Intelligent bed chair comfort level quantitative evaluation method based on multi-dimensional feature fusion

Abstract

The invention discloses an intelligent bed chair comfort level quantitative evaluation method based on multi-dimensional feature fusion, and the method comprises the steps: an offline model training stage: collecting a pressure signal of the surface of a bed chair, processing the pressure signal to extract multi-dimensional pressure feature parameters, and obtaining user individual parameters and environment parameters at the same time; generating a fusion supervision signal; using a machine learning algorithm, taking the data as training data, and training to obtain a comfort evaluation model; and an online scoring stage: collecting pressure signals on the surfaces of the beds and the chairs in real time, processing the pressure signals to extract real-time multi-dimensional pressure characteristic parameters, combining the user individual parameters and the environmental parameters, inputting the parameters into the comfort evaluation model, and outputting a real-time comfort score. According to the method, a machine learning model fusing pressure, user and environment characteristics and subjective and objective dual-supervision signals is trained offline, and the model is applied online, so that real-time and quantitative evaluation of the comfort level of the intelligent bed chair is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology for rehabilitation and nursing, and in particular to a quantitative evaluation method for the comfort of intelligent beds and chairs based on multi-dimensional feature fusion. Background Technology

[0002] With the increasing demand for rehabilitation and nursing care, integrated smart beds and chairs have become a hot product, aiming to help the elderly and disabled live and rehabilitate more comfortably. These smart devices can assist users in changing positions between beds and wheelchairs, providing personalized care and rehabilitation support. However, most current smart beds and chairs focus on mechanical and hardware design, such as structural innovation, motion stability, and safety features, while lacking effective solutions for comfort, a key factor in improving the user experience. Comfort is a complex feeling involving physiological perception and psychological expectations, extremely difficult to measure and control directly; quantifying it into signals that machines can recognize is a technological challenge.

[0003] Existing technologies primarily employ two approaches to address comfort issues. One is a pre-programmed control scheme where the device stores typical modes (e.g., reading mode, rest mode) and adjusts the user's posture through fixed movement trajectories and speeds. This method is simple, but because comfort relies entirely on the designer's experience, it cannot cater to the needs of different individuals, resulting in a simplistic and passive user experience. The other approach is based on simple pressure threshold detection. Some high-end products or research equipment install several pressure sensors in key areas of the mattress to detect a fixed electrical threshold. When the pressure exceeds the threshold, the system issues an alarm or stops movement. While this approach provides some safety protection, its comfort evaluation is too simplistic and lacks detailed differentiation, failing to adapt to the personalized needs of different users.

[0004] This situation leads to several major problems with existing technologies: First, comfort evaluation relies too heavily on subjective feelings and lacks objective standards; second, the movement modes of the devices lack personalized adjustments, making it difficult to meet the needs of different users; third, the effects are delayed and inaccurate, intervening only when discomfort occurs; and fourth, it fails to fully utilize multi-dimensional pressure information, such as pressure distribution and dynamic change trends, thus failing to provide more accurate comfort assessments. Therefore, this invention aims to establish a new method that can objectively quantify comfort by combining multi-dimensional pressure characteristics, individual user differences, and environmental information to achieve dynamic and personalized comfort evaluation of integrated smart beds and chairs in different posture changes. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology. To achieve the above objective, a quantitative evaluation method for the comfort of intelligent beds and chairs based on multi-dimensional feature fusion is adopted to solve the problems mentioned in the background technology.

[0006] A quantitative evaluation method for the comfort of smart beds and chairs based on multi-dimensional feature fusion includes the following steps: Offline model training phase: pressure signals from the bed and chair surfaces are collected and processed to extract multi-dimensional pressure feature parameters, while user-specific parameters and environmental parameters are also acquired; a fusion supervision signal is generated based on subjective ratings and electromyography data; a comfort evaluation model is trained using machine learning algorithms with the multi-dimensional pressure feature parameters, user-specific parameters, environmental parameters, and the fusion supervision signal as training data. Online rating phase: Pressure signals from the bed and chair surfaces are collected in real time, and the pressure signals are processed to extract real-time multi-dimensional pressure feature parameters. These parameters are then combined with individual user parameters and environmental parameters and input into the comfort evaluation model to output a real-time comfort score.

