User sleeping posture recognition system and method for intelligent pillow

Abstract

The invention relates to the technical field of intelligent health monitoring and signal processing, in particular to a user sleeping posture recognition system and method for an intelligent pillow, and the method comprises the steps: collecting the real-time pressure distribution data of a pressure sensor array arranged in the intelligent pillow; based on signal attenuation characteristics of the real-time pressure distribution data, current material physical state parameters are solved; in combination with the current material physical state parameter and the real-time pressure distribution data, calculating a material memory noise entropy representing a historical ghost superposition degree; comparing the material memory noise entropy with a preset interference threshold to generate a user sleeping posture state result; an identification mode is adaptively switched, and a fuzzy security boundary locking strategy based on sequential logic reasoning is executed; executing a static image accurate identification strategy; according to the method, catastrophic misjudgment of the neural network model during data ghosting is avoided, and denoising of a physical level and robustness enhancement of a logic level are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring and signal processing technology, specifically to a smart pillow system and method for recognizing user sleeping posture. Background Technology

[0002] In smart home health monitoring applications, smart pillows generally use flexible pressure sensor arrays embedded under viscoelastic filling materials such as memory foam to collect data on the pressure distribution on the user's head in a non-contact manner, and then achieve real-time monitoring of sleeping posture through data analysis. To identify a user's sleeping posture, existing technologies typically employ a classification architecture based on static image features. This involves directly treating real-time pressure data collected by sensors as a two-dimensional grayscale image and inputting it into a deep learning model such as a convolutional neural network to extract geometric shape features, thereby determining whether the user is lying on their back, side, or stomach. While this approach has a certain level of accuracy under ideal conditions where the user remains still for an extended period and the ambient temperature is constant, the significant physical hysteresis and temperature-sensitive characteristics of memory foam itself cause delays and nonlinear drift in the rebound response of the filling material when the user frequently turns over or when the ambient temperature fluctuates. This results in the pressure afterimages from past moments overlapping with the current real signal in a time series, creating a noisy image similar to double exposure. Existing algorithms lack the ability to dynamically perceive and decouple signals from the physical medium. When faced with high-noise afterimages or changes in material hardness, they often lead to serious misjudgments due to feature extraction distortion, failing to meet the need for low-latency, high-reliability early warning for high-risk postures such as prone positions. Therefore, how to effectively eliminate the hysteresis noise caused by material memory effect under non-ideal viscoelastic media and dynamic environmental interference, and improve the robustness and accuracy of sleeping posture recognition in complex dynamic scenarios, has become an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a smart pillow system and method for recognizing user sleeping postures, avoiding interference from pressure image afterimages caused by the rebound hysteresis of viscoelastic materials and the drift of medium physical properties caused by environmental temperature fluctuations. Furthermore, it can adaptively switch between fuzzy logic and precise recognition strategies based on material memory noise entropy, thereby achieving highly reliable sleeping posture monitoring and risk warning in complex dynamic scenarios. Specifically, the technical solution of this invention is as follows: Smart pillows use methods to recognize users' sleeping postures, including: Collect real-time pressure distribution data from a pressure sensor array installed inside the smart pillow; Based on the signal attenuation characteristics of the real-time pressure distribution data, the current material physical state parameters are calculated using a preset material signal dynamic compensation model. By combining the current material physical state parameters with the real-time pressure distribution data, the material memory noise entropy, which characterizes the degree of superposition of historical afterimages, is calculated. The material memory noise entropy is compared with a preset interference threshold, and the recognition mode is adaptively switched according to the comparison result to generate the user's sleeping posture status result. The adaptive switching recognition mode is configured as follows: when the material memory noise entropy is greater than the preset interference threshold, a fuzzy safety boundary locking strategy based on temporal logic reasoning is executed; when the material memory noise entropy is less than or equal to the preset interference threshold, a static image accurate recognition strategy is executed.

[0004] Preferably, the calculation of the current material physical state parameters includes extracting the attenuation rate of the pressure peak and the hysteresis time constant from the real-time pressure distribution data; The attenuation rate and the hysteresis time constant are input into the material signal dynamic compensation model to infer the current hardness and resilience of the smart pillow filling material. The material signal dynamic compensation model is pre-constructed based on the stress relaxation characteristics of viscoelastic materials at different temperatures.

