Intelligent temperature control mattress system based on multi-modal sensing and adaptive learning

Abstract

The invention discloses an intelligent temperature control mattress system based on multi-mode sensing and adaptive learning, and the system comprises a data collection module which is used for collecting a mattress surface vibration signal, a mattress surface temperature and a mattress surrounding environment parameter; the data analysis module is used for performing deep analysis on the time-frequency domain features according to a biological recognition model to obtain human body or non-human body classification, performing self-learning to obtain body movement changes, accurately recognizing various typical sleeping postures and filtering out error data; and the model prediction module is used for monitoring the physiological parameters of the user according to the spectral analysis, inputting the physiological parameters, the collected temperature and the parameters of the surrounding environment of the mattress into a temperature control model composed of three neural networks of LSTM, GNN and Transform for comprehensive analysis and decision making, and outputting an optimal temperature control strategy. According to heart rate changes or body heat fluctuations, the sleep stage and the sleep and waking time are intelligently recognized, and therefore sleep environment adjustment better conforming to the human body rhythm is achieved.

Description

TECHNICAL FIELD

[0001] The present application belongs to an intelligent temperature control mattress system based on multi-modal sensing and adaptive learning. BACKGROUND

[0002] In recent years, the field of intelligent mattresses and intelligent temperature control mattresses has formed a clear path of technological evolution. Early products achieve temperature control adjustment through timing programs or manual settings. Further iterations of products mainly use capacitive or piezoresistive sensors to detect user usage and achieve basic temperature control through simple threshold judgments, but there is a high false trigger rate. The fundamental reason is that the sensing system cannot distinguish the differences in biomechanical characteristics of humans, pets, and heavy objects, and the temperature control is achieved through simple threshold values, resulting in low control accuracy and poor comfort. The latest technology introduces multiple detection schemes and more accurate control algorithms, which improves temperature control accuracy, but fails to establish a heat flux density dynamic model, resulting in local temperature fluctuations when the body is turned over. The overall sign detection sensor system is complex, with decreased stability and high cost. In addition, the temperature control algorithm lacks explicit adaptive learning ability and cannot generate a personalized database, which not only fails to provide better sleep comfort, but also cannot guarantee user data privacy.

[0003] There are several major systemic defects in existing technologies: 1. Poor biometric reliability: Traditional capacitive / pressure sensors have decreased biometric detection performance in different temperature and humidity environments and different mattress material conditions, or require additional wearable smart devices for sign monitoring, which is due to the insufficient robustness of single sensing modalities to complex interference.

[0004] 2. Dynamic response performance, thermal inertia control accuracy is insufficient and control lag: In the case of body turning and other state changes, the response delay of traditional solutions is large, and the local temperature fluctuation amplitude exceeds 3℃, which is much higher than the human comfort perception threshold (0.8℃). This is mainly due to the inability to establish an accurate heat flux density-circulating flow dynamic model and the lack of efficient real-time control algorithms.

[0005] 3. Lack of specific implementation of temperature control and individualization: Most existing solutions use fixed temperature curves or simple zoning control, which cannot be dynamically adjusted according to the user's real-time physiological state (such as heart rate variability, body surface temperature gradient, etc.). Clinical tests show that the efficiency of this type of solution for improving sleep quality is less than 30%, which is much lower than the 72% efficiency of individualized adjustment solutions. Existing solutions that control temperature based on sign data use real-time data for calculation, lacking the ability to follow and learn long-term changes in user body state. In addition, existing technologies lack specific implementation of individualized temperature control 4. Privacy protection is missing: the main problems are that biometric data and user identity information are not effectively isolated, there is a lack of differential privacy protection mechanism under the edge computing architecture, and the dependence on cloud platform calculation is large. In addition, the data storage and file encryption settings are unreasonable, which cannot meet the higher privacy protection requirements. These problems make the system have a serious risk of data leakage.

[0006] Therefore, an intelligent temperature control mattress system based on multi-modal sensing and adaptive learning is provided. SUMMARY

[0007] To solve the above problems in the prior art, the present application provides an intelligent temperature control mattress system based on multi-modal sensing and adaptive learning, which intelligently identifies sleep stages, sleep and wake-up timing according to heart rate changes or body heat fluctuations, thereby realizing a sleep environment adjustment that is more in line with human rhythm.

[0008] The technical solution to achieve the above-mentioned purpose is: An intelligent temperature control mattress system based on multi-modal sensing and adaptive learning, comprising: A data acquisition module for acquiring mattress surface vibration signals, mattress surface temperature and mattress surrounding environment parameters, wherein the vibration signals are processed with low-pass filtering and wavelet packet noise reduction to obtain time-domain vibration waveform and frequency spectrum feature data; A data analysis module for deep analysis of time-frequency domain features according to a biometric recognition model to obtain human or non-human classification and self-learn accurate recognition of various typical sleeping postures, and filter out error data, wherein the typical sleeping postures include supine, lateral recumbency and prone position; A model prediction module for monitoring physiological parameters of the user according to frequency spectrum analysis, inputting the physiological parameters, collected temperature and mattress surrounding environment parameters into a temperature control model composed of LSTM (Long Short-Term Memory Network), GNN (Graph Neural Network) and Transformer (Neural Network Model Based on Self-Attention Mechanism) three kinds of neural networks for comprehensive analysis and decision-making, and outputting the optimal temperature control strategy, wherein the physiological parameters include the user's heart rate and respiration rate; An instruction generation module for generating PWM (Pulse Width Modulation) control instructions by comprehensively considering the current physiological state, environmental conditions, historical data and temperature control strategy according to the model prediction control algorithm; A temperature control module for driving the impeller pump speed regulation system and the heater of the mattress according to the PWM control instructions to control the water flow and temperature; A data security storage module for adopting AES-128 real-time encryption and SHA-256 hash desensitization double protection, automatically covering all original data only after 7 days of local secure storage, and retaining necessary vectorization features for model optimization; a model optimization module, configured to continuously optimize the control parameters and thus the model through a federated learning mechanism.

[0009] Preferably, the data acquisition module comprises: a PVDF (polyvinylidene fluoride) flexible sensor array unit, configured to acquire the bed surface vibration signals in real time, and to obtain time-domain vibration waveform and frequency-domain spectrum feature data through low-pass filtering and wavelet packet denoising processing; an infrared temperature sensor, configured to form a multi-modal temperature monitoring network with a thermistor, and to calculate the bed surface heat distribution by combining an advanced heat field reconstruction algorithm to accurately model the bed surface temperature distribution; an environmental temperature and humidity sensor, configured to monitor the surrounding environmental parameters and provide environmental compensation basis.

[0010] Preferably, in the PVDF flexible sensor array unit, the vibration signals are subjected to preliminary denoising by a fourth-order Butterworth low-pass filter to remove high-frequency signals and electromagnetic interference, and the transfer function is: ; In the formula, is the frequency component, is the cut-off frequency; the acquired signals are input Laplace transform is performed first: ; In the formula, is the Laplace transform, is the time of the acquired signals; filtering is realized by frequency domain multiplication through the transfer function in the frequency domain, and the output signal in the frequency domain after filtering is: ; the signal in the time domain is output through inverse Laplace transform: ; In the formula, is the inverse Laplace transform; The Daubechies4 wavelet basis function is selected, the wavelet basis is suitable for processing the local features of biological signals, the decomposition layer is set to 5 layers of decomposition, i.e., the signal frequency band is divided into sub-frequency bands, and the signal-to-noise ratio of the reconstructed signal after threshold denoising is improved to above 42 dB: ; In the formula, is the scale, is the position of translation, is an integer set, Daubechies4 wavelet base function, is the wavelet coefficient, representing the energy of the signal at this frequency band and time point; After the wavelet packet decomposition is completed, threshold denoising is performed, and the threshold is automatically confirmed through heuristic threshold , for each sub-band coefficient Soft threshold processing is performed to obtain the denoised wavelet coefficient , the judgment condition is: ; After threshold denoising, the significant feature signal characteristics are finally retained, and random noise is suppressed, and the denoised coefficient is linearly combined with the Daubechies4 wavelet base function to reconstruct the signal : ; In the PVDF flexible sensor array unit, the frequency of the human body sign signal fluctuates over time, and short-time Fourier transform is used to generate a time-frequency spectrum matrix, which represents the signal intensity at time , and frequency A time-frequency heat map is generated, and each feature can form a highlighted area: ; In the formula, is the filtered PVDF signal time waveform, is the Hanning window function, is a complex exponential function used to calculate frequency components and detect the frequency signals present in the current segment; In the infrared temperature sensor, the temperature sensor is arranged on the control host, and the sensor array is arranged with a density of 4x4. The target temperature value is obtained by input value calculation , and the temperature measurement formula is: ; In the formula, is the sensor original reading, is the sensor material coefficient, is the skin emissivity of the measurement point, is the calibration offset; After the sensor output is CRC checked, the temperature field is reconstructed through the heat conduction equation: ; In the formula, is the offset of pipe heat dissipation; In the environmental temperature and humidity sensor, the environmental parameters are collected, and the compensation coefficient calculation formula is: ; wherein, is the ambient temperature compensation coefficient, is the ambient temperature.