[0007] As a further aspect of the present invention, the offline model training phase specifically includes the following steps: S11. Signal acquisition and processing of array pressure sensing unit: The timing tags on both sides of the bed / chair are synchronized through a synchronization device, the raw signals of the pressure sensor array are acquired, and calibration and filtering are performed. S12. Determination of human lying posture: Based on the filtered pressure data, calculate the global pressure center coordinates and the pressure ratio of the left and right halves, and combine historical buffer data to determine whether the human body is in a supine position, and collect subsequent data when the human body is determined to be in a supine position. S13. Multidimensional pressure feature data extraction: For multiple key areas of the bed / chair surface, calculate the static pressure distribution characteristics and dynamic pressure change characteristic parameters, and construct a multidimensional pressure feature matrix; S14. Dual-supervised signal fusion of subjective rating and electromyography (EMG) data: Extract EMG features from key areas, use logistic functions to map EMG features to objective comfort ratings, and fuse them with subjective ratings through a dynamic weighting algorithm to generate a fused supervision label signal; S15. Offline model training to obtain a comfort evaluation model: The multi-dimensional pressure feature parameters, user individual parameters, environmental parameters, and the fused supervision label signal are used as sample data, and machine learning algorithms are used to train the model to obtain a comfort evaluation model.

[0008] As a further aspect of the present invention, the online scoring stage specifically includes the following steps: S21. Multidimensional pressure feature extraction in the online application stage: Simultaneously collect pressure signals from both sides of the bed / chair, and use the same feature extraction method as in the offline stage to obtain real-time static and dynamic pressure feature parameters; S22. Construct a real-time online feature vector: Combine the real-time pressure feature parameters, pre-configured individual user parameters, and real-time collected environmental parameters into an online feature vector; S23. Comfort rating during online application phase: The online feature vector is input into the comfort evaluation model, and the model outputs a quantified comfort rating in real time.

[0009] As a further aspect of the present invention: the filtering process in step S11 employs a region-adaptive Wiener filtering method based on prior knowledge, specifically including the following steps: The bed / chair surface is divided into multiple fixed regions. For the pressure unit in each region, the mean and variance of its local neighborhood are calculated, and the filtering intensity is adaptively adjusted based on the noise variance estimate of each region.

[0010] As a further aspect of the present invention: the specific steps for determining the human body's bed-lying posture in step S12 include: Calculate the global pressure center coordinates (CoP_x(t), CoP_y(t)) and the pressure ratio R_lr(t) between the left and right halves; By comparing the lateral offset of the pressure center coordinate with the threshold Threshold_x, the deviation of the left and right half-zone pressure ratio from 1 with the threshold Threshold_ratio, and combining the comparison of the longitudinal coordinate of the pressure center with the sitting-up posture threshold Threshold_sit, and the relationship between the left and right half-zone pressure ratio and the side-lying posture threshold Threshold_side, a preliminary judgment of the bed rest posture can be made. By utilizing a historical buffer that stores preliminary judgment results over a recent period, the final lying position is confirmed when the judgment results in the buffer are consistent for a consecutive predetermined period of time. When the final lying position is not a supine position, a voice reminder is triggered or the drive control system is used to force the user to lie in a supine position.

[0011] As a further aspect of the present invention: in step S13, the static pressure distribution characteristics include the average pressure within the region and the distribution uniformity index, and the dynamic pressure change characteristics include the regional average pressure change rate and the total pressure change rate; the multi-dimensional pressure feature matrix has 25 dimensions, which is obtained by calculating the average pressure, distribution uniformity index and average pressure change rate for each of the 8 partitions, and calculating a total pressure change rate.

[0012] As a further aspect of the present invention: the specific steps for fusing the subjective score and electromyographic data dual-supervision signals in step S14 include: By attaching electromyography (EMG) sensors to key muscle areas of the left / right shoulder, left / right hip, and left / right thigh, 6-channel EMG signals are extracted, and their root mean square values ​​are calculated to obtain an EMG feature dataset. The electromyography feature dataset is mapped to an objective discomfort score using the logistic function Sigmoid, and further converted into an electromyography objective comfort score S_emg(t) of 1-10. A dynamic weighted fusion algorithm is used to fuse subjective scores S_sub(t) and objective electromyographic comfort scores S_emg(t) to generate a fusion supervision label signal S_comb(t). The weight α of the subjective score is adjusted according to the decisiveness of the user's rating or the historical consistency between the subjective score and the electromyographic score.

[0013] As a further aspect of the present invention: in step S15, the sample data is obtained through multiple sets of experiments for different users. For users whose electromyography scores are highly consistent with their subjective scores, a large amount of sample data can be obtained using only the electromyography scores when constructing the sample dataset. The machine learning algorithm adopts the gradient boosting regression tree (GBRT) algorithm and divides the sample dataset into a training set and a validation set according to the proportion β, where the value of β ranges from [0.7, 0.9].

[0014] As a further aspect of the present invention: the individual user parameters include weight, age, height, and blood pressure; the environmental parameters include ambient temperature and ambient humidity.

[0015] As a further aspect of the present invention: the individual user parameters are input and pre-configured into the system controller via a mobile APP; the environmental parameters are collected in real time via temperature and humidity sensors.