[0005] Preferably, the calculation of material memory noise entropy includes estimating the theoretical residual amount of the pressure pattern at the current moment based on the current material physical state parameters; The energy proportion of the theoretical residual amount in the real-time pressure distribution data is calculated, and the energy proportion is defined as the material memory noise entropy. The material memory noise entropy is used to quantify the degree of signal superposition caused by material hysteresis in the current image.

[0006] Preferably, the fuzzy safety boundary locking strategy includes abandoning the classification and recognition of instantaneous static images, and instead extracting the pressure centroid change vector of the real-time pressure distribution data within a preset time window; Based on the pressure center of gravity change vector, a time-series motion trajectory is constructed, and it is determined whether the time-series motion trajectory falls into a preset non-prone safe area. If the patient falls into the non-prone safe sleeping area, the output will be locked to a safe sleeping position; otherwise, a high-risk warning will be triggered.

[0007] Preferably, the static image accurate recognition strategy includes inputting the real-time pressure distribution data into a preset convolutional neural network model, wherein the convolutional neural network model is configured to output specific supine, lateral, or prone classification results.

[0008] A smart pillow system for recognizing user sleeping postures, applied to the smart pillow method for recognizing user sleeping postures as described in any one of claims 1 to 5; and, A pressure acquisition device is used to monitor pressure changes on the surface of the smart pillow and output real-time pressure distribution data; The core processing unit establishes a communication connection with the pressure acquisition device to transmit real-time data, and includes: a signal compensation module that communicates with the pressure acquisition device; an entropy analysis module that receives the state parameters output by the signal compensation module; a strategy decision module that selects an identification path based on the output result of the entropy analysis module; and a state output module that generates the final result according to the selected identification path. The signal compensation module receives the real-time pressure distribution data, analyzes the signal attenuation characteristics, calculates the current material physical state parameters, and transmits them to the entropy analysis module. The entropy analysis module combines the pressure data to calculate the material memory noise entropy and sends the result to the strategy decision module. The strategy decision module compares the material memory noise entropy with a preset interference threshold. If it is greater than the preset interference threshold, the fuzzy safety boundary locking strategy is invoked. If it is less than or equal to the preset interference threshold, the static image accurate recognition strategy is invoked. The user's sleeping posture status result is output through the status output module.

[0009] Preferably, the signal compensation module includes a feature extraction unit and a state inference unit; The feature extraction unit is used to capture the nonlinear decay curve of the pressure signal, and the state inference unit has a built-in material viscoelastic mechanical model, which is used to invert the real-time hardness and resilience of the material based on the nonlinear decay curve without the assistance of a temperature sensor.

[0010] Preferably, the strategy decision-making module includes a temporal reasoning unit and an image classification unit; The temporal reasoning unit is configured to perform vector analysis in a material memory noise environment to eliminate high-risk poses, and the image classification unit is configured to perform high-precision pose classification in a low-noise environment.

[0011] Preferably, the status output module includes a risk assessment unit and an alarm execution unit; The risk assessment unit receives the user's sleeping posture status in real time. When the identification result is a prone position or a high-risk unknown state that lasts for more than a preset safety time limit, a trigger command is generated. The alarm execution unit responds to the trigger command by controlling external devices to perform a wake-up operation or send a remote alarm signal.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs an adaptive recognition architecture with self-awareness of the physical medium state. It quantifies the signal lag by calculating the material memory noise entropy. In harsh scenarios such as low temperature or rapid flipping that produces afterimages, it actively switches to temporal logic reasoning, effectively solving the problem of double exposure of images caused by material rebound lag. It avoids catastrophic misjudgment of neural network models when data ghosting occurs, and achieves denoising at the physical level and robustness enhancement at the logical level. 2. This invention analyzes the nonlinear attenuation characteristics of pressure signals and inversely infers the real-time hardness and resilience of materials without the need for temperature sensors. This mechanism eliminates additional hardware costs and the effects of heat conduction hysteresis, solves the problem of algorithm failure caused by the drift of pillow physical properties due to ambient temperature fluctuations, ensures that the system can still accurately obtain physical parameters when the hardness of the medium changes drastically, and guarantees all-weather recognition stability. 3. During periods of high noise where visual features fail, this invention utilizes a fuzzy safety boundary locking strategy to take over the safety defenses; by extracting the pressure center of gravity change vector and constructing a temporal motion trajectory, it determines in the logical space whether the user has fallen into a non-prone safety zone; this biomechanical-based reasoning mechanism ensures that the system will not miss dangers due to unclear images during moments of signal confusion such as when the user turns over, greatly reducing the risk of suffocation caused by prone positioning. 4. This invention achieves decoupling of computational paths and intelligent allocation of computing resources; it runs low-power time-series inference when there is high noise and performs high-precision image classification when the signal is clean; this dynamic switching mechanism avoids the ineffective waste of computing power, perfectly matches the embedded application scenario powered by batteries, and significantly improves the energy efficiency ratio and extends the battery life of the device while ensuring monitoring continuity and response speed. Attached Figure Description

