feature in the g-th group. Only when the coefficients of all features in a group are not 0 will all features in the same group be retained; otherwise, all features in the same group will be removed.

[0063] x b,g,p This represents the value of the p-th feature within the g-th group of the b-th sample;

[0064] λ represents the penalty intensity, with a common value range of 0.0001 to 10. The larger the value, the heavier the penalty and the easier it is to remove features; the smaller the value, the lighter the penalty and the more difficult it is to remove features.

[0065] α represents the mixed weights, and its value is between 0 and 1. The goal is to ensure that features within the region are preserved or removed as a whole, avoiding the splitting of physical collaborative characteristics;

[0066] α∑ p∈g (β g,p 2 The purpose is to balance the coefficients of highly correlated features within a region in order to avoid any one feature becoming overly dominant.

[0067] As an improvement, there are four types of movements: side-lying rollover, supine rollover, gross motor movement, and minor motor movement.

[0068] As an improvement, the mattress features air bladders in five zones: head, shoulders, waist, hips, and legs.

[0069] The beneficial effects of the sleeping posture recognition method based on dynamic air pressure of an airbag mattress of the present invention are: by collecting air pressure data of each zone during the actual various movements of the subject, calculating the time feature value and frequency feature value of dynamic air pressure changes, and adopting a unique feature screening method, the recognition accuracy can be effectively improved. Attached Figure Description

[0070] Figure 1 This is a flowchart of a sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to Embodiment 1 of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be explained and described below. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0072] See Figure 1 The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to embodiments of the present invention includes: The airbag mattress has multiple zones, each zone is equipped with an air pressure sensor, and the sleeping posture recognition method based on dynamic air pressure of the airbag mattress includes:

[0073] Step S1: Collect the air pressure values ​​of the air pressure sensor in each zone during the subject's various actions at a preset sampling frequency, record the time point of the maximum air pressure value, and the total duration of each type of action is random.

[0074] Step S2: Calculate the time characteristic value and frequency characteristic value of the dynamic change of air pressure for each subject in each type of action;

[0075] Step S3: Group the feature values ​​according to the partition. Only when the coefficients of all features in the same group are not equal to 0, are all feature values ​​in the same group retained; otherwise, all feature values ​​in the same group are removed. q is retained for each action category. c Strongly correlated regional characteristics;

[0076] Step S4, in q c Calculate the mean μ_q of p feature values ​​in a group feature. c and standard deviation σ_q c ;

[0077] Step S5: After the user lies down, monitor the air pressure values ​​of each zone within the mattress in real time. Calculate the air pressure characteristic value of each zone within the window time t, and determine whether the air pressure characteristic value is within [μ_q]. c -θ*σ_q c Within the range of θ, if so, it is determined that a corresponding rolling or body movement event occurred within the window time, where θ is the percentile of the normal distribution.

[0078] The sleeping posture recognition method based on dynamic air pressure of an airbag mattress in this embodiment of the invention collects air pressure data of each zone during various actual movements of the subject, calculates the time and frequency feature values ​​of dynamic air pressure changes, groups the feature values ​​according to the zone, and retains all feature values ​​in the same group only when the coefficient of all features in the same group is not equal to 0; otherwise, all feature values ​​in the same group are discarded. q is retained for each type of movement. c Strongly correlated regional features, in q c Calculate the mean μ_q of p feature values ​​in a group feature. cand standard deviation σ_q c This can effectively improve the accuracy of recognition.

[0079] Example 1

[0080] See Figure 1 The sleeping posture recognition method based on dynamic air pressure of an airbag mattress in Embodiment 1 of the present invention includes the following steps.

[0081] First, define the parameters.

[0082] Suppose an airbag mattress has i partitioned airbags and i corresponding pressure sensors, with a pressure sensor sampling rate of f. s ;

[0083] The final output event category (action category) C a It includes four categories {C1, C2, C3, C4}, where 'a' represents the subscript index, and 'a = {1, 2, 3, 4}.

[0084] C1 = Rolling over from side-lying to flat-lying, C2 = Rolling over from flat-lying to side-lying, C3 = Large body movement, C4 = Small body movement.

[0085] Airbag mattresses typically have five zones, corresponding to the head, shoulders, waist, hips, and legs of the human body, i=5.