[0016] Compared with the prior art, the present invention has the following technical advantages: Using the above technical solution, in the offline phase, pressure signals from the surface of the bed and chair are collected and processed to extract multi-dimensional pressure distribution and variation characteristics. These characteristics are then combined with individual user parameters and environmental parameters. Furthermore, the system innovatively integrates subjective user ratings with objective electromyographic signals reflecting muscle tension to generate a more reliable supervisory label, thereby training a machine learning model (i.e., a comfort evaluation model). Subsequently, in the online phase, this trained model is used to collect real-time pressure, user, and environmental data, quickly outputting a quantitative and personalized comfort score. This achieves an objective, accurate, and real-time evaluation of the comfort of the smart bed and chair. Attached Figure Description

[0017] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the steps of the quantitative evaluation method according to an embodiment of this application; Figure 2 This is a flowchart illustrating the overall technical solution of the evaluation method disclosed in this application. Figure 3 This is a flowchart of the first-stage comfort evaluation model construction method according to an embodiment of this application; Figure 4 This is a flowchart of the signal acquisition and processing of the array pressure sensing unit according to an embodiment of this application; Figure 5 This is a schematic diagram illustrating the bed / chair surface division according to an embodiment of this application; Figure 6 This is a flowchart of the human bed rest posture determination algorithm according to an embodiment of this application. Detailed Implementation

[0018] 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.

[0019] Please refer to Figure 1 and Figure 2 In this embodiment of the invention, a quantitative evaluation method for the comfort of a smart bed-chair based on multi-dimensional feature fusion includes the following steps: like Figure 2 As shown, it mainly includes two stages: comfort evaluation model training and online real-time comfort scoring.

[0020] The first stage, the comfort evaluation model training stage, involves processing experimentally collected data, constructing a sample dataset, and training the model. In the second stage, based on the above-mentioned training and evaluation model, real-time online multi-dimensional feature data is collected and input into the evaluation model to score comfort.

[0021] Step S1, Offline Model Training Stage: Collect pressure signals from the surface of the bed and chair, process the pressure signals to extract multi-dimensional pressure feature parameters, and simultaneously obtain user individual parameters and environmental parameters; generate a fusion supervision signal based on subjective ratings and electromyography data; use machine learning algorithms, with the multi-dimensional pressure feature parameters, user individual parameters, environmental parameters and the fusion supervision signal as training data, to train a comfort evaluation model; In this embodiment, the user's individual parameters include weight, age, height, and blood pressure; the environmental parameters include ambient temperature and ambient humidity.

[0022] In this embodiment, individual user parameters are input and pre-configured into the system controller via a mobile app; environmental parameters are collected in real time by temperature and humidity sensors.

[0023] like Figure 3 As shown in the figure, the offline stage of comfort evaluation model construction includes pressure sensing unit signal acquisition and processing, judgment of human bed-lying posture, extraction and calculation of multi-dimensional pressure feature parameters, dual-supervised signal fusion of subjective scores and objective electromyographic data, and training of comfort evaluation model. S11. Signal acquisition and processing of array pressure sensing units: The timing tags on both sides of the bed / chair are synchronized through a synchronization device to acquire the raw signals of the pressure sensor array and perform calibration and filtering. Specifically, the effective pressure unit signals of each area are acquired and calibrated, and the calibrated signals are filtered using a region adaptive Wiener filtering method based on prior knowledge. The filtering process in step S11 employs a region-adaptive Wiener filtering method based on prior knowledge. The specific steps include: The bed / chair surface is divided into multiple fixed regions. For the pressure unit in each region, the mean and variance of its local neighborhood are calculated, and the filtering intensity is adaptively adjusted based on the noise variance estimate of each region.

[0024] (1) Establish a relative coordinate system: take the center point of the head of the bed as the origin O, the direction from the head of the bed to the foot of the bed as the positive y-axis, the direction from the head of the bed to the side of the wheelchair as the positive x-axis, and the direction from the ground to the sky as the positive z-axis. Establish a spatial coordinate system. Since the pressure value of the z-axis is not valuable for the research of this technical solution, the pressure is only considered in the XOY plane direction, that is, the x and y axis values.

[0025] like Figure 4 The diagram shown is a schematic representation of the bed / chair surface division. (2) Signal acquisition to obtain pressure matrix Vraw_k(i,j,t): Two M*(N / 2) dimensional flexible pressure sensor mattresses are integrated into the bed side and wheelchair side of the integrated bed and chair respectively, splicing them into an M*N dimensional smart mattress sensor. The sampling time is synchronized through the bed side infrared receiver and transmitter and the chair side infrared receiver. Then the system reads the original voltage signal of the sensor array at a frequency f to obtain the M*N original pressure unit signal matrix Vraw_k(i,j,t) at time t, where i is the row index, corresponding to the i-th of the M sensing units in the bed length direction (y-axis); j is the column index, corresponding to the j-th of the N sensing units in the bed width direction (x-axis); k is the bed / chair surface area index, corresponding to the k-th area Rgn_k in the 8 partitions. If the variable mentioned below does not have the subscript k, it is assumed that this variable represents the entire effective area of ​​the bed surface.