[0013] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] Example 1: Please see Figure 1 The methods by which smart pillows recognize users' sleeping postures include: Real-time pressure distribution data is collected from a pressure sensor array installed inside the smart pillow; based on the signal attenuation characteristics of the real-time pressure distribution data, the current material physical state parameters are calculated using a preset material signal dynamic compensation model; By combining current material physical state parameters with real-time pressure distribution data, the material memory noise entropy, which characterizes the degree of superposition of historical afterimages, is calculated. The material memory noise entropy is compared with a preset interference threshold, and the recognition mode is adaptively switched according to the comparison result to generate the user's sleeping posture status result. Among them, the adaptive switching recognition mode is configured as follows: when the material memory noise entropy is greater than the preset interference threshold, a fuzzy safety boundary locking strategy based on temporal logic reasoning is executed. When the material memory noise entropy is less than or equal to the preset interference threshold, the static image accurate recognition strategy is executed.

[0016] This embodiment details the specific execution logic for achieving highly reliable sleeping posture recognition on a non-ideal viscoelastic medium, aiming to solve the problem of double exposure of pressure images caused by the rebound hysteresis of pillow material when users frequently turn over or when the ambient temperature changes. The system performs a data acquisition step by using a high-density flexible piezoresistive sensor array, such as a 64x64 grid, embedded beneath the smart pillow's filling material (hereinafter referred to as viscoelastic material), to capture pressure signals at a preset high-frequency sampling rate, such as 50Hz, generating time-series pressure image frames. The system performs a temperature-free physical state inversion. By analyzing the shape of the falling edge of the pressure signal, i.e. the speed at which the pressure disappears when the user leaves or moves, it inversely infers the current viscosity of the material and thus calculates the current physical state parameters of the material. Calculate the material memory noise entropy using a specialized algorithm combined with physical parameters. This parameter is used to quantify how much energy in the current pressure image is false afterimage signal that does not belong to the user entity at the current moment, and essentially reflects the signal-to-noise ratio of the physical medium; The system executes a decision based on the calculated entropy value using a bimodal cognitive architecture. If the material memory noise entropy is greater than a preset interference threshold (e.g., 0.6), the system determines that it is currently in a high-noise, high-hysteresis environment. At this point, the static image features are distorted, so a fuzzy safety boundary locking strategy is executed, abandoning precise classification and focusing on hazard logic reasoning. If the material memory noise entropy is less than or equal to the preset interference threshold, the system determines that the material is in good condition and the image is clear. Then, a precise static image recognition strategy is executed, and a high-precision model is called for fine-grained classification. The data is sourced from real-time data collected by a pressure sensor array. Its physical meaning is a two-dimensional pressure distribution matrix at the current time step, with units of N / cm². The source is calculated by the entropy analysis module, and its physical meaning is the normalized coefficient that characterizes the energy proportion of historical afterimages in the current image. It is dimensionless. The source is the system's preset storage; its physical meaning is the critical signal-to-noise ratio threshold that distinguishes whether image features are reliable; it is dimensionless. This embodiment introduces material memory noise entropy as a decision switch to construct an adaptive recognition architecture that is self-aware of the physical medium state. In harsh scenarios where low temperatures cause pillows to harden or users rapidly turn over, resulting in numerous afterimages, this method can proactively reduce dimensions and use temporal logic to safeguard safety boundaries, avoiding catastrophic false alarms from AI models in data riddled with ghosting. In particular, in preventing the risk of suffocation caused by prone sleeping, it achieves both physical-level denoising and logical-level robustness enhancement.