[0086] Then, collect the data.

[0087] Several participants were invited to perform four types of movements on a mattress in sequence: side-lying rollover, supine rollover, gross motor movement, and minor motor movement. The number of movements in each type was the same, n.

[0088] The total duration T_C of each action event in each category a-n The sampling interval ΔT of the barometric pressure sensor is fixed, but the sampling points t in each category of action are random. a-n =T_C 1-n / ΔT;

[0089] Simultaneously collect the air pressure values ​​P from i barometers during the occurrence of the above four types of events. i _C a-n (k), where k is the index of the sampling point, k = {1, ..., t} a-n}

[0090] For example: Suppose n = 10, T_C 1-9 This represents the total duration of the 9th rolling movement in the side-lying rolling movement category; assuming T_C 1-9 =30 seconds, ΔT=0.1 seconds, air pressure sampling point t_C during the 9th rolling action in the side-lying rolling category. 1-n =30 / 0.1 = 300. Therefore, P2_C1-9 (150) represents the shoulder air pressure value corresponding to the 150th sampling point during the 9th rolling action in the side-lying rolling category.

[0091] Secondly, feature calculation.

[0092] This embodiment needs to obtain the time characteristics and frequency characteristics of air pressure dynamic changes during several turning / body movements. The former reflects the dynamic changes of air pressure over time, and the latter reflects the frequency distribution of air pressure changes.

[0093] For example, if when rolling over from a side-lying position, most people tend to move their shoulders first, then use their hips for leverage, and finally adjust their legs and back to complete the entire rolling over motion, then it is very likely that the air pressure in the shoulders will rise sharply first, followed by the air pressure in the hips, and finally the air pressure in the legs and back will change slightly.

[0094] The following characteristic values ​​were calculated sequentially for each subject in each category:

[0095] (1) Rate of change of air pressure V i _C a-n : Indicates the maximum rate of change of air pressure within a window of time, such as the V-shaped pressure in the lumbar and hip area when turning over in a side-lying position. i It is likely larger than a physical event, measured in kilopascals per second (kPa), and its calculation formula is:

[0096]

[0097] In the formula, |P i _C a-n (k+1)-P i _C a-n (k)| represents the absolute pressure difference between time k and time k+1 in the i-th partition of the n-th event in the a-th event;

[0098] max k This indicates that the maximum value of the difference is taken within the window period;

[0099] ΔT is the sampling interval;

[0100] (2) Maximum pressure change ΔP i _C a-n : Represents the total change in air pressure within the complete time of a single action (or a single event), measured in kilopascals (kPa). The formula for its calculation is:

[0101] ΔP i _C a-n =max k |P i _C a-n (k)|-min k |P i_C a-n (k)|

[0102] In the formula, max k |P i _C a-n (k)| represents the maximum air pressure value of the i-th partition in the n-th event of the a-th type of event;

[0103] min k |P i _C a-n (k)| represents the minimum air pressure value of the i-th partition in the n-th event of the a-th type of event;

[0104] (3) The relative time T of the maximum air pressure i-max _C a-n ,T i-min _C a-n The relative position of the time when the air pressure reaches its maximum and minimum values ​​within the window time, without units, with a value range of [0,1], is calculated using the following formula;

[0105]

[0106] In the formula, k i-max _C a-n This represents the sampling point corresponding to the maximum air pressure value of the i-th partition in the n-th event of the a-th type of event;

[0107] k i-min _C a-n This represents the sampling point corresponding to the minimum air pressure value of the i-th partition in the n-th event of type a;

[0108] t_C a-n -1 indicates the total number of sampling intervals within the window time;

[0109] (4) Time difference ΔT between the start of regional air pressure changes ij _C a-n : Indicates the time difference between the i-th and j-th airbags when a significant pressure change begins to occur, in seconds.

[0110] When P i _C a-n (k)-P i _C a-n When (1)>δ, k i-start _C a-n Let δ be the starting index of the change in the i-th partition; δ is the preset threshold, which is the threshold for defining "significant change". There is no specific restriction here. δ is determined by factors such as air pressure and the depth of the airbag in the mattress. The closer the airbag is to the surface of the mattress, the more sensitive the airbag is to changes in body movement on the bed, and the corresponding threshold should be set smaller.