[0026] (3) Signal calibration output conversion value P_k(i,j,t): Each pressure sensing unit (i, j) is calibrated offline using a standard pressure device. The conversion coefficient gain and zero bias are denoted as GA_k[i,j] and ZO_k[i,j], respectively. At time t, the original signal of each pressure unit is converted in real time to obtain the calibrated pressure matrix P_k(i,j,t) at time t.

[0027] (4) Signal filtering output P_f_k(i,j,t): The region adaptive Wiener filtering method based on prior knowledge is used to filter the pressure matrix. The method of this invention performs filtering, and compared to traditional median or Gaussian filtering, it can adaptively adjust the filtering intensity according to the pressure characteristics of each region, thus better preserving signal details. Specifically, the bed / chair surface is divided into eight fixed regions Rgn_k (k=1,2,...,8) including left / right backrest, left / right buttocks, left / right thighs, and left / right calves. Then, for the pressure unit within the k-th region, the mean μ_loc_k and variance σ_loc_k² of its local neighborhood are calculated. Filtered

[0028] in, This is the noise variance estimate for the k-th region Rgn_k.

[0029] S12. Determination of Human Bedridden Posture: Based on filtered pressure data, calculate the global pressure center coordinates and the pressure ratio of the left and right halves. Combined with historical buffer data, determine whether the human body is in a supine position. If the human body is determined to be in a supine position, subsequent data collection is performed. Specifically, the global pressure center and the pressure ratio of the left and right halves are analyzed and calculated in the time domain. By analyzing the pressure center trajectory and regional distribution characteristics, combined with the system's historical buffer data, the posture of the bedridden human body is quickly determined. The specific steps for determining the human body's bed rest posture in step S12 include: Calculate the global pressure center coordinates (CoP_x(t), CoP_y(t)) and the pressure ratio R_lr(t) between the left and right halves; By comparing the lateral offset of the pressure center coordinate with the threshold Threshold_x, the deviation of the left and right half-zone pressure ratio from 1 with the threshold Threshold_ratio, and combining the comparison of the longitudinal coordinate of the pressure center with the sitting-up posture threshold Threshold_sit, and the relationship between the left and right half-zone pressure ratio and the side-lying posture threshold Threshold_side, a preliminary judgment of the bed rest posture can be made. By utilizing a historical buffer that stores preliminary judgment results over a recent period, the final lying position is confirmed when the judgment results in the buffer are consistent for a consecutive predetermined period of time. When the final lying position is not a supine position, a voice reminder is triggered or the drive control system is used to force the user to lie in a supine position.

[0030] like Figure 5 The diagram shows the flowchart of the algorithm for determining the human body's posture while lying in bed. To ensure the reliability of comfort data, scoring must be performed in a supine position. Therefore, the human posture is determined before data collection. The set of bedridden postures is assumed to be Pos(t) = {SUPINE (upright), LEFT_SIDE (left lateral), RIGHT_SIDE (right lateral), SITTING (sitting), UNCERTAIN (uncertain). Based on the above filtered data... Data is used to calculate the pressure ratio R_lr(t) of the left and right halves of the global pressure center CoP(t). The bed posture is determined in real time based on the pressure center trajectory and regional distribution characteristics. Historical buffer data is used to further confirm the final posture. When the user is in a supine position, a voice prompt initiates signal acquisition and posture change; when the user is not in a supine position, a voice prompt continues until the user adjusts to a supine position, at which point posture change begins and data acquisition starts. The specific implementation method is as follows: First, calculate the global pressure center coordinates CoP(t) (CoP_x(t), CoP_y(t)).

[0031]

[0032] (2) Then, calculate the pressure ratio R_lr(t) between the left and right halves. ①The longitudinal centerline of the bed surface (j = N / 2) serves as the left and right boundaries; ② Calculate the total pressure in the left half: P_left(t) = Where j∈(0, N / 2], k={1,3,5,7}; ③ Calculate the total pressure in the right half: P_right(t) = Where j∈(N / 2,N],k={2,4,6,8}; ④ Finally, the pressure ratio of the left and right halves is calculated: R_lr(t) = P_left(t) / P_right(t).

[0033] (3) Secondly, a preliminary real-time assessment of the bed rest posture is performed. ① Define the threshold for lateral displacement of the pressure center as Threshold_x, the threshold for left-right pressure balance as Threshold_ratio, the threshold for sitting up as Threshold_sit, and the threshold for side-lying as Threshold_side. The initial values ​​of each of these thresholds are obtained through experimental data calibration, and then fine-tuned based on feedback during actual use.