[0017] Example 2: The current solution of material physical state parameters includes extracting the decay rate of pressure peaks and hysteresis time constants from real-time pressure distribution data; By inputting the decay rate and hysteresis time constant into the material signal dynamic compensation model, the current hardness and resilience of the smart pillow filling material are inferred in reverse. The material signal dynamic compensation model is pre-constructed based on the stress relaxation characteristics of viscoelastic materials at different temperatures.

[0018] This embodiment is a further specification of the current material physical state parameter calculation steps in Embodiment 1; the process uses a nonlinear decay model to quantify the real-time relaxation rate of the material; The system monitors pixels in the pressure sensor array in real time where the values ​​drop sharply, that is, it captures the falling edge of the step response; Using the modified stress relaxation attenuation formula: ; in, To monitor pressure peak The relative time from which the calculation begins; By fitting data from multiple consecutive frames, the steady-state support pressure can be calculated in real time. Lag time constant And the first derivative of the pressure, i.e., the decay rate; The calculated eigenvalues ​​are input into the material signal dynamic compensation model. This model pre-stores the rheological spectrum of a specific type of polyurethane material, which covers the stress relaxation characteristics of viscoelastic materials at different temperatures, such as 5°C to 35°C, and at different aging degrees, including hardness values. With rebound rate value The solution is obtained using a pre-calibrated empirical regression equation, and the specific calculation model is as follows: Rebound rate calculation model: ; in, This represents the theoretical maximum resilience of the material. These are constants calibrated experimentally based on specific polyurethane materials; It has dimensions The coefficients are used to make the denominator dimensionless; Hardness calculation model: ; in, are the empirical regression fit coefficients, where Includes a conversion factor that maps the pressure unit N to the hardness unit; the system will use the conversion factor calculated in Example 2. By directly substituting into the above formula, the current hardness value of the smart pillow's filling material can be inferred without directly measuring the temperature. and rebound rate value In response to an increase in the hysteresis time constant, the system determines that the ambient temperature has decreased or the material has hardened, and thus adjusts the subsequent calculation parameters. The source is real-time sampling by the sensor, and its physical meaning is the instantaneous pressure value monitored at the current moment, with the unit being N; The source is historical cached data, and its physical meaning is the peak pressure before the action occurs, with the unit being N; The source is a fitting calculation, and its physical meaning is the steady-state support pressure value after the material relaxes, with the unit being N; The source is a fitting calculation, and its physical meaning is the hysteresis time constant, which characterizes the relaxation rate of the material under the current physical conditions, and the unit is s; The source is model inversion, and the physical meaning is the current hardness value of the material, with the unit being Shore A. This embodiment implements a seamless sensing mechanism; without increasing the hardware cost of the temperature sensor, it endows the algorithm with the ability to sense the state of the physical medium through signal processing; even in scenarios where the physical properties of the pillow drift from soft as dough to hard as a brick due to drastic fluctuations in bedroom ambient temperature, this solution can still accurately obtain material parameters, solve the problem of algorithm failure caused by the drift of medium properties, and ensure all-weather recognition stability.

[0019] Example 3: The calculation of material memory noise entropy includes estimating the theoretical residual amount of the pressure pattern at the current moment based on the current physical state parameters of the material; The energy proportion of the theoretical residual amount in the real-time pressure distribution data is calculated, and the energy proportion is defined as the material memory noise entropy. The material memory noise entropy is used to quantify the degree of signal superposition caused by material hysteresis in the current image. This embodiment is a micro-quantitative description of the material memory noise entropy calculation step in Embodiment 1; the step aims to transform the abstract signal hysteresis interference into a calculable numerical index. The system is based on the time lag constant obtained in the previous embodiment. Pressure image of the previous moment Physical simulations of attenuation are performed by applying attenuation functions based on material properties. This is specifically expressed as a discretized exponential decay model: ; in, For pixel coordinates in a two-dimensional image space, The sampling interval is... By directly reusing the real-time lag time constant calculated in Example 2, a virtual image that the sensor should theoretically read at this moment is generated, i.e., a ghost image. ; The system performs differential operations on the residual energy of the afterimage to calculate the energy present in the afterimage image. However, it is already in the current real-time pressure image. The hysteresis energy that disappears in the middle is defined as the pixel-level difference accumulation: ; The summation symbol This indicates traversing all pixels of the image; The lag energy Divide by the total energy of the current image The normalized material memory noise entropy is obtained. ,in To prevent the removal of minute quantities; In response to A value approaching 1 indicates that the current image energy contains a large amount of lingering historical artifacts, meaning the user has left but the pressure imprint remains, and the system determines the signal-to-noise ratio to be extremely low; in response to If the value approaches 0, it indicates that the afterimage basically overlaps with the current image or there is no afterimage, meaning that the current image is not significantly affected by historical lag energy, and the system determines that the image is a reliable pure pose. The source is a physical simulation calculation, and the physical meaning is the theoretical residual amount image, with the unit being N / cm². The source is a preset algorithm, and its physical meaning is the material decay function, which is dimensionless. The source is integral calculation, the physical meaning is material memory noise entropy, the value range is [0,1], and it is dimensionless; This embodiment constructs a digital twin-like afterimage, accurately distinguishing whether an image that appears to be lying on its side is actually lying on its side or an afterimage that has not yet disappeared. In complex dynamic scenarios where users continuously and rapidly turn over, this calculation mechanism provides a solid mathematical basis for subsequent strategy switching, effectively preventing posture misjudgment caused by signal superposition and improving the system's analytical depth of dynamic behavior.