[0111] The time difference ΔT between the start and end of the zone air pressure change ij _C a-n The calculation formula is:

[0112] ΔT ij _C a-n =|k i-start _C a-n -k j-start _C a-p |*ΔT;

[0113] (5) The average value of adjacent pressure differences μ oi _C s-n and standard deviation σ pi _C a-n : Reflects the stability of air pressure changes, unit (kPa); the change in small body movement events is theoretically minimal;

[0114] (6) Correlation coefficient of inter-regional air pressure r ij _C a-n The degree of linear correlation between the air pressure changes in the two zones; for example, when turning over in a side-lying position, the shoulder and hip areas may change simultaneously, so the correlation coefficient of the shoulder and hip areas may be higher when turning over in a side-lying position.

[0115] (7) The dominant frequency of air pressure change f i,dom _C a-n The main frequency components of air pressure changes (e.g., a 5-10 second turning event may have a lower dominant frequency: 0.2-0.5Hz, while a 1-3 second small body movement may have a higher dominant frequency: 0.5-2Hz) are calculated as follows:

[0116] a. Generate a pressure difference sequence: ΔP i (k)_C a-n =P i (k)_C a-n -P i (k-1)_C a-n ;

[0117] b. Discrete Fourier Transform (DFT) based on the air pressure difference sequence:

[0118]

[0119] In the formula, f i (m) represents the Fourier transform result of the pressure difference sequence in region i, and is a complex index;

[0120] m represents the frequency point index, with values ​​m = 1, 2, 3, ..., t. a-n ;

[0121] j represents the imaginary unit, j 2 =-1;

[0122] c. Based on f i (m) Calculate the power spectral density (PSD) and the dominant frequency f in each event category. i,dom _C a-n :

[0123]

[0124] f i,dom _C a-n =arg max fm∈[0,fs / 2] (PSD i (fm));

[0125] It should be noted that while existing methods perform a Direct Fourier Transform (DFT) on the raw air pressure values, this embodiment replaces the raw values ​​with a difference sequence. This allows for more accurate capture of dynamic changes related to "rolling over / body movement," thereby improving the frequency characteristics' ability to distinguish between the four types of events. The reason is that due to the cushioning effect of the mattress's comfort layer, changes in the raw air pressure values ​​caused by rolling over / body movement are often "averaged." In this case, the Fourier transform of the raw air pressure values ​​is "submerged" by static components, making the frequency characteristics of the dynamic signal indistinct. The difference sequence, however, amplifies the relative amplitude of dynamic changes.

[0126] (8) Pressure change energy: The total energy of pressure change in each region during each event;

[0127]

[0128] f low-i This indicates the first frequency to the left of the dominant frequency that satisfies PSD. i Frequency (f)≤τ

[0129] f high-i This indicates the first frequency to the right of the dominant frequency that satisfies PSD. i Frequency (f)≤τ

[0130] τ represents the power threshold, [f low-i f high-i [Indicates the effective energy frequency band range]

[0131] Δf represents the frequency resolution, Δf = f s / (t a-n -1)

[0132] It should be noted that the existing formula is used to calculate the total energy E. iThis is achieved by calculating the power spectral density (PSD) across the entire frequency band, i.e., the total energy of the signal (including all frequency components, regardless of their relevance). This embodiment dynamically calculates the local energy range near the dominant frequency by setting a bandwidth threshold τ for the signal power spectrum. This method, based on the signal's own power distribution characteristics (rather than a fixed value), can more accurately capture effective energy frequency bands, eliminate redundant frequency bands, and dynamically adapt to signal characteristics. For example, for a turning event (with a sharp main peak, such as f...). i,dom =1Hz), the effective energy frequency band range may be [0.8, 1.2] Hz (narrow bandwidth), but for large dynamic events (with a flat main peak, such as f) i,dom =3Hz), the effective energy frequency band range may be [2.8,4.2]Hz (narrow bandwidth).

[0133] Next, feature selection.