[0011] Preferably, in the data analysis module, the pre-processed time-frequency domain feature matrix of the PVDF sensor , i.e. the result of 4 sensor channels, 256-dimensional features, and STFT, needs to be checked for input values, and the abnormal values are limited within ; For each sensor channel, the data needs to be standardized for evaluating the standard deviation distance of the sample point to the overall mean : ; wherein, is the mean, is the standard deviation; Feature splicing fusion and projection After grouping the left and right partitions of the sensor, the features are longitudinally spliced and merged, is a learnable weight matrix, is a linear rectifier ReLU activation function, and finally the sensor input realizes linear projection dimension reduction from 256 dimensions to 128 dimensions, and the meaningless negative response is removed through the ReLU activation function; The final output result is: The sensor realizes 2 groups combined with 2 projection modes, and the feature dimension after simultaneous fusion is 128; The human body detection branch adopts an improved ResNet18 (convolutional neural network (CNN) architecture) architecture, and finally outputs the classification results of respiration, heart rate, and body sign: ; wherein, is the classification weight matrix, which depends on the discrimination rule of non-human features, and positive weights are given to high-frequency noise and negative weights are given to feature features, is the classification bias vector; The input is the first 128-dimensional feature of the output of the feature fusion layer: ; The first step is to calculate the residual block to learn the subtle difference pattern of human and non-human vibration: ; wherein, is the residual function, which is performed in two ​1D convolution (convolution operation on one-dimensional signal) calculation and ReLU (rectified linear unit) calculation, the first layer convolution extracts local time-frequency patterns, ReLU activation removes negative responses, the second layer convolution refines features, and residual connection preserves original features; The second step attention mechanism calculation automatically amplifies important frequency band signals: ; In the formula, The function is used to compress the weight to the interval (0, 1), Global average pooling is performed; After processing, the feature is enhanced: ; In the formula, is the feature matrix weighted by attention, is the Hadamard product, is the outer product operation, which expands the vector to matrix through a 4-dimensional all-1 vector; Finally, the classification probability is output: Global feature aggregation, average the features of the four sensors to get a 128-dimensional data feature vector representing the entire mattress range: ; Sum and average to extract spatial dimension aggregation to eliminate the influence of the number of sensors; Finally, we get: ; Finally, use Softmax for normalization processing; Sleep posture recognition uses a spatio-temporal convolutional network, the input is the first 128 features of the output of the feature fusion layer, and the input data is first spatially convolved: ; The convolution kernel covers 5 time points to capture local body pressure change features, and outputs 64 feature channels to extract body pressure features of different parts, while spatial downsampling reduces the complexity of calculation; LSTM is used to model the 64-dimensional spatial features at the corresponding time to obtain the spatio-temporal fusion features at the current time: ; In the formula, is the 64-dimensional spatial feature at the current time, is the hidden state at the previous time; Discrete body posture state is fused into time-continuous change state to suppress instantaneous misjudgment, and finally 6-class posture probability distribution is output: ; where, is the body pose classification weight, is the classification bias vector, and finally normalized by Softmax; The anomaly filtering module is implemented by a variational autoencoder. The encoding process first maps the input raw time-frequency feature matrix to the latent space statistics through a three-layer fully connected network: ; ; where, is the best compressed representation of the signal in the latent space, describes the uncertainty of each dimension; The latent variables are sampled by the reparameterization trick: ; This operation allows gradient backpropagation while maintaining randomness, where each component of is independently sampled from a standard normal distribution; The decoder maps to , whose training objective is guided by the following loss function: ; where, is the variational autoencoder loss function, is the weight coefficient of the KL term; The reconstruction loss forces the model to accurately recover the biometric feature components of the input signal, while the KL divergence term constrains the latent space distribution to be close to the standard normal distribution to avoid overfitting; When the reconstruction error exceeds the dynamic threshold, the current signal is determined to be interference and the filtering mechanism is triggered.

[0012] Preferably, in the model prediction module, the temperature control model includes a hydraulic model and a thermodynamic model, the hydraulic model is used to output an optimal impeller pump speed regulation strategy, and the thermodynamic model is used to output an optimal heater power adjustment strategy, and the two strategies constitute an optimal temperature control strategy. In the model prediction module, high-precision sign monitoring is achieved through multi-modal signal fusion. First, the PVDF sensor signal is preprocessed, where is the number of sampling points; After eliminating high-frequency noise and baseline drift by 0.1-3Hz band-pass filtering, the Welch power spectrum estimation algorithm is used to calculate the signal frequency domain features: ; In the formula, For frequency index, For the number of segments, The time-domain waveform of the filtered PVDF signal is shown below. For the Hanning window function, Complex exponential function; By finding spectral peaks Determine heart rate. Determine the respiratory rate; The sleep stage classifier uses a three-channel LSTM network to process temporal features: ; In the formula, For normalized temperature characteristics, For heart rate variability, The activity index; The output includes: a 5-dimensional probability distribution of sleep stages. Corresponding to the stages of wakefulness, REM, and N1-N3; Metabolic rate calculation integrates user physiological parameters and environmental data: ; In the formula, For weight, For height, For age, Weighted average of temperatures from 4 sensors value, The sleep stage coefficients are: wakefulness = 0, REM = 0.1, N3 = 0.3; This output forms a feedback loop with the temperature control; when a temperature is detected... A decrease of 15% and When this happens, the energy-saving mode is triggered to reduce heating power.

[0013] Preferably, in the model prediction module, the temperature control model composed of three neural networks, LSTM, GNN and Transformer, performs comprehensive analysis and decision-making to generate the optimal temperature control strategy. Its input feature vector It is composed of three parts of features: physiological parameter vector Including real-time heart rate respiratory rate Sleeping position coding Temperature field data Temperature measurements corresponding to the 16 zones of the mattress, environmental compensation vector. Integrated ambient temperature, relative humidity and user activity intensity coefficient ; The three sub-feature spaces are first normalized to zero mean and unit variance by a feature scaling layer: ; where the covariance matrix and the mean are updated by online sliding window statistics; The normalized features are input into three neural networks in parallel: The LSTM temporal network models the time-varying characteristics of the parameters through a gating mechanism , whose hidden state captures the dynamic pattern in the last 2 minutes, and outputs the control quantity , where is the body pose classification weight based on the LSTM temporal network, is the classification bias vector based on the LSTM temporal network; The GNN spatial network constructs a temperature field graph structure , where the features of the nodes include the partition temperature and the distance from the human core , and the edge weight encodes the spatial correlation of heat conduction, where is the normalized distance from the th partition to the human core, and the spatial control quantity is generated by two layers of graph convolution, where is the degree matrix, is the adjacency matrix with self-loop, is the node feature matrix, is the trainable graph convolution weight matrix, is the original adjacency matrix, is the identity matrix; The Transformer fusion network concatenates the above outputs into , calculates the cross-network feature correlation matrix through the multi-head attention mechanism, and finally outputs , which focuses on the key decision dimension, where is the query vector, is the key vector, is the scaling factor, is the value matrix; The arbitration adopts a dynamic weighting strategy that is adaptive to temperature: ; where the weight is: ; where is the temperature-sensitive coefficient; The mean square error of the last 5 decisions of each network is dynamically adjusted: ; In the formula, is the network The decision accuracy at time , is the time step index, is the current time, is the network predicted temperature, is the actual measured temperature; The arbitration output contains 16 partition target temperatures and corresponding actuator control amounts, forming a closed loop with the instruction generation module of the back end.