[0020] Example 4: The fuzzy safety boundary locking strategy includes abandoning the classification and recognition of instantaneous static images, and instead extracting the pressure centroid change vector of real-time pressure distribution data within a preset time window; Construct a time-series motion trajectory based on the pressure center of gravity change vector, and determine whether the time-series motion trajectory falls into the preset non-prone safe area; If the sleeping position is not in a safe prone position, the output will be locked to a safe sleeping position; otherwise, a high-risk warning will be triggered.

[0021] This embodiment elaborates on the fuzzy security boundary locking strategy triggered when the system determines that it is in a high-noise, high-risk state; this strategy constitutes the core defense mechanism for dealing with the vast majority of high-risk steady states; In response to the system entering this mode, the algorithm no longer focuses on the shape features of static images, but instead extracts the center of gravity of pressure from real-time pressure distribution data within a preset time window. Movement trajectory, its coordinates Calculated based on the weighted first moment formula: ; In this context, the summation symbol ∑ represents traversing all rows and columns of the sensor array. The pressure value is the unit in the i-th row and j-th column of the pressure sensor array. For spatial coordinate index, The physical spacing between sensor units, in cm; based on this, the pressure center of gravity change vector is calculated. This filters out subtle noise and focuses on macroscopic mass transfer. The system constructs a time-series motion trajectory based on the pressure center of gravity change vector and compares the trajectory in the logical space to see if it falls into the preset non-prone safe area; the system presets a reference origin self-calibration step: the system maintains a long-term moving average center of gravity coordinate. ; If and only if the material memory noise entropy If the duration exceeds 5 seconds, the system determines that the user is in a standard supine static position, and at this time, it utilizes the current center of gravity. Update reference origin: ,in ; Based on this, it is specifically defined as taking the current reference origin. The polar coordinate constraint region is defined based on biomechanical characteristics, specifically as the polar coordinate constraint region with the initial supine center of gravity recorded when the system is powered on or when the user first lies down, as the origin. ; in, Polar radius, Polar angle, The preset safe displacement radius threshold, for example, is set to a value between 8cm and 15cm. For a safe angle range, for example, the value... Because the trajectory of the center of gravity shift from supine or lateral to prone has a specific direction and amplitude; based on this, dynamic transition logic based on safety boundary determination and high-risk locking is executed; In response to the trajectory falling into a non-prone safe area, even if the image is blurry, the system will still lock the output as a safe sleeping position to prevent false alarms due to poor visibility; in response to the trajectory characteristics not conforming to the safe area determination and the image continuing to be blurry, the system will immediately trigger a high-risk warning. The source is continuous frame differential calculation, and the physical meaning is the displacement vector of the pressure center of gravity per unit time, with the unit being cm / s; Non-prone safe zone: The source is a preset biomechanical model. Its physical meaning is a set of coordinates in the logical space that represent low-risk posture changes. It is dimensionless. In this embodiment, during the blind spot time when visual features fail, physical logic reasoning takes over the system's safety defenses. At the most dangerous moment when the user turns over, that is, when the pillow has not yet rebounded and caused sensor data to become chaotic, this strategy ensures that the system will not fail due to poor visibility, but will survive through behavioral logic, greatly reducing the false alarm rate under extreme conditions, and reflecting the algorithm design philosophy of prioritizing safety.