[0034] Define condition T1: abs(CoP_x(t) - N / 2) <Threshold_x Define condition T2: abs(R_lr(t) - 1) <Threshold_ratio ② If both conditions 1 and 2 are met, the following formula can be used to preliminarily determine whether the posture is sitting up or lying down.

[0035] ③ If conditions 1 and 2 are not satisfied simultaneously: The following formula can be used to make a preliminary judgment on whether the patient is in a left lateral decubitus position.

[0036] Determine whether the patient is in a right lateral decubitus position using the following formula.

[0037] If R_lr(t) does not satisfy the above left lateral position and right lateral position logic, it is marked as UNCERTAIN (uncertain), which usually occurs during posture transformation or special postures.

[0038] (4) Next, the final lying position is confirmed based on historical buffer data. The system establishes a short-term history buffer to store single posture judgment values, such as storing the preliminary judgment results of the most recent 5 seconds. It checks whether the preliminary judgment results of 5 consecutive seconds in the buffer are completely the same. If they are the same, the current posture is determined to be that posture, and the final posture result is output. If not, the previously confirmed posture is maintained, or UNCERTAIN is output.

[0039] (5) Finally, data collection is performed after a voice prompt. When the system is ready to collect training data, if the posture is SUPINE (lying supine), a voice prompt will indicate that data collection has begun. If the posture is other, such as SITTING, LEFT_SIDE, RIGHT_SIDE, or UNCERTAIN, a voice prompt will be triggered, such as "It has been detected that you are not currently in a supine position. To ensure accurate comfort evaluation, please slowly lie down, keeping your face up." Alternatively, the system may use a drive control system to forcibly adjust the user to a supine position.

[0040] S13. Multidimensional pressure feature data extraction: For multiple key areas of the bed / chair surface, calculate static pressure distribution features and dynamic pressure change feature parameters to construct a multidimensional pressure feature matrix; specifically, for 8 key areas of the bed / chair surface, calculate 3 types of static pressure distribution features and 2 types of dynamic pressure change feature parameters to construct a 25-dimensional pressure feature matrix. In step S13, the static pressure distribution characteristics include the average pressure and distribution uniformity index within the region, and the dynamic pressure change characteristics include the regional average pressure change rate and the total pressure change rate. The multi-dimensional pressure feature matrix has 25 dimensions and is obtained by calculating the average pressure, distribution uniformity index, and average pressure change rate for each of the eight partitions, and by calculating a total pressure change rate.

[0041] Specifically, multi-dimensional pressure feature parameter extraction and calculation; Given the pressure distribution matrix P_f_k(t), for eight Regn regions, calculate three types of static distribution features and two types of dynamic change feature parameters to construct a 25-dimensional pressure feature matrix F_pressure(t) for the eight regions. The specific method is as follows: (1) First, calculate the average pressure in the k-th region of the static pressure characteristics. and distribution uniformity index ;

[0042] Among them, It is the total number of effective pressure units within region Regn_k, and [i,j]∈Regn_k.

[0043] (2) Next, calculate the average pressure change rate F_roc_avg_k(t) of the kth region in the dynamic pressure change characteristics to characterize the regional pressure stability during attitude change.

[0044] in, It is the pressure signal sampling period.

[0045] (3) Then, calculate the total pressure change rate F_roc_total(t) of all effective sensing units in the dynamic pressure change characteristics to characterize the global pressure stability during attitude change;

[0046] (4) Finally, construct a 25-dimensional pressure feature vector F_pressure(t), which is 8 sub-regions × (1 average pressure + 1 uniformity index + 1 average pressure change rate) + 1 total pressure change rate = 25 dimensions; F_pressure(t)={ , , , }; S14. Dual-supervised signal fusion of subjective rating and electromyography (EMG) data: Extract EMG features from key areas, use logistic functions to map EMG features to objective comfort ratings, and fuse them with subjective ratings through a dynamic weighting algorithm to generate a fused supervision label signal; The specific steps for fusing subjective scores and electromyographic data under dual supervision in step S14 include: By attaching electromyography (EMG) sensors to key muscle areas of the left / right shoulder, left / right hip, and left / right thigh, 6-channel EMG signals are extracted, and their root mean square values ​​are calculated to obtain an EMG feature dataset. The electromyography feature dataset is mapped to an objective discomfort score using the logistic function Sigmoid, and further converted into an electromyography objective comfort score S_emg(t) of 1-10. A dynamic weighted fusion algorithm is used to fuse subjective scores S_sub(t) and objective electromyographic comfort scores S_emg(t) to generate a fusion supervision label signal S_comb(t). The weight α of the subjective score is adjusted according to the decisiveness of the user's rating or the historical consistency between the subjective score and the electromyographic score.