[0014] Preferably, in the instruction generation module, multi-step advanced optimization control is realized by establishing a dynamic state space model of the mattress temperature field, and a discrete state space equation is established based on the first law of thermodynamics: ; ; ; In the formula, the state vector is the temperature value of 16 partitions, the control input is the estimated value of 16 waterway flow and the corresponding heating sheet power instruction, is the water pump flow setting value, is the power instruction for temperature control, the disturbance vector quantifies the influence of environmental temperature and humidity and user activity intensity, is the user real-time activity intensity coefficient, is the model uncertainty, is the temperature state transition matrix, is the control input matrix, is the disturbance input matrix; is a symmetric strip matrix, and its non-zero elements are: ; In the formula, is the thermal conductivity of the material, is the material density, is the material volume, is the specific heat capacity, is the contact area of adjacent partitions, is the spacing; The control matrix is obtained by fitting the fluid heat transfer experimental data: ; wherein, is the first water flow at the first is the maximum allowed temperature rise; Solve the following optimization problem in each control period: ; The constraints are: ; ; ; ; wherein, is the prediction horizon, is the control horizon, is the weight matrix, is to ensure the control smoothness, is the penalty coefficient, is the slack variable, is the physical lower limit of the actuator, is the physical upper limit of the actuator, is the first temperature change of the first is the control period, is the personalized temperature comfort interval; The reference temperature is dynamically generated by the multi-network arbitration: ; wherein, is the basic temperature field, is the physiological compensation term, is the sleep posture compensation matrix, is the real-time physiological vector, is the resting reference value, is the sleep posture state code; The control quantity output by the model predictive control is converted into a PWM signal and heating power, and the signals in the state estimator are updated for initialization in the next period, forming a complete closed-loop control with the temperature control module in the back end.

[0015] Preferably, in the temperature control module, the impeller pump adopts a dynamic speed regulation strategy to control the operating noise below 30dB while ensuring the required flow; the heater is adjusted by closed-loop PID to ensure that the water temperature control accuracy reaches ±0.2℃; the waterway pressure and water temperature data generated during the execution are fed back to the temperature control module in real time for online updating of the hydraulic model and the thermodynamic model, forming a complete control closed loop; The output optimized instructions are converted into actual equipment drive signals, and temperature field control is achieved by precisely adjusting the water circuit system and heating elements; For PWM control of impeller pumps, the first step is to determine the flow rate requirement. Calculate the target rotational speed: ; In the formula, This is the pump characteristic coefficient. This represents the current pipeline pressure difference, i.e., the pressure compensation item. Speed ​​commands are converted into PWM duty cycles via a PI controller: ; ; In the formula, For speed error, This represents the historical value of the rotational speed error. For the target speed, This refers to the actual rotational speed; The heater power control adopts a feedforward-feedback composite strategy: feedforward term Directly compensate for environmental heat loss, among which, For the target temperature, The ambient temperature; Feedback PID item Eliminate steady-state error, For temperature deviation, ,in, This is the measured temperature of the current area; To prevent overheating, the final power limit is: ; All commands are transmitted via internal bus communication at a frequency of 100Hz, while real-time feedback from sensors is received, forming a closed-loop control. Real-time feedback optimization achieves continuous improvement in closed-loop control accuracy through a hydraulic-thermal coupled observer. Its core lies in dynamically correcting the deviation between model prediction and actual execution. The observer, built based on a state-space model, operates on a two-stage mechanism of temperature prediction and measurement correction, with the following equation: ; In the formula, This is the estimated mattress temperature state vector. This is the actual temperature sensor reading. Here is the observer gain matrix, which is dynamically updated by solving the Riccati equations online: ; In the formula, is a state transition matrix, is a state transition matrix transpose, is a solution matrix of algebraic Riccati equation, is a measurement noise covariance matrix, is a transpose of observation model matrix, is an observation model matrix, is a process noise covariance; The impeller pump hydraulic control first establishes a dynamic three-coupling model of flow-rate-speed-pressure: ; Speed and flow-rate mapping relationship through online identification of pipeline characteristic parameters 、 Dynamic correction, pressure compensation term Real-time reflection of system impedance changes; To suppress water hammer effect, use pressure differential feedback Combined with second-order Butterworth filter Smooth control command, Target speed, Rate of change of pipeline pressure, Transformation operator in discrete control; The motor drive end eliminates electromagnetic interference through back electromotive force compensation ; The heater control adopts a multi-modal PID algorithm to achieve a balance between fast response and steady-state accuracy: ; User comfort learning automatically adjusts the zone temperature setting through optimizing the objective function: ; Where, Target temperature value, Temperature change rate, Measured foot area temperature value, Measured head area temperature value; Every 30 minutes, the hydraulic model is updated by recursive least squares method, and the conduction coefficient, convection coefficient and metabolic influence coefficient of the thermodynamic model are corrected based on finite element analysis.

[0016] Preferably, in the data security storage module, all sensor raw data is encrypted in real time by hardware-accelerated AES-128 algorithm during the data acquisition stage, and the encryption process can be represented as: ; Where, For using AES-128-CTR encryption algorithm, For sensor data, For device unique key, For counter mode, For SHA-256 based keyed-hash message authentication code, For device fingerprint generated by physical unclonable function, For timestamp accurate to millisecond; User identity information is desensitized by double-layer hashing: In the formula, is an intercept function that retains the first 16 bits of the hash value, is a MAC address, is a random salt value.

[0017] Preferably, in the model optimization module, the federated learning mechanism realizes cross-device collaborative optimization through triple privacy protection: First, the temperature field is compressed into 4-dimensional latent features using an autoencoder: In the formula, is a temperature field encoder; Its reconstruction loss Retains key heat distribution patterns: In the formula, is a decoder for reconstructing the temperature field, is the temperature; Then, the physiological parameters are reduced to 2-dimensional principal components through PCA: In the formula, is a physiological parameter projection matrix, is the corresponding physiological parameter; Finally, the control instruction is SHA-256 hashed and desensitized: In the formula, is a hash desensitization operation; The uploaded gradient information is disturbed by Laplace noise Satisfies the differential privacy requirement, controls the information leakage risk of the privacy budget, and the central server generates a global model after aggregating the gradients of each device: In the formula, is the number of participating devices, is the device in the ​​​​​​Train the model parameters after the wheel; And transfer global knowledge to local model through knowledge distillation Its distillation loss Balance model performance and complexity: In the formula, The divergence loss is, The output probability difference distribution is, The local model trainable parameters.

[0018] Compared with the prior art, the beneficial effects of the present application are: the present application intelligently identifies sleep stages, sleep and wake-up opportunities according to heart rate changes or body heat fluctuations, thereby realizing a sleep environment adjustment that is more in line with human rhythm, solving the problem of insufficient precision of intelligent temperature control strategy, reducing the complexity and cost of the intelligent temperature control system architecture, and increasing data security; Through multi-modal sensor data collection of users and environment, the sleep temperature management of users is realized, sleep stages, sleep and wake-up opportunities are intelligently identified according to heart rate changes or body heat fluctuations, thereby realizing a sleep environment adjustment that is more in line with human rhythm; the biological recognition precision is improved through the combination of low-cost PVDF sensor array and body heat sensor and deep neural network; a thermal-hydraulic coupling dynamic model is constructed to realize high-precision temperature regulation; an incremental learning framework is adopted to establish personalized temperature strategy, and the quality of sleep and sleep comfort are improved according to the data increment of stable period; an edge privacy protection framework is set up to improve the security of private data. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 is a module diagram of an intelligent temperature control mattress system based on multi-modal sensing and adaptive learning according to the present application; Figure 2 is a detailed architecture diagram of the intelligent temperature control mattress system according to the present application; Figure 3 is a specific module diagram of the data acquisition module according to the present application; Figure 4 is a network architecture diagram of the biometric identification module according to the present application; Figure 5 is a realization framework diagram of the temperature control model composed of LSTM, GNN and Transformer three kinds of neural networks through multi-network arbitration system for comprehensive analysis and decision-making according to the present application; Figure 6 is a framework diagram of edge encryption and data storage according to the present application;​ Figure 7 This is a schematic diagram of the federated learning framework and multi-model co-evolution architecture in this invention. Detailed Implementation

[0020] 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.