[0022] Example 5: The static image accurate recognition strategy includes inputting real-time pressure distribution data into a preset convolutional neural network model, which is configured to output specific supine, lateral, or prone classification results. This embodiment describes a static image accurate recognition strategy under normal conditions; this strategy is activated in a stable environment where the material memory noise entropy is below a threshold. Confirm that the signal-to-noise ratio of the real-time pressure distribution data collected by the current sensor meets the requirements; input the clear pressure topology map into the preset convolutional neural network model CNN; the model has been trained with a large amount of high-definition pressure data and has the ability to extract fine-grained features; The model outputs specific classification results for supine, lateral, or prone positions, and can further analyze specific head orientation or neck support. CNN model: Sourced from pre-trained storage, physically meaning a deep learning classifier, dimensionless; This embodiment ensures the system's high-precision recognition capability under normal conditions; by combining it with the aforementioned fuzzy locking strategy for both peacetime and wartime use, this solution can provide accurate posture analysis in scenarios where users are sleeping soundly, providing a high-quality data foundation for value-added services such as comfort assessment, and achieving a balance between safety assurance and precise service.

[0023] Example 6: Please see Figure 2 A smart pillow for recognizing user sleeping posture, which is applied to any of the smart pillow methods for recognizing user sleeping posture in Examples 1 to 5; and a pressure acquisition device for monitoring pressure changes on the surface of the smart pillow and outputting real-time pressure distribution data. The core processing unit establishes a communication connection with the pressure acquisition device to transmit real-time data, and includes: a signal compensation module that communicates with the pressure acquisition device. The entropy analysis module receives the state parameters output by the signal compensation module; the strategy decision module selects the recognition path based on the output results of the entropy analysis module. The status output module generates the final result based on the selected identification path; among them, the signal compensation module receives real-time pressure distribution data, analyzes the signal attenuation characteristics and calculates the current material physical state parameters, which are then transmitted to the entropy analysis module. The entropy analysis module combines the pressure data to calculate the material memory noise entropy and sends the result to the strategy decision module. The strategy decision module compares the material memory noise entropy with the preset interference threshold. If it is greater than the preset interference threshold, the fuzzy safety boundary locking strategy is invoked. If it is less than or equal to the preset interference threshold, the static image accurate recognition strategy is invoked. The user's sleeping posture status result is output through the status output module.

[0024] This embodiment provides a smart pillow system for recognizing user sleeping postures. This system serves as the hardware and architecture carrier for the above method, constructing a closed-loop adaptive processing environment. The system consists of a pressure acquisition device made of a multi-layer flexible circuit board and a pressure-sensitive ink coating, as well as a core processing device, such as a low-power MCU, which acts as the brain. In terms of function, the signal compensation module receives raw sensor data, performs physical noise reduction, and calculates material physical state parameters such as hardness and resilience. The entropy analysis module, acting as the system's confidence evaluator, calculates the material memory noise entropy by combining pressure data and material parameters, and sends this key indicator to the strategy decision-making module. The strategy decision-making module, acting as the commander, dynamically adjusts the data flow based on the entropy value. When the entropy value exceeds the preset interference threshold, the fuzzy security boundary locking strategy is invoked for logical reasoning. In response to an entropy value being less than or equal to the threshold, a static image accurate recognition strategy is invoked for deep learning classification. The status output module generates the final result based on the selected identification path and sends it through the wireless module; The architecture constructed in this embodiment no longer simply stacks algorithms, but allows hardware modules to be deeply coupled through entropy values. In scenarios where embedded devices have limited computing resources, the system can intelligently allocate computing resources. Logical reasoning has a small amount of computation while CNN has a large amount of computation. The dynamic allocation mechanism enables low-power chips to efficiently process complex nonlinear material data, significantly improving the energy efficiency ratio and response speed of the device.

[0025] Example 7: The signal compensation module includes a feature extraction unit and a state inference unit; The feature extraction unit is used to capture the nonlinear decay curve of the pressure signal, and the state inference unit has a built-in material viscoelastic mechanical model, which is used to invert the real-time hardness and resilience of the material based on the nonlinear decay curve without the assistance of a temperature sensor.

