Non-inductive user health state monitoring interaction method based on smart home

Abstract

The invention discloses a non-inductive user health state monitoring interaction method based on smart home, and relates to the technical field of electrical digital data processing, and the method comprises the steps: deploying a non-contact sensor network composed of a millimeter wave radar