[0021] like Figure 1 , 2 As shown, an intelligent temperature-controlled mattress system based on multimodal sensing and adaptive learning includes: a data acquisition module 1, a data analysis module 2, a model prediction module 3, an instruction generation module 4, a temperature control module 5, a data security storage module 6, and a model optimization module 7.

[0022] Data acquisition module 1 is used to acquire vibration signals from the mattress surface, mattress surface temperature, and environmental parameters around the mattress. The vibration signals are processed with low-pass filtering and wavelet packet noise reduction to obtain time-domain vibration waveforms and frequency-domain spectral characteristic data.

[0023] like Figure 3 As shown, the data acquisition module 1 includes: a PVDF flexible sensor array unit 11, an infrared temperature sensor 12, and an ambient temperature and humidity sensor 13.

[0024] The PVDF flexible sensor array unit 11 is used to acquire vibration signals from the mattress surface in real time. With the help of low-pass filtering and wavelet packet noise reduction processing, time-domain vibration waveforms and frequency-domain spectral characteristic data are obtained.

[0025] In this embodiment, the vibration signal is initially denoised using a fourth-order Butterworth low-pass filter to remove high-frequency signals and electromagnetic interference. Its transfer function is: ; In the formula, For frequency components, The cutoff frequency; Input acquired signal First, perform the Laplace transform: ; In the formula, For Laplace transform, The time for signal acquisition; In the frequency domain, filtering is achieved through frequency multiplication using the transfer function. The filtered output signal is represented in the frequency domain as follows: ; The signal in time domain is outputted by inverse Laplace transform again: ; In the formula, is inverse Laplace transform; Daubechies4 wavelet base function is selected, the wavelet base is suitable for processing local characteristics of biological signals, the decomposition layer is set to 5 layers of decomposition, i.e. the signal frequency band is divided into sub-frequency bands, and the signal-to-noise ratio of the reconstructed signal after threshold denoising is improved to more than 42dB: ; In the formula, is scale, is the position of translation, is an integer set, is Daubechies4 wavelet base function, is wavelet coefficient, indicating the energy of the signal at the frequency band and time point; After completing wavelet packet decomposition, threshold denoising is performed, and the threshold is automatically confirmed by heuristic threshold Soft threshold processing is performed on each sub-band coefficient , and the judgment condition is: ; After threshold denoising, the significant characteristic signal features are finally retained, and the random noise is suppressed, and the denoised coefficient is linearly combined with the Daubechies4 wavelet base function to reconstruct the signal : ; In the embodiment, the frequency of the human body characteristic signal fluctuates over time, short-time Fourier transform is adopted to generate a time-frequency spectrum matrix, and the numerical value represents the signal intensity at time , and the frequency is A time-frequency heat map is generated, and each feature can form a highlighted area: ; In the formula, is the time-domain waveform of the filtered PVDF signal, is the Hanning window function, is a complex exponential function, which is used to calculate the frequency component and detect the frequency signal existing in the current segment.

[0026] Infrared temperature sensor 12, for precise modeling of the bed surface temperature distribution by forming a multi-modal temperature monitoring network with thermistors, combined with advanced thermal field reconstruction algorithms, calculating the bed surface heat distribution.

[0027] In the embodiment, temperature sensors are arranged on the control host, and the sensor array is arranged at a density of 4x4. The target temperature value is obtained by input value calculation , and the temperature measurement formula is: ; In the formula, is the sensor raw reading, is the sensor material coefficient, is the skin emissivity of the measurement point, is the calibration offset; After the sensor output is checked by CRC, the temperature field is reconstructed by the heat conduction equation: ; In the formula, is the offset of pipe heat dissipation; The environmental temperature and humidity sensor collects environmental parameters, and the compensation coefficient calculation formula is: ; In the formula, is the environmental temperature compensation coefficient, is the environmental temperature.

[0028] The environmental temperature and humidity sensor 13 is used to monitor the surrounding environmental parameters and provide environmental compensation basis.

[0029] The data analysis module 2 is used to deeply analyze the time-frequency domain features according to the biological recognition model, obtain the human or non-human classification, and self-learn the accurate recognition of various typical sleeping postures. The error data is filtered out, such as Figure 4 as shown, wherein the typical sleeping postures include supine, lateral recumbency and prone.

[0030] In the embodiment, the time-frequency domain feature matrix of the input PVDF sensor after preprocessing , that is, 4 sensor channels, 256-dimensional features, and the result of STFT. The input value needs to be checked, and the abnormal value is limited within : ; For each sensor channel, the data needs to be standardized for evaluating the standard deviation distance of the sample point to the overall mean value : ; In the formula, is the average value, is the standard deviation; Feature concatenation fusion and projection , the features are longitudinally spliced and merged after grouping the left and right partitions of the sensor, is a learnable weight matrix, is a linear rectifier ReLU activation function, and finally the sensor input realizes linear projection dimension reduction from 256 dimensions to 128 dimensions, and the meaningless negative response is removed through the ReLU activation function; The final output result is: The sensor realizes 2 groups combined with 2 projection modes, and the feature dimension after fusion is 128; The human body detection branch adopts the improved ResNet18 architecture, and finally outputs the classification results of respiration, heart rate and body sign: ; In the formula, is a classification weight matrix, which depends on the discrimination rule of non-human features. Positive weight is given to high-frequency noise, and negative weight is given to characteristic features, is a classification bias vector; The input is the first 128-dimensional feature of the output of the feature fusion layer: ; The first step is to calculate the residual block, which learns the subtle difference pattern of human and non-human vibration through residual learning: ; In the formula, is a residual function, which performs two-layer 1D convolution calculation and ReLU calculation. The first layer of convolution extracts local time-frequency patterns, and ReLU activation removes negative responses. The second layer of convolution refines the features, and the residual connection retains the original features; The second step is to calculate the attention mechanism to automatically amplify important frequency band signals: ; In the formula, function is used to compress the weight to the interval (0, 1), is the global average pooling processing; After processing, the features are enhanced: ; In the formula, is the feature matrix weighted by attention, is the Hadamard product, is the outer product operation, which expands the vector to a matrix through a 4-dimensional all-1 vector; Finally, the classification probability output is: Global feature aggregation, average the features of the 4 sensors to get a 128-dimensional data feature vector representing the entire mattress range: ; Sum and average to extract spatial dimensions after aggregation to eliminate the influence of the number of sensors; Finally, we get: ; Finally, use Softmax for normalization processing; Sleep posture recognition uses a spatio-temporal convolutional network, the input is the first 128-dimensional feature of the output of the feature fusion layer, and the input data is first subjected to spatial convolution: ; The convolution kernel covers 5 time points to capture local body pressure change features, and outputs 64 feature channels to extract body pressure features of different parts, while spatial downsampling reduces the complexity of calculation; LSTM is used for time modeling of the 64-dimensional spatial features corresponding to the moment, and the spatio-temporal fusion features of the current moment are obtained: ; In the formula, is the 64-dimensional spatial feature of the current moment, is the hidden state of the previous moment; Discrete body posture state is fused into time-continuous change state to suppress instantaneous misjudgment, and finally 6-class posture probability distribution is output: ; In the formula, is the body posture classification weight, is the classification bias vector, and finally Softmax is used for normalization processing; The anomaly filtering module is implemented through a variational autoencoder. Its encoding process first maps the input original time-frequency feature matrix to the statistical quantity of the latent space through a three-layer fully connected network: ; ; In the formula, is the best compressed representation of the signal in the latent space, describes the uncertainty of each dimension; The latent variable is sampled through the reparameterization trick: ; This operation allows gradient backpropagation while maintaining randomness, where each component of is independently sampled from a standard normal distribution; decoder Will Reconstructed Its training objective is guided by the following loss function: ; In the formula, The loss function of the variational autoencoder is... These are the weighting coefficients for the KL terms; The reconstruction loss forces the model to accurately recover the biological features of the input signal, while the KL divergence term constrains the latent space distribution to approximate a standard normal distribution. To avoid overfitting; When reconstruction error When the dynamic threshold is exceeded, the current signal is determined to be interference and the filtering mechanism is triggered.