[0026] This embodiment details the signal compensation module in the system; the original design purpose of this module is to solve the dynamic drift problem of nonlinear material properties. The feature extraction unit is equipped with differentiating circuits or digital filters, specifically designed to capture the nonlinear decay curve of the pressure signal, with a focus on the tail features of the signal. The state inference unit calls the built-in material viscoelasticity model, preferably using the standard linear solid model or the generalized Maxwell model, to accurately describe the stress relaxation behavior of polymer materials, replacing the simple model that is only applicable to creep description; on this basis, without the assistance of temperature sensors, the unit directly inverts the real-time hardness and resilience of the material based on the curve features such as the time constant captured by the feature extraction unit. Standard linear solid model: Sourced from preset storage, its physical meaning is a viscoelastic constitutive model consisting of two springs and a damper, dimensionless; This embodiment significantly reduces hardware BOM costs, eliminates the thermistors and their wiring, and solves the problem of thermal conduction lag in temperature sensors. Given the physical contradiction of slow temperature conduction and instantaneous pressure feedback inside the pillow core, this solution achieves real-time response to environmental changes through signal inversion, ensuring the real-time performance and accuracy of compensation parameters.

[0027] Example 8: The strategy decision-making module includes a temporal reasoning unit and an image classification unit; The temporal inference unit is configured to perform vector analysis in a material memory noise environment to eliminate high-risk poses, and the image classification unit is configured to perform high-precision pose classification in a low-noise environment.

[0028] This embodiment refines the strategy decision-making module in the system and achieves decoupling of the calculation path; The temporal inference unit is configured to operate in a material memory noise interference environment, i.e. High-frequency activation involves performing vector analysis, calculating the displacement and velocity vectors of the pressure center of gravity, and employing an elimination logic to rule out high-risk postures. Simultaneously, or alternatively, the image classification unit is configured to operate in a low-noise environment. Low-frequency activation enables high-precision pose classification using feature matching logic; the system switches between the two units based on instructions from the preceding module. This embodiment significantly reduces the average power consumption of the system by decoupling computation; when the user frequently turns over and causes high noise, the system only runs low-computational-power time-series inference; only when the user is in a stable sleep and the noise is low does it run high-computational-power image classification; this mechanism is perfectly suited for battery-powered embedded scenarios, greatly extending the device's battery life while ensuring monitoring continuity.

[0029] Example 9: The status output module includes a risk assessment unit and an alarm execution unit; The risk assessment unit receives the user's sleeping posture status in real time. When the identification result is a prone position or a high-risk unknown state that lasts for more than the preset safety time limit, a trigger command is generated. The alarm execution unit responds to the trigger command by controlling external devices to perform a wake-up operation or send a remote alarm signal.

[0030] This embodiment defines the state output module in the system in a scenario-based manner; this module is responsible for converting the calculation results into physical intervention. The risk assessment unit maintains a risk score pool and receives the user's sleeping posture status results in real time. In response to the recognition result being a direct prone position, or a high-risk unknown state determined by a fuzzy strategy and lasting for more than a preset safety time limit, such as 60 seconds, the unit generates a trigger command. The alarm execution unit responds to the trigger command and controls external devices to perform wake-up operations, such as driving the miniature vibration motor in the pillow for physical intervention, or sending a remote alarm signal to the guardian's mobile client. Safety time limit: Source is user setting or system default, physical meaning is the maximum duration of time that a high-risk state is allowed, in seconds; This embodiment constructs a graded intervention mechanism; by combining the fuzzy locking strategy of the preceding steps, this module effectively filters out instantaneous false alarms caused by material memory effect, and only triggers an alarm when the risk is confirmed to persist; in infant or elderly care scenarios, this mechanism not only ensures timely wake-up when a real danger occurs, but also minimizes the alarm desensitization effect caused by false alarms, solving the pain point of users turning off the device due to frequent false alarms.

[0031] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for recognizing a user's sleeping posture by an intelligent pillow, characterized in that: The method comprises: collecting real-time pressure distribution data of a pressure sensor array arranged in the smart pillow; based on the signal attenuation characteristics of the real-time pressure distribution data, using a preset material signal dynamic compensation model to calculate the current material physical state parameter; combining the current material physical state parameter and the real-time pressure distribution data, calculating the material memory noise entropy representing the superposition degree of historical residual image; comparing the material memory noise entropy with a preset interference threshold, and adaptively switching the recognition mode according to the comparison result to generate a user sleeping posture state result; wherein the adaptive switching recognition mode is configured to: when the material memory noise entropy is greater than the preset interference threshold, execute a fuzzy safe boundary locking strategy based on time sequence logic reasoning; when the material memory noise entropy is less than or equal to the preset interference threshold, execute a static image accurate recognition strategy.