[0134] The reasons for feature selection are as follows: 1. The actions of different events are different among the four types of events, so the weights of different feature values ​​are also different. For example, the correlation between the shoulders and hips may be more important in the rolling over event, and the dispersion of air pressure changes may be more important in the body movement event; 2. There is a strong correlation between the features in the four types of events, that is, "there is a lot of redundant information". Without feature selection (i.e., using all features for sleeping posture recognition), the data will be sparsely distributed in high-dimensional space, and the model will find it difficult to learn a stable "feature-event" mapping pattern from sparse data.

[0135] The traditional approach to feature selection is to calculate the correlation coefficient between features, set a fixed threshold (such as 0.7), and perform a removal procedure on highly correlated features. The limitations of this approach are that it relies on subjective thresholds, ignores the co-correlation of multiple features (it can only handle pairwise correlations and cannot handle scenarios where three or more features are correlated), and may lose key information.

[0136] The approach in this embodiment is as follows: based on elastic network regression, the feature values ​​are grouped according to body regions, and it is set that all feature values ​​in the same group can be retained only when the coefficients of all features in the same group are not equal to 0; otherwise, all features in the same group are removed.

[0137] The formula for feature selection is:

[0138]

[0139] The purpose of this formula is to find a set of parameters (β0, β...). g,p This minimizes the subsequent expression;

[0140] M represents the total sample size, which is the total number of samples for the four types of events (e.g., if there are 20 samples for each of the four types of events, M = 80).

[0141] b represents the sample index, which ranges from 1 to M, representing the b-th sample;

[0142] y b This represents the event label of the b-th sample (e.g., y1 = 1 indicates that the sleeping position of the first sample is side-lying and turning over).

[0143] β0 represents a constant term, which is the baseline value for predicting the time label when all feature values ​​are 0;

[0144] G represents the total number of body regions;

[0145] g represents the feature group index, which is a group divided according to "body region", with a value range of 1 to G;

[0146] p represents the feature index, indicating the p-th feature;

[0147] p∈g represents the p specific features within the g-th group;

[0148] β g,p This represents the coefficient (core parameter) of the p-th feature in the g-th group. Only when the coefficients of all features in a group are not 0 will all features in the same group be retained; otherwise, all features in the same group will be removed.

[0149] x b,g,p This represents the value of the p-th feature within the g-th group of the b-th sample;

[0150] λ represents the penalty intensity (customizable), with a common value range of 0.0001 to 10. A larger value indicates a heavier penalty and makes it easier to remove features, while a smaller value indicates a lighter penalty and makes it more difficult to remove features.

[0151] α represents the mixed weight (customizable), with a value between 0 and 1;

[0152] The goal is to ensure that features within the region are preserved or removed as a whole, avoiding the splitting of physical collaborative characteristics;

[0153] α∑ p∈g (β g,p 2 The purpose is to balance the coefficients of highly correlated features within a region, that is, to avoid any one feature from becoming overly dominant.

[0154] The optimal λ and α can be calculated using existing cross-validation methods.

[0155] The significance of this embodiment is as follows: 1. The features of a single body region are a coordinated response to the region's state. The method of group determination can completely preserve the physical information of the region's dynamics, ultimately improving the accuracy and interpretability of sleeping posture recognition; 2. The ultimate goal of sleeping posture recognition is to link mattress zone adjustment (e.g., softening the shoulder airbag when turning over in a side-lying position). Since the mattress zones of the region combination are completely corresponding, after preserving the features of the same group, it is possible to directly determine whether the region needs adjustment by "preserving a certain region group". This avoids situations where "the shoulder air pressure change rate is preserved, but the shoulder main frequency is eliminated" but the mattress adjustment cannot be matched (i.e., it is uncertain whether the region needs to be softened as a whole), thus providing a clear logic for subsequent mattress adjustment.

[0156] It's important to clarify that the core of sleep posture recognition is not "the usefulness of a single feature," but rather "the synergistic usefulness of multiple features within a region." For example, retaining only the "maximum change in lumbar pressure" might fail to distinguish between "lying flat and turning over" (lumbar pressure change + high frequency) and "user pressing on the lower back" (lumbar pressure change + low frequency). Retaining both "lumbar pressure + frequency" within the same group allows for accurate identification of lying flat and turning over through "high pressure + high frequency." Therefore, "retaining features within the same group" ensures a "complete description of regional dynamics," rather than leading the model into the misconception of "sparseness for the sake of sparseness."