[0031] Model prediction module 3 monitors the user's physiological parameters based on spectral analysis. These physiological parameters, along with collected temperature and ambient mattress parameters, are input into a temperature control model composed of LSTM, GNN, and Transformer neural networks integrated by a multi-network arbitration system. The model performs comprehensive analysis and decision-making, outputting the optimal temperature control strategy. Figure 5 As shown, the physiological parameters include the user's heart rate and respiratory rate.

[0032] In this embodiment, the temperature control model includes a hydraulic model and a thermodynamic model. The hydraulic model is used to output the optimal impeller pump speed regulation strategy, and the thermodynamic model is used to output the optimal heater power regulation strategy. The two strategies together constitute the optimal temperature control strategy.

[0033] In this embodiment, high-precision vital sign monitoring is achieved through multimodal signal fusion. First, the PVDF sensor signal is analyzed. Preprocessing is performed, in which This represents the number of sampling points; After eliminating high-frequency noise and baseline drift through 0.1-3Hz bandpass filtering, the frequency domain characteristics of the signal are calculated using the Welch power spectrum estimation algorithm. ; In the formula, For frequency index, For the number of segments, The time-domain waveform of the filtered PVDF signal is shown below. For the Hanning window function, Complex exponential function; By finding spectral peaks Determine heart rate. Determine the respiratory rate; The sleep stage classifier uses a three-channel LSTM network to process temporal features: ; wherein, is the normalized temperature feature, is the heart rate variability, is the motion index; The output contains: 5-dimensional sleep stage probability distribution corresponding to wake, REM, N1-N3 stages; The metabolic rate calculation fuses user physiological parameters and environmental data: ; wherein, is the body weight, is the height, is the age, is the 4-sensor temperature weighted average value, is the sleep stage coefficient, wake = 0, REM = 0.1, N3 = 0.3; This output forms a feedback loop with the temperature control, when a 15% decrease and is detected, the energy saving mode is triggered to reduce the heating power.

[0034] In the embodiment, the temperature control model composed of three kinds of neural networks, LSTM, GNN and Transformer, makes comprehensive analysis and decision, and generates the optimal temperature control strategy; Its input feature vector is composed of three parts of features: physiological parameter vector , including real-time heart rate , respiration rate and sleep posture encoding , temperature field data , corresponding to the temperature measurement values of the 16 partitions of the mattress, environmental compensation vector , integrating environmental temperature, relative humidity and user activity intensity coefficient ; The three sub-feature spaces are first normalized to zero mean and unit variance by the feature scaling layer: ; Wherein, the covariance matrix and the mean are updated by online sliding window statistics; The normalized features are input into the three neural networks in parallel: The LSTM time series network models the time-varying characteristics of the physiological parameters through the gating mechanism , whose hidden state captures the dynamic pattern of the last 2 minutes, and outputs the control quantity In the formula, For pose classification weights based on LSTM temporal networks, This is the classification bias vector based on the LSTM temporal network; GNN spatial network constructs temperature field map structure ,node Features include zone temperature and distance from the human body core edge weight The spatial correlation of encoding heat conduction, where, For the first The normalized distance from each partition to the human core is used to generate spatial control quantities through two layers of graph convolution. ,in, For degree matrix, Given an adjacency matrix with self-loops, The node feature matrix, For trainable graph convolution weight matrices, This is the original adjacency matrix. It is the identity matrix; The Transformer fusion network concatenates the above outputs into... The cross-network feature correlation matrix is ​​calculated using a multi-head attention mechanism. Final output Focusing on key decision-making dimensions, among which, For query vector, For key vectors, Scaling factor It is a value matrix; The arbitration employs a temperature-adaptive dynamic weighting strategy: ; Wherein, weight: ; In the formula, This refers to the temperature sensitivity coefficient. Dynamically adjusted based on the mean square error of the last 5 decisions of each network: ; In the formula, For the network At any moment Decision accuracy at that time For time step index, For the current moment, For the network Predicted temperature, The actual measured temperature; The arbitration output The target temperature and corresponding actuator control amount of 16 partitions form a closed loop with the instruction generation module 4 of the back end.

[0035] The instruction generation module 4 is used to generate PWM control instructions by comprehensively considering the current physiological state, environmental conditions, historical data, and temperature control strategy according to the model predictive control algorithm.

[0036] In the embodiment, multi-step advanced optimization control is realized by establishing a dynamic state space model of the mattress temperature field, and a discrete state space equation is established based on the first law of thermodynamics: ; ; ; In the formula, the state vector is the temperature value of 16 partitions, the control input is the estimated value of 16 waterway flow and the corresponding heating sheet power instruction, is the water pump flow setting value, is the power instruction of temperature control, the disturbance vector quantifies the influence of environmental temperature and humidity and user activity intensity, is the real-time activity intensity coefficient of the user, is the model uncertainty, is the temperature state transition matrix, is the control input matrix, is the disturbance input matrix; is a symmetric strip matrix, and the non-zero elements are: ; In the formula, is the thermal conductivity of the material, is the material density, is the material volume, is the specific heat capacity, is the contact area of adjacent partitions, is the spacing; The control matrix is obtained by fitting the fluid heat transfer experimental data: ; In the formula, is the water flow at the th position, is the maximum allowed temperature rise; The following optimization problem is solved at each control period: ; The constraint is: ; ; ; ; In the formula, To predict the time domain, To control the time domain, This is the weight matrix. To ensure smooth control, The penalty coefficient is... As slack variables, This is the physical lower limit of the actuator. For the physical connection of the actuator, For the first Temperature change in the zone To control the cycle, For personalized temperature comfort range; The connection with multi-network arbitration is reflected in the reference temperature. Dynamic generation mechanism: ; In the formula, Based on the basic temperature field, This is a physiological compensation item. For sleeping posture compensation matrix, This is a real-time physiological vector. This is the resting baseline value. Encoding sleeping posture status; The control quantity output by the model predictive control is converted into a PWM signal and heating power, and the signal in the state estimator is updated for the initialization of the next cycle, forming a complete closed-loop control with the back-end temperature control module 5.

[0037] Temperature control module 5 is used to drive the impeller pump speed regulation system and heater of the mattress according to PWM control instructions, and to control the water flow and temperature.

[0038] In this embodiment, the impeller pump adopts a dynamic speed regulation strategy to control the operating noise below 30dB while ensuring the required flow rate; the heater ensures that the water temperature control accuracy reaches ±0.2℃ through closed-loop PID regulation; the water pressure and water temperature data generated during the execution process are fed back to the temperature control module in real time for online updates of the hydraulic model and thermodynamic model, forming a complete control closed loop. The output optimized instructions are converted into actual equipment drive signals, and temperature field control is achieved by precisely adjusting the water circuit system and heating elements; For PWM control of impeller pumps, the first step is to determine the flow rate requirement. Calculate the target rotational speed: ; where, is the pump characteristic coefficient, is the current differential pressure, i.e. the pressure compensation term; The speed command is converted to PWM duty cycle by a PI controller: ; ; where, is the speed error, is the history of the speed error, is the target speed, is the actual speed; The heater power control employs a feedforward-feedback compound strategy: The feedforward term directly compensates for the environmental heat loss, where, is the target temperature, is the ambient temperature; The feedback PID term eliminates the steady-state error, is the temperature deviation, where, is the measured temperature of the current zone; To prevent overheating, the final power limit is: ; All commands are sent through the internal bus communication at a frequency of 100 Hz, while receiving real-time feedback from sensors, forming a closed-loop control; Real-time feedback optimization achieves continuous improvement of closed-loop control accuracy through a hydraulic-thermal coupling observer. The core is to dynamically correct the deviation between model prediction and actual execution. The observer based on the state space model works in a two-stage mechanism of temperature prediction-measurement correction, and the equation is: ; where, is the estimated mattress temperature state vector, is the actual temperature sensor reading, is the observer gain matrix, which is dynamically updated by solving the Riccati equation online: ; where, is the state transition matrix, is the transpose of the state transition matrix, is the solution matrix of the algebraic Riccat equation, is the measurement noise covariance matrix, is the transpose of the observation model matrix, is the observation model matrix, is the process noise covariance; The impeller pump hydraulic control first establishes a dynamic three-coupling model of flow-rate-speed-pressure: ; The speed and flow-rate mapping relationship is dynamically corrected through online identification of pipeline characteristic parameters , The pressure compensation term reflects the real-time change of system impedance; To suppress water hammer effect, pressure differential feedback combined with a second-order Butterworth filter smooths the control command, where is the target speed, is the rate of change of pipeline pressure, is the transformation operator in discrete control; The motor drive end eliminates electromagnetic interference through back electromotive force compensation ; The heater control adopts a multi-modal PID algorithm to achieve a balance between fast response and steady-state accuracy: ; User comfort learning automatically adjusts the zone temperature setting through optimization of the objective function: ; where is the target set temperature value, is the temperature change rate, is the measured temperature value of the foot area, is the measured temperature value of the head area; Every 30 minutes, the hydraulic model is updated by recursive least squares, and the conduction coefficient, convection coefficient, and metabolic influence coefficient of the thermodynamic model are corrected based on finite element analysis.