2. The smart pillow user sleeping posture recognition method according to claim 1, wherein: the calculation of the current material physical state parameter includes extracting the decay rate and the lag time constant of the pressure peak value in the real-time pressure distribution data; input the decay rate and the lag time constant into the material signal dynamic compensation model to inversely infer the current hardness value and the rebound rate value of the smart pillow filling material, and the material signal dynamic compensation model is pre-constructed based on the stress relaxation characteristics of viscoelastic materials at different temperatures.

3. The smart pillow user sleeping posture recognition method according to claim 1, wherein: the calculation of the material memory noise entropy includes, based on the current material physical state parameter, estimating the theoretical residual amount of the pressure pattern at the current time; calculate the energy proportion of the theoretical residual amount in the real-time pressure distribution data, and define the energy proportion as the material memory noise entropy, which is used to quantify the signal superposition degree caused by material hysteresis in the current image. 4.The smart pillow user sleeping posture recognition method of claim 1, characterized in that: The fuzzy safe boundary locking strategy includes abandoning the classification and recognition of instantaneous static image, and instead extracting the pressure gravity center change vector of the real-time pressure distribution data within a preset time window; based on the pressure gravity center change vector, construct a time sequence motion trajectory, and judge whether the time sequence motion trajectory falls into a preset non-prone safe region; if it falls into the non-prone safe region, lock the output as a safe sleeping posture state, otherwise trigger a high-risk warning.

5. The smart pillow user sleeping posture recognition method according to claim 1, wherein: the static image accurate recognition strategy includes inputting the real-time pressure distribution data into a preset convolutional neural network model, and the convolutional neural network model is configured to output specific supine, lateral or prone classification results.

6. A smart pillow system for recognizing user sleeping posture, characterized in that: The smart pillow user sleeping posture recognition method according to any one of claims 1-5 is used, and a pressure collecting device for monitoring the pressure change on the surface of the smart pillow and outputting real-time pressure distribution data. The core processing device is in communication connection with the pressure acquisition device to transmit real-time data, and comprises: a signal compensation module in communication connection with the pressure acquisition device; an entropy value analysis module receiving state parameters output by the signal compensation module; a strategy decision module selecting an identification path based on the output result of the entropy value analysis module; and a state output module generating a final result according to the selected identification path; The signal compensation module receives the real-time pressure distribution data, analyzes signal attenuation characteristics, and calculates current material physical state parameters and transmits them to the entropy value analysis module. The entropy value analysis module calculates material memory noise entropy in combination with pressure data and sends the result to the strategy decision module. The strategy decision module compares the material memory noise entropy with a preset interference threshold value. If the material memory noise entropy is greater than the preset interference threshold value, a fuzzy safe boundary locking strategy is called. If the material memory noise entropy is less than or equal to the preset interference threshold value, a static image accurate identification strategy is called. The state output module outputs a user sleeping posture state result.

7. The intelligent pillow user sleeping posture recognition system according to claim 6, characterized in that: The signal compensation module comprises a feature extraction unit and a state inference unit. The feature extraction unit is used to capture the nonlinear attenuation curve of the pressure signal, and the state inference unit is internally provided with a material viscoelastic mechanics model, which is used to inversely calculate the real-time hardness and resilience of the material according to the nonlinear attenuation curve without the assistance of a temperature sensor.

8. The intelligent pillow user sleeping posture recognition system according to claim 6, characterized in that: The strategy decision module comprises a time sequence reasoning unit and an image classification unit. The time sequence reasoning unit is configured to perform vector analysis to exclude high-risk posture risks in a material memory noise interference environment, and the image classification unit is configured to perform high-precision posture classification in a low-noise environment.

9. The intelligent pillow user sleeping posture recognition system according to claim 6, characterized in that: The state output module comprises a risk assessment unit and an alarm execution unit. The risk assessment unit receives a user sleeping posture state result in real time, and generates a trigger instruction when the identification result is a prone position or a high-risk unknown state duration exceeds a preset safety time limit. The alarm execution unit controls an external device to perform a wake-up operation or send a remote alarm signal in response to the trigger instruction.

Images