[0047] Specifically, the fusion of subjective scores and electromyographic data under dual supervision. Severe localized compression and high shear force can cause discomfort to users, leading to involuntary muscle tension in an attempt to change position, resulting in elevated RMS electromyography (EMG) values. Therefore, EMG activity levels are positively correlated with discomfort. This approach uses subjective ratings as the first supervisory signal and EMG signal ratings as the second supervisory signal. By setting rules, subjective and EMG scores are fused to generate a more reliable standard supervisory label. This enhanced label is then used to train the comfort rating model. The specific implementation method is as follows: (1) First, construct the user subjective evaluation supervised dataset S_sub(t). In this invention, the supervised dataset includes subjective ratings (S_sub) given by users in different postures and objective comfort ratings (S_emg) obtained through electromyography (EMG) signals. For the subjective rating dataset, subjective evaluation indicators, standard scales, and weighting coefficients for comfort at different key body parts are established. Then, multiple users are arranged to conduct multiple sets of experiments based on the standard scales to obtain subjective data, which are then processed to obtain S_sub. The S_sub scale ranges from 1 to 10 points, with higher scores indicating greater comfort.

[0048] (2) Then, the objective comfort score of electromyography S_emg(t) is generated by quantification. ① Electromyography (EMG) signal feature extraction. An EMG signal sensor was attached to the key muscle areas of the left / right shoulder, left / right hip, and left / right thigh. Six-channel signals were extracted, and the root mean square value within a set time window was calculated to obtain the EMG feature dataset of the key areas: E_z(t)={E_shoulder_R(right shoulder), E_shoulder_L(left shoulder), E_hip_R(right hip), E_hip_L(left hip), E_thigh_R(right thigh), E_thigh_L(left thigh)}, where the subscript z∈[1,6].

[0049] ② Establish an electromyography-comfort mapping model. A sigmoid function is used for mapping, compressing the output to the (0, 1) interval, and converting the electromyography feature values ​​into an objective discomfort score, Discft_emg(t).

[0050] Among them, the value of w_0 ensures that the discomfort score is close to 0 when the muscles are completely relaxed, and the weight of each electromyographic site on the comfort effect of w_z is determined from the experimental data through regression analysis.

[0051] ③ Calculate the objective comfort score for electromyography at time t, S_emg(t). S_emg(t) is a score of 1-10, consistent with the scale of S_sub(t).

[0052] (3) Finally, the fused label supervision signal S_comb(t) is calculated. A dynamic weighted fusion method is used to fuse subjective scores S_sub(t) and objective scores S_emg(t).

[0053] Here, α is the reliability weight of subjective rating. The setting principle is as follows: if the user is decisive when giving subjective rating, the weight of S_sub α is high; if the user is hesitant, the weight of S_emg is reduced and the weight of S_emg is increased; for some users, their subjective rating and electromyography rating are always highly consistent, so a fixed high weight can be used (e.g., α=0.9); for users whose two ratings are often inconsistent, their subjective evaluation can be considered unreliable, and their α value is reduced (e.g., α=0.3).

[0054] S15. Offline model training to obtain a comfort evaluation model: The multi-dimensional pressure feature parameters, user individual parameters, environmental parameters, and the fused supervision label signal are used as sample data, and machine learning algorithms are used to train the model to obtain a comfort evaluation model.

[0055] Specifically, machine learning algorithms are applied, using the Gradient Boosting Regression Tree (GBRT) method. The above-mentioned multidimensional stress feature parameters, user individual parameters, environmental parameter variables, and fused supervision signal sample data are input for learning and training to obtain a comfort evaluation model.

[0056] In step S15, the sample data is obtained through multiple sets of experiments for different users. For users whose electromyography scores are highly consistent with their subjective scores, a large amount of sample data can be obtained using only the electromyography scores when constructing the sample dataset. The machine learning algorithm adopts the gradient boosting regression tree (GBRT) algorithm and divides the sample dataset into a training set and a validation set according to the proportion β, where the value of β ranges from [0.7, 0.9].

[0057] Specifically, define the model training input dataset, including the pressure feature vector F_pressure, user parameters BDY (weight W, age A, height H, blood pressure BP), and environmental ENV (temperature T_env, humidity H_env). Define the supervised dataset, namely the supervised label S_comb, which integrates subjective ratings and electromyography (EMG) scores.

[0058] U sample datasets were obtained from multiple experiments targeting different users. Since electromyography (EMG) data is readily and quickly available, for users whose EMG scores and subjective scores were highly consistent, the weight of the subjective scores was ignored, and only the EMG scores were used to obtain as much sample data as possible. Finally, U input feature sets X_train={F_pressure, BDY, ENV} and U target label sets Y_train={S_comb} were constructed. The evaluation model is obtained by training on sample data. The gradient boosting regression tree (GBRT) machine learning algorithm is used to learn the mapping relationship from X_train to Y_train, where β*U sample datasets are used for training and (1-β)*U samples are used for validation to obtain the comfort evaluation model MOD_A, where β generally takes the value range of [0.7, 0.9].