[0039] The data security storage module 6 is used to adopt AES-128 real-time encryption and SHA-256 hash desensitization double protection, automatically overwrite all original data after 7 days of local secure storage, and retain necessary vectorization features for model optimization.

[0040] As shown in Figure 6 , in the data acquisition stage, all sensor raw data are encrypted in real time through hardware-accelerated AES-128 algorithm, and the encryption process can be represented as: ; where For using AES-128-CTR encryption algorithm, For sensor data, For device unique key, For counter mode, For SHA-256 based keyed-hash message authentication code, For device fingerprint generated by physically unclonable function, For time stamp accurate to millisecond; User identity information is desensitized by double-layer hashing: ; In the formula, is, is the MAC address, is a random salt value.

[0041] The model optimization module 7 is used to continuously optimize the control parameters through the federated learning mechanism, thereby optimizing the model.

[0042] As Figure 7 shown, the federated learning mechanism realizes cross-device collaborative optimization through triple privacy protection: First, the temperature field is compressed into 4-dimensional latent features by using the autoencoder: ; In the formula, is the temperature field encoder; The reconstruction loss of the autoencoder retains the key heat distribution pattern: ; In the formula, is the decoder for reconstructing the temperature field, is the temperature; Then, the physiological parameters are reduced to 2-dimensional principal components by PCA: ; In the formula, is the physiological parameter projection matrix, is the corresponding physiological parameter; Finally, the control instruction is subjected to SHA-256 hashing desensitization: ; In the formula, is the hashing desensitization operation; The uploaded gradient information is disturbed by Laplace noise to meet the differential privacy requirement, control the information leakage risk of the privacy budget, and the central server generates a global model after aggregating the gradients of all devices: ; In the formula, the number of participating devices, the model parameters of the device after the round of training; and migrate global knowledge to local models through knowledge distillation with distillation loss balance model performance and complexity: wherein, divergence loss, output probability difference distribution, local model trainable parameters.

[0043] The system first collects the mattress surface vibration signals in real time through the sensing layer PVDF flexible sensor array, cooperates with low-pass filtering and wavelet packet noise reduction processing, effectively eliminates environmental noise interference, provides high-quality time-domain vibration waveform and frequency-domain spectrum feature data for subsequent analysis, and further identifies human body sign data and human sleep state. At the same time, the system uses a multi-modal temperature monitoring network composed of infrared temperature sensors and thermistors, combined with advanced thermal field reconstruction algorithms, to realize accurate modeling of the temperature distribution on the mattress surface. The environmental temperature and humidity sensor continuously monitors the surrounding environmental parameters to provide necessary environmental compensation basis for the temperature control strategy.

[0044] At the data processing level, the system performs in-depth analysis on the time-frequency domain features collected by the PVDF sensor through a special biological recognition model, which not only can distinguish human body from other interference sources with extremely high accuracy, but also can accurately identify a variety of typical sleeping postures through self-learned body movement changes. Based on the spectrum analysis of the PVDF signal, the real-time monitoring of the user's heart rate, respiratory rate and other key physiological parameters is further realized. These physiological data, together with temperature field information and environmental parameters, are input into a multi-network arbitration system, and a hybrid model composed of LSTM, GNN and Transformer three kinds of neural networks is used for comprehensive analysis and decision-making to ensure the optimality of the temperature control strategy.

[0045] The core control part of the system adopts advanced model predictive control (MPC) algorithm, which generates accurate PWM control instructions by considering the current physiological state, environmental conditions and historical data. These instructions drive the impeller pump speed regulation system and the heater power regulation module at the same time, realizing accurate control of water flow and temperature. The impeller pump adopts a dynamic speed regulation strategy to control the operating noise below 30dB while ensuring the required flow; the heating system ensures the water temperature control accuracy to ±0.2℃ through closed-loop PID regulation. The waterway pressure and water temperature data generated during execution are fed back to the control system in real time for online updating of the hydraulic model and thermodynamic model, forming a complete control closed loop.

[0046] ​To continuously optimize system performance, the present scheme designs a unique incremental learning framework. The system takes 7 days as a learning cycle, continuously optimizes the control parameters through the federal learning mechanism, and at the same time strictly protects the user privacy. In terms of data security, the system adopts AES-128 real-time encryption and SHA-256 hash desensitization double protection, all original data are automatically overwritten after being stored locally for 7 days, only necessary vectorized features are reserved for model optimization. This full closed-loop design from data collection to decision execution to continuous learning enables the system to achieve accurate temperature control and accurate biological feature recognition while maintaining low response time, significantly better than existing technical solutions.

[0047] Finally, it should be noted that: the above is only the preferred embodiment of the present application, and is not used to limit the present application, although the present application is described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical scheme recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent temperature control mattress system based on multi-modal sensing and adaptive learning, characterized in that, The method comprises the following steps: A data acquisition module is used to collect bed surface vibration signals, bed surface temperature, and bed surrounding environment parameters, wherein the vibration signals are subjected to low-pass filtering and wavelet packet noise reduction processing to obtain time-domain vibration waveform and frequency-domain spectrum feature data; A data analysis module is used to perform deep analysis on the time-frequency domain features according to a biological recognition model, to obtain human body or non-human body classification, to self-learn accurate recognition of various typical sleeping postures from body movement changes, and to filter out error data, wherein the typical sleeping postures include supine, lateral recumbency, and prone position; A model prediction module is used to monitor physiological parameters of a user according to frequency spectrum analysis, to input the physiological parameters, collected temperature, and bed surrounding environment parameters into a temperature control model composed of LSTM, GNN, and Transformer three kinds of neural networks for comprehensive analysis and decision-making, and to output an optimal temperature control strategy, wherein the physiological parameters include user heart rate and respiration rate; An instruction generation module is used to generate PWM control instructions by comprehensively considering current physiological state, environmental conditions, historical data, and temperature control strategy according to a model prediction control algorithm; A temperature control module is used to drive the impeller pump speed regulation system and the heater of the bed according to the PWM control instructions to control water flow and temperature; A data security storage module is used to adopt AES-128 real-time encryption and SHA-256 hash desensitization double protection, to automatically overwrite all original data after 7 days of local secure storage, and to retain necessary vectorization features for model optimization; A model optimization module is used to continuously optimize control parameters through a federal learning mechanism, and to optimize the model.

2. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 1, wherein, The data acquisition module comprises: A PVDF flexible sensor array unit is used to collect bed surface vibration signals in real time, and the vibration signals are subjected to low-pass filtering and wavelet packet noise reduction processing to obtain time-domain vibration waveform and frequency-domain spectrum feature data; An infrared temperature sensor is used to form a multi-modal temperature monitoring network with a thermistor, to accurately model the bed surface temperature distribution by combining an advanced thermal field reconstruction algorithm, and to calculate the bed surface heat distribution; An environmental temperature and humidity sensor is used to monitor surrounding environment parameters to provide environmental compensation basis. 3.The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 2, wherein, In the PVDF flexible sensor array unit, the vibration signals are subjected to preliminary noise reduction by a fourth-order Butterworth low-pass filter to remove high-frequency signals and electromagnetic interference, and the transfer function is: ； wherein is the frequency component, is the cut-off frequency; Inputting the collected signal First, Laplace transform is performed: ； wherein is the Laplace transform, is the time of acquisition of the signal; In the frequency domain, the filter is realized by the transfer function through frequency domain multiplication, and the output signal in the frequency domain after filtering is: ； Then the signal in the time domain is output through the inverse Laplace transform: ； wherein is the inverse Laplace transform; The Daubechies4 wavelet base function is selected, the wavelet base is suitable for processing the local characteristics of biological signals, the decomposition layer is set to 5 layers of decomposition, that is, the signal frequency band is divided into sub-frequency bands, and the signal-to-noise ratio of the reconstructed signal after threshold denoising is improved to above 42dB: ； wherein is the scale, is the position of the translation, is the set of integers, is the Daubechies 4 wavelet basis function, is the wavelet coefficient, representing the energy of the signal at that frequency band and time point; After the wavelet packet decomposition is completed, threshold denoising is performed, and the threshold is automatically confirmed through a heuristic threshold For each subband coefficient Soft threshold processing is performed to obtain the denoised wavelet coefficients The condition is: ； After threshold denoising, the significant signal features are finally reserved, and the random noise is suppressed. Then the denoised coefficients are used to reconstruct the signal With Daubechies4 wavelet basis function Linear combination is performed to reconstruct the signal : ； In the PVDF flexible sensor array unit, the frequency of the human body signal fluctuates over time. Short-time Fourier transform is used to generate a time-frequency spectrum matrix, and the numerical value is represented at time , the signal strength of the frequency is . A time-frequency heat map is generated, and each feature can form a highlighted area: ； wherein is the filtered PVDF signal time domain waveform, is the Hanning window function, is the complex exponential function used to calculate the frequency components, detecting the frequency signals present in the current segment; In the infrared temperature sensor, a temperature sensor is arranged on a control host, a sensor array is arranged at a 4*4 density, and a target temperature value is obtained by input value calculation , and a temperature measurement formula is ； wherein is the sensor raw reading, is the sensor material coefficient, is the skin emissivity of the measurement point, is the calibration offset; After the sensor output is subjected to CRC check, the temperature field is reconstructed by the heat conduction equation: ； In the formula, is the offset for heat dissipation from the pipe; In the environmental temperature and humidity sensor, the environmental parameters are collected, and the compensation coefficient calculation formula is: ； In the formula, is the ambient temperature compensation coefficient, is the ambient temperature.

4. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 3, wherein, In the data analysis module, the time-frequency domain feature matrix of the input PVDF sensor after preprocessing , i.e. 4 sensor channels, 256-dimensional features, the result of STFT, the input value needs to be checked, and the abnormal value is limited within : ； For each sensor channel, the data needs to be normalized for assessing the standard deviation distance of the sample point to the population mean : ； wherein is the average value, is the standard deviation; Feature stitching fusion and projection , after grouping the left and right partitions of the sensor, the features are longitudinally stitched and merged, is a learnable weight matrix, is a linear rectifier ReLU activation function, and finally the sensor input realizes linear projection dimension reduction from 256 dimensions to 128 dimensions, and the meaningless negative response is removed through the ReLU activation function; The final output result is: The sensor realizes 2 groups combined with 2 projection modes, and the fused feature dimension is 128; The human body detection branch adopts an improved ResNet18 architecture to finally output respiration, heart rate, and body movement sign classification results: ； wherein is a classification weight matrix, depending on the discrimination rule for non-body features, positive weights are assigned to high frequency noise, negative weights are assigned to feature features, is a classification bias vector; The input is the first 128 features of the output of the feature fusion layer: ； The first step is to calculate the residual block to learn the subtle difference pattern of human body and non-human body vibration: ； In the formula, is a residual function, two layers 1D convolution and ReLU, the first layer of convolution extracts local time-frequency patterns, and ReLU activation removes negative responses. The second layer of convolution refines the features, and the residual connection preserves the original features. The second step of the attention mechanism calculation automatically amplifies important frequency band signals: ； In the formula, The function is used to compress the weight to the interval (0, 1), For global average pooling processing; After processing, the features are enhanced: ； wherein is the attention-weighted feature matrix, is the Hadamard product, is the outer product operation, which extends the vector to the matrix; Finally, the classification probability is output: Global feature aggregation averages the features of the four sensors to obtain a 128-dimensional data feature vector representing the entire mattress range: ； Summation and averaging eliminate the influence of the number of sensors after spatial dimension aggregation; Finally, we get: ； Finally, use Softmax for normalization processing; Sleep posture recognition uses a spatio-temporal convolutional network. The input is the first 128 features of the output of the feature fusion layer. The input data is first spatially convolved: ； The convolution kernel covers 5 time points to capture local body pressure change features. The output is 64 feature channels to extract body pressure features of different parts, while spatial downsampling reduces computational complexity. LSTM is used to model the 64-dimensional spatial features at the corresponding time to obtain the spatio-temporal fusion features at the current time: ； In the formula, is a 64-dimensional space feature at the current time, is a hidden state at the previous time. Discrete body posture states are fused into time-continuous change states to suppress instantaneous misjudgments, and finally output 6-class posture probability distribution: ； In the formula, is a body pose classification weight, is a classification bias vector, and finally normalized using Softmax. The anomaly filtering module is implemented by a variational autoencoder, and the encoding process thereof first converts an input original time-frequency feature matrix The statistical quantity of the latent space is mapped through a three-layer fully connected network: ； ； wherein is the best compressed representation of the signal in latent space, describes the uncertainty in each dimension; The latent variable is sampled through the reparameterization trick: ； This operation allows gradient backpropagation while preserving the randomness, where each component is independently sampled from a standard normal distribution; Decoder will be reconstructed as with the training objective guided by the following loss function: ； wherein is the variational autoencoder loss function, is the weight coefficient of the KL term; The reconstruction loss forces the model to accurately recover the biometric component of the input signal, while the KL divergence term constrains the latent space distribution to be close to the standard normal distribution to avoid overfitting; When the reconstruction error exceeds a dynamic threshold, the current signal is determined to be interference and a filtering mechanism is triggered.

5. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 4, wherein, In the model prediction module, the temperature control model includes a hydraulic model and a thermodynamic model. The hydraulic model is used to output the optimal impeller pump speed regulation strategy, and the thermodynamic model is used to output the optimal heater power adjustment strategy. The above two strategies form the optimal temperature control strategy. In the model prediction module, high-precision sign monitoring is realized through multi-modal signal fusion. First, the PVDF sensor signal is preprocessed , wherein is the sampling point number; After eliminating high-frequency noise and baseline drift through 0.1-3Hz band-pass filtering, the Welch power spectrum estimation algorithm is used to calculate the signal frequency domain features: ； wherein is a frequency index, is a segment number, is a filtered PVDF signal time-domain waveform, is a Hanning window function, is a complex exponential function; By finding spectral peaks determining a heart rate, determining a respiration rate; The sleep stage classifier uses a three-channel LSTM network to process time series features: ； wherein is a normalized temperature feature, is heart rate variability, is a motion index; The output includes: 5-dimensional sleep stage probability distribution corresponding to wake, REM, N1-N3 stages; The metabolic rate calculation combines user physiological parameters and environmental data: ； wherein, is the body weight, is the height, is the age, is the 4-sensor temperature weighted average Values, = 0.3 for sleep stage coefficients, wake = 0, REM = 0.1, N3 = 0.3; This output forms a feedback loop with the temperature control, when it is detected that the temperature has dropped 15% and the energy saving mode is triggered to reduce the heating power.

6. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 5, wherein, In the model prediction module, the temperature control model composed of LSTM, GNN, and Transformer neural networks performs comprehensive analysis and decision-making to generate the optimal temperature control strategy. Its input feature vector is composed of three parts: physiological parameter vector , including real-time heart rate , respiratory rate and sleep posture encoding , temperature field data , corresponding to the temperature measurement value of 16 partitions of the mattress 16, environmental compensation vector , integrating environmental temperature, relative humidity and user activity intensity coefficient ; The three sub-feature spaces are first normalized to zero mean and unit variance by the feature scaling layer: ； where the cross-covariance matrix and the mean is updated by an online sliding window statistics. The normalized features are input into three neural networks in parallel: The LSTM time series network models the time-varying characteristics of the parameters through a gating mechanism , the hidden state of which captures the dynamic pattern of the last 2 minutes and outputs the control quantity , wherein is a body posture classification weight based on the LSTM time series network, is a classification bias vector based on the LSTM time series network. GNN spatial network construction temperature field graph structure , the features of the nodes include partition temperature and distance from the human core , the edge weight encodes the spatial correlation of heat conduction, wherein, is the normalized distance from the th partition to the human core, and the spatial control quantity is generated by two-layer graph convolution , wherein, is the degree matrix, is the adjacency matrix with self-loop, is the node feature matrix, is the trainable graph convolution weight matrix, is the original adjacency matrix, is the unit matrix; The transformer fusion network splices the above outputs as , calculates a cross-network feature correlation matrix through a multi-head attention mechanism, and finally outputs focusing on the key decision dimension, wherein is a query vector, is a key vector, is a scaling factor, is a value matrix; The arbitration uses a temperature-adaptive dynamic weighting strategy: ； Where the weight is: ； In the formula, is the temperature sensitivity coefficient; The mean square error of the last 5 decisions of each network is dynamically adjusted: ； wherein is the network At time the decision accuracy, is the time step index, is the current time, is the network predicted temperature, is the actually measured temperature; The arbitration output The target temperature and corresponding actuator control amount of 16 partitions form a closed loop with the instruction generation module of the back end.

7. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 6, wherein, In the instruction generation module, multi-step ahead optimal control is achieved by establishing a dynamic state space model of the mattress temperature field. The discrete state space equation is established based on the first law of thermodynamics: ； ； ； where state vector is the temperature value of 16 zones, control input is the water flow rate estimation value of 16 water paths and the corresponding heating sheet power command, is the water pump flow rate set value, is the temperature control power command, disturbance vector quantifies the influence of environmental temperature and humidity and user activity intensity, is the user real-time activity intensity coefficient, is the model uncertainty, is the temperature state transition matrix, is the control input matrix, is the disturbance input matrix; is a symmetric band matrix whose non-zero elements: ； wherein is the thermal conductivity of the material, is the density of the material, is the volume of the material, is the specific heat capacity, is the contact area of adjacent partitions, is the spacing; Control matrix Then, the fitting equation is obtained according to the fluid heat transfer experimental data: ； wherein is the first water flow at the point, is the maximum allowed temperature rise; At each control period, the following optimization problem is solved: ； The constraints are: ； ； ； ； wherein, is a prediction horizon, is a control horizon, is a weight matrix, to ensure control smoothness, is a penalty coefficient, is a slack variable, is a physical lower limit of the actuator, is a physical upper limit of the actuator, is a first temperature change amount of the zone, is a control period, is a personalized temperature comfort interval; The interface with the multi-network arbitration is reflected in the dynamic generation mechanism of the reference temperature ​ ； wherein, is a base temperature field, is a physiological compensation term, is a sleep position compensation matrix, is a real-time physiological vector, is a resting reference value, is a sleep position state encoding; The control output of the model predictive control is converted into PWM signals and heating power, and the signal in the state estimator is updated for the next period initialization, forming a complete closed-loop control with the back-end temperature control module.

8. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 7, wherein, In the temperature control module, the impeller pump adopts a dynamic speed regulation strategy to control the operating noise below 30 dB while ensuring the required flow rate. The heater is adjusted by a closed-loop PID to ensure water temperature control accuracy of ±0.2℃. The waterway pressure and water temperature data generated during the execution process are fed back to the temperature control module in real time for online updating of the hydraulic model and thermodynamic model, forming a complete control closed loop. The output optimization instructions are converted into actual device driving signals to achieve temperature field control through precise adjustment of the waterway system and heating elements. For PWM control of the impeller pump, first, the target rotation speed is calculated according to the flow demand target rotation speed: ； wherein is the pump characteristic coefficient, is the current differential line pressure, i.e. the pressure compensation term; The speed command is converted to PWM duty cycle through a PI controller: ； ； In the formula, is a rotational speed error, is a historical value of the rotational speed error, is a target rotational speed, is an actual rotational speed; The heater power control adopts a feedforward-feedback composite strategy: feedforward term directly compensating for environmental heat loss, wherein, is a target temperature, is an ambient temperature; feedback PID term eliminate steady state error, for temperature deviation, wherein, is the measured temperature of the current zone; To prevent overheating, the final power limit is: ； All commands are sent through internal bus communication at a frequency of 100 Hz, while receiving real-time feedback from sensors to form a closed-loop control. Real-time feedback optimization achieves continuous improvement of closed-loop control accuracy through a hydraulic-thermal coupling observer. The core is to dynamically correct the deviation between model prediction and actual execution. The observer based on the state space model works in a two-stage mechanism of temperature prediction and measurement correction, with the equation being: ； wherein is the estimated mattress temperature state vector, is the actual temperature sensor readings, is the observer gain matrix, which is dynamically updated by solving a Riccati equation on-line: ； wherein is the state transition matrix, is the transpose of the state transition matrix, is the solution matrix of the algebraic Riccati equation, is the measurement noise covariance matrix, is the transpose of the observation model matrix, is the observation model matrix, is the process noise covariance; The hydraulic control of the impeller pump first establishes a dynamic three-coupling model of flow rate-speed-pressure: ； Rotational speed The mapping relationship between flow rate and pressure is dynamically corrected by online identification of pipeline characteristic parameters , The pressure compensation term reflects the change of system impedance in real time; To suppress water hammer effects, a pressure differential feedback is used in combination with a second order Butterworth filter the control command is smoothed, for the target rotational speed, for the rate of change of the line pressure, is a transformation operator in discrete control; Motor drive end through back electromotive force compensation Eliminate electromagnetic interference; The heater control uses a multi-modal PID algorithm to achieve a balance between fast response and steady-state accuracy: ； User comfort learning automatically adjusts the regional temperature settings through optimization of the objective function: ； wherein Ttarget is a target temperature value, Tspeed is a temperature change rate, Tfoot is a measured temperature value of the foot region, Thead is a measured temperature value of the head region; The hydraulic model is updated every 30 minutes through recursive least squares, and the conduction coefficient, convection coefficient, and metabolic influence coefficient of the thermodynamic model are corrected based on finite element analysis. 9.The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 8, wherein, In the data security storage module, all sensor raw data is encrypted in real time through hardware-accelerated AES-128 algorithm during data acquisition: ； wherein is the AES-128-CTR encryption algorithm, is the sensor data, is the device unique key, is the counter mode, is the SHA-256 based keyed-hash message authentication code, is the device fingerprint generated by the physically unclonable function, is the time stamp accurate to milliseconds; User identity information is processed through double-layer hash desensitization: ； wherein, is a truncation function that retains the first 16 bits of the hash value, is a MAC address, is a random salt value.

10. The intelligent temperature control mattress system based on multi-modal sensing and adaptive learning of claim 9, wherein, In the model optimization module, the federated learning mechanism realizes cross-device collaborative optimization through triple privacy protection: First, the temperature field is compressed to 4-dimensional latent features using an autoencoder: ； In the formulae, is a temperature field encoder; its reconstruction loss preserving key heat distribution patterns: ； In the formula, a decoder for reconstructing a temperature field, is the temperature; Then, the physiological parameters are reduced to 2-dimensional principal components through PCA: ； wherein is a physiological parameter projection matrix, is a corresponding physiological parameter; Finally, the control instructions are SHA-256 hashed and desensitized: ； In the formula, is a hash de-sensitization operation; The uploaded gradient information is disturbed by Laplacian noise Satisfy the differential privacy requirement, control the risk of information leakage of privacy budget, and the central server aggregates the gradients of each device to generate a global model: ； In the formula, is the number of participating devices, is the model parameter of the device in the i-th round of training; and transfer global knowledge to local models through knowledge distillation its distillation loss balance model performance and complexity: ； wherein, is the divergence loss, is the output probability discrepancy distribution, is the local model trainable parameters.

Images