[0157] Finally, feature application.

[0158] After the preceding calculations, q is retained for each type of event. c The group has strongly correlated regional features, and c is the event category index.

[0159] In q c Calculate the mean μ_q of p feature values ​​in a group feature. c and standard deviation σ_q c .

[0160] After the user lies down, the air pressure values ​​in the mattress zones are monitored in real time. The air pressure characteristic value within a time window (t) is calculated, and it is determined whether the characteristic value falls within the range of [μ_q]. c -θ*σ_q c Within the specified range, if so, it is determined that a corresponding rolling / movement event occurred within the window time. For example, the occurrence of a side-lying rolling event is usually due to the prolonged maintenance of the side-lying position, and the rolling movement is caused by shoulder pressure discomfort. Considering the fatigue of the shoulder muscles due to pressure, the movement characteristics may be that the movement first uses the buttocks to gain strength, and then the shoulder is leveled. In terms of characteristic values, this is reflected in the time difference of the start of the change in air pressure in the shoulder and buttocks, the correlation coefficient between the air pressure in the shoulder and buttocks, etc.

[0161] The beneficial effects of the sleeping posture recognition method based on dynamic air pressure of an airbag mattress in Embodiment 1 of the present invention are as follows:

[0162] It calculates data related to air pressure difference, breaking through the limitation of insignificant static air pressure changes and solving the problem of small static air pressure difference before and after turning over and low recognition accuracy caused by the comfort layer buffer.

[0163] Improved individual adaptability: The calibration process is independent of user weight and initial airbag pressure, adapting to different user groups;

[0164] It can distinguish between rolling over and body movement: accurately identify four types of events: rolling over while lying on the side, rolling over while lying on the back, large body movement, and small body movement, reducing missed and false judgments;

[0165] Reduce hardware costs: Use the mattress's existing zoned pressure sensors, eliminating the need for additional sensor hardware.

[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the content described in the above specific embodiments. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. A sleeping posture recognition method based on dynamic air pressure of an airbag mattress, wherein the airbag mattress has multiple zones, each zone is equipped with an air pressure sensor, characterized in that: Sleep posture recognition methods based on dynamic air pressure in airbag mattresses include: Step S1: Collect the air pressure values ​​of the air pressure sensor in each zone during the subject's various actions at a preset sampling frequency, record the time point of the maximum air pressure value, and the total duration of each type of action is random. Step S2: Calculate the time characteristic value and frequency characteristic value of the dynamic change of air pressure for each subject in each type of action; Step S3: Group the feature values ​​according to the partition. Only when the coefficients of all features in the same group are not equal to 0, are all feature values ​​in the same group retained; otherwise, all feature values ​​in the same group are removed. q is retained for each action category. c Strongly correlated regional characteristics; Step S4, in q c Calculate the mean μ_q of p feature values ​​in a group feature. c and standard deviation σ_q c ; Step S5: After the user lies down, monitor the air pressure values ​​of each zone within the mattress in real time. Calculate the air pressure characteristic value of each zone within the window time t, and determine whether the air pressure characteristic value is within [μ_q]. c -θ*σ_q c Within the range of θ, if so, it is determined that a corresponding rolling or body movement event occurred within the window time, where θ is the percentile of the normal distribution.

2. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 1, characterized in that: In step S2, the following air pressure characteristic values ​​are calculated for each subject for each type of action: maximum air pressure change rate, maximum air pressure change value, time difference of air pressure change start in each zone, mean and standard deviation of air pressure difference, inter-zone air pressure correlation coefficient, air pressure change frequency, and air pressure change energy.

3. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 2, characterized in that: In step S2, the maximum rate of change of air pressure V i _C a-n The calculation formula is expressed as: In the formula, |P i _C a-n (k+1)-P i _C a-n (k)| represents the absolute value of the pressure difference between time k and time k+1 in the i-th partition of the n-th event in the a-th type of event, max k This indicates the maximum absolute value of the pressure difference within the window time, where ΔT is the sampling interval.

4. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 2, characterized in that: In step S2, the maximum pressure change ΔP i _C a-n The calculation formula is expressed as: ΔP i _C a-n = max k |P i _C a-n (k)|-min k |P i _C a-n (k)| In the formula, max k |P i _C a-n (k)| represents the maximum air pressure value of the i-th partition in the n-th event of the a-th type of event, min k |P i _C a-n (k)| represents the minimum air pressure value of the i-th partition in the n-th event of the a-th type of event.

5. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 2, characterized in that: In step S2, the starting time difference ΔT between the changes in air pressure in each zone ij _C a-n This represents the time difference (in seconds) between the start of a significant pressure change in the airbags of the i-th and j-th zones. i _C a-n (k)-P i _C a-n When (1)>δ, k i-start _C a-n Let ΔT be the starting index of the change in the i-th partition, and let ΔT be the starting time difference of the pressure change in the partition. ij _C a-n The calculation formula is expressed as: ΔT ij _C a-n =|k i-start _C a-n -k j-start _C a-n |*ΔT In the formula, δ is the preset threshold and ΔT is the duration of a single sampling interval.

6. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 2, characterized in that: In step S2, the dominant frequency of air pressure change f i,dom _C a-n The calculation formula process includes: a. Generate a pressure difference sequence: ΔP i (k)_C a-n =P i (k)_C a-n -P i (k-1)_C a-n ; b. Discrete Fourier Transform based on the air pressure difference sequence: In the formula, f i (m) represents the Fourier transform result of the pressure difference sequence in region i, where m represents the frequency index, with values ​​m = 1, 2, 3, ..., t. a-n j represents the imaginary unit, j 2 =-1; c. Based on f i (m) Calculate the power spectral density and the dominant frequency f in each category of events. i,dom _C a-n , is represented as: f i,dom _C a-n =arg max fm∈[0,fs / 2] (PSD i (fm)); The energy of pressure change is the total energy of pressure change in each region during each event, expressed as: In the formula, f low-i This indicates the first frequency to the left of the dominant frequency that satisfies PSD. i The frequency (f) ≤ τ, f high-i This indicates the first frequency to the right of the dominant frequency that satisfies PSD. i (f)≤τ, where τ represents the power threshold, [f low-i f high-i ] represents the effective energy frequency range, Δf represents the frequency resolution, Δf = f s / (t a-n -1).

7. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 1, characterized in that: In step S3, based on the elastic network regression, the feature values ​​are grouped according to body regions to learn a stable "feature-event" mapping pattern.

8. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 7, characterized in that: In step S3, the formula for calculating feature grouping is expressed as follows: In the formula, The purpose of this formula is to find a set of parameters β0, β g,p Minimize the subsequent expression; M represents the total sample size, which is the total number of samples for the four types of events; b represents the sample index, which ranges from 1 to M, representing the b-th sample; y b This represents the event label for the b-th sample; β0 represents a constant term, which is the baseline value for predicting the time label when all feature values ​​are 0; G represents the total number of body regions; g represents the feature group index, which is a group divided according to body regions, with a value range of 1 to G; p represents the feature index, indicating the p-th feature; p∈g represents the p specific features within the g-th group; β g,p This represents the coefficient of the p-th feature in the g-th group. Only when the coefficients of all features in a group are not 0 will all features in the same group be retained; otherwise, all features in the same group will be removed. x b,g,p This represents the value of the p-th feature within the g-th group of the b-th sample; λ represents the penalty intensity, with a common value range of 0.0001 to 10. The larger the value, the heavier the penalty and the easier it is to remove features; the smaller the value, the lighter the penalty and the more difficult it is to remove features. α represents the mixed weights, and its value is between 0 and 1. The goal is to ensure that features within the region are preserved or removed as a whole, avoiding the splitting of physical collaborative characteristics; α∑ p∈g (β g,p 2 The purpose is to balance the coefficients of highly correlated features within a region in order to avoid any one feature becoming overly dominant.

9. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 1, characterized in that: There are four types of movements: side-lying rollover, supine rollover, gross motor movement, and fine motor movement.

10. The sleeping posture recognition method based on dynamic air pressure of an airbag mattress according to claim 1, characterized in that: The mattress has air bladders in five zones: head, shoulders, waist, hips, and legs.

Images