[0059] Step S2, Online Scoring Stage: Pressure signals from the bed / chair surface are collected in real time. These signals are processed to extract real-time multi-dimensional pressure feature parameters. These parameters, combined with individual user parameters and environmental parameters, are input into the comfort evaluation model to output a real-time comfort score. This specifically includes the following steps: S21. Multidimensional pressure feature extraction in the online application stage: Simultaneously collect pressure signals from both sides of the bed / chair, and use the same feature extraction method as in the offline stage to obtain real-time static and dynamic pressure feature parameters; Specifically, after the timing of signal acquisition on both sides of the bed / chair is synchronized, the original array signal of bed / chair surface pressure during the online application phase is obtained. Using the multi-dimensional pressure feature extraction method constructed in the first phase, a total of 25 dimensions of static and dynamic pressure feature parameters are obtained. S22. Construct a real-time online feature vector: Combine the real-time pressure feature parameters, pre-configured individual user parameters, and real-time collected environmental parameters into an online feature vector; Specifically, a feature vector is constructed for the online application stage, including stress features, individual user parameters, and environmental variable parameters. This vector is then input into the comfort evaluation model trained in the first stage to obtain an online comfort score.

[0060] S23. Comfort rating during online application phase: The online feature vector is input into the comfort evaluation model, and the model outputs a quantified comfort rating in real time.

[0061] Specifically, based on the comfort evaluation model MOD_A trained above, online real-time stress characteristics, user parameters, and environmental data are input to obtain the comfort score S_onlie(t) at time t in real time. Specifically: (1) Using the multidimensional pressure feature extraction method established during the training phase, calculate the pressure feature vector F_pressure(t) of the online phase at time t; (2) Construct a real-time online feature vector F_online(t), F_online(t) = [pressure feature F_pressure(t), weight W, age A, height H, daily blood pressure BP, ambient temperature T_env(t), ambient humidity H_env(t)]; where weight W, age A, height H, and blood pressure BP are input and configured into the system controller via a mobile APP, and ambient temperature T_env(t) and ambient humidity H_env(t) are collected in real time by temperature and humidity sensors; (3) Based on the comfort evaluation model, calculate the comfort score S_onlie(t) in real time. Input the constructed online feature vector F_online(t) into the trained comfort evaluation model MOD_A, and the model outputs a quantitative and personalized comfort score S_onlie(t)=MOD_A(F_online(t)) in real time.

[0062] The beneficial effects of this invention are: The existing equipment has preset parameters such as attitude change speed and angle, which cannot adapt to different users. The input of this solution model includes parameters such as user weight, age, and height, and the attitude adjustment is dynamically adjusted according to the individual.

[0063] Existing comfort assessments often rely on subjective descriptions from users or caregivers, resulting in coarse-grained evaluations. The solution of this invention can output a continuous quantitative score that can distinguish subtle differences in comfort.

[0064] Existing technical solutions generally only utilize the instantaneous peak pressure information, while the solution of this invention comprehensively utilizes multi-dimensional information such as the spatial distribution (uniformity) and temporal dynamics (rate of change) of pressure.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents, all of which should be included within the scope of protection of the invention.

Claims

1. A quantitative evaluation method for the comfort of intelligent beds and chairs based on multi-dimensional feature fusion, characterized in that, Includes the following steps: Offline model training phase: pressure signals from the bed and chair surfaces are collected and processed to extract multi-dimensional pressure feature parameters, while user-specific parameters and environmental parameters are also acquired; a fusion supervision signal is generated based on subjective ratings and electromyography data; a comfort evaluation model is trained using machine learning algorithms with the multi-dimensional pressure feature parameters, user-specific parameters, environmental parameters, and the fusion supervision signal as training data. Online rating phase: Pressure signals from the bed and chair surfaces are collected in real time, and the pressure signals are processed to extract real-time multi-dimensional pressure feature parameters. These parameters are then combined with individual user parameters and environmental parameters and input into the comfort evaluation model to output a real-time comfort score.

2. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 1, characterized in that, The offline model training phase specifically includes the following steps: S11. Signal acquisition and processing of array pressure sensing unit: The timing tags on both sides of the bed / chair are synchronized through a synchronization device, the raw signals of the pressure sensor array are acquired, and calibration and filtering are performed. S12. Determination of human lying posture: Based on the filtered pressure data, calculate the global pressure center coordinates and the pressure ratio of the left and right halves, and combine historical buffer data to determine whether the human body is in a supine position, and collect subsequent data when the human body is determined to be in a supine position. S13. Multidimensional pressure feature data extraction: For multiple key areas of the bed / chair surface, calculate the static pressure distribution characteristics and dynamic pressure change characteristic parameters, and construct a multidimensional pressure feature matrix; S14. Dual-supervised signal fusion of subjective rating and electromyography (EMG) data: Extract EMG features from key areas, use logistic functions to map EMG features to objective comfort ratings, and fuse them with subjective ratings through a dynamic weighting algorithm to generate a fused supervision label signal; S15. Offline model training to obtain a comfort evaluation model: The multi-dimensional pressure feature parameters, user individual parameters, environmental parameters, and the fused supervision label signal are used as sample data, and machine learning algorithms are used to train the model to obtain a comfort evaluation model.

3. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 1, characterized in that, The online scoring stage specifically includes the following steps: S21. Multidimensional pressure feature extraction in the online application stage: Simultaneously collect pressure signals from both sides of the bed / chair, and use the same feature extraction method as in the offline stage to obtain real-time static and dynamic pressure feature parameters; S22. Construct a real-time online feature vector: Combine the real-time pressure feature parameters, pre-configured individual user parameters, and real-time collected environmental parameters into an online feature vector; S23. Comfort rating during online application phase: The online feature vector is input into the comfort evaluation model, and the model outputs a quantified comfort rating in real time.

4. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 2, characterized in that, The filtering process in step S11 employs a region-adaptive Wiener filtering method based on prior knowledge. The specific steps include: The bed / chair surface is divided into multiple fixed regions. For the pressure unit in each region, the mean and variance of its local neighborhood are calculated, and the filtering intensity is adaptively adjusted based on the noise variance estimate of each region.

5. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 2, characterized in that, The specific steps for determining the human bed-lying posture in step S12 include: Calculate the global pressure center coordinates (CoP_x(t), CoP_y(t)) and the pressure ratio R_lr(t) between the left and right halves; By comparing the lateral offset of the pressure center coordinate with the threshold Threshold_x, the deviation of the left and right half-zone pressure ratio from 1 with the threshold Threshold_ratio, and combining the comparison of the longitudinal coordinate of the pressure center with the sitting-up posture threshold Threshold_sit, and the relationship between the left and right half-zone pressure ratio and the side-lying posture threshold Threshold_side, a preliminary judgment of the bed rest posture can be made. By utilizing a historical buffer that stores preliminary judgment results over a recent period, the final lying position is confirmed when the judgment results in the buffer are consistent for a consecutive predetermined period of time. When the final lying position is not a supine position, a voice reminder is triggered or the drive control system is used to force the user to lie in a supine position.

6. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 2, characterized in that, In step S13, the static pressure distribution characteristics include the average pressure and distribution uniformity index within the region, and the dynamic pressure change characteristics include the average pressure change rate and the total pressure change rate within the region. The multi-dimensional pressure feature matrix has 25 dimensions and is obtained by calculating the average pressure, distribution uniformity index, and average pressure change rate for each of the eight partitions, and by calculating a total pressure change rate.

7. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 2, characterized in that, The specific steps for fusing subjective scores and electromyographic data dual-supervision signals in step S14 include: By attaching electromyography (EMG) sensors to key muscle areas of the left / right shoulder, left / right hip, and left / right thigh, 6-channel EMG signals are extracted, and their root mean square values ​​are calculated to obtain an EMG feature dataset. The electromyography feature dataset is mapped to an objective discomfort score using the logistic function Sigmoid, and further converted into an electromyography objective comfort score S_emg(t) of 1-10. A dynamic weighted fusion algorithm is used to fuse subjective scores S_sub(t) and objective electromyographic comfort scores S_emg(t) to generate a fusion supervision label signal S_comb(t). The weight α of the subjective score is adjusted according to the decisiveness of the user's rating or the historical consistency between the subjective score and the electromyographic score.

8. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 2, characterized in that, In step S15, the sample data is obtained through multiple sets of experiments for different users. For users whose electromyography scores are highly consistent with their subjective scores, a large amount of sample data can be obtained using only the electromyography scores when constructing the sample dataset. The machine learning algorithm adopts the gradient boosting regression tree (GBRT) algorithm and divides the sample dataset into a training set and a validation set according to the proportion β, where the value of β ranges from [0.7, 0.9].

9. A method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion according to any one of claims 1, 2, or 3, characterized in that, The individual user parameters include weight, age, height, and blood pressure; the environmental parameters include ambient temperature and ambient humidity.

10. The method for quantitative evaluation of the comfort of a smart bed and chair based on multi-dimensional feature fusion as described in claim 3, characterized in that, The individual user parameters are input and pre-configured into the system controller via a mobile app; the environmental parameters are collected in real time by temperature and humidity sensors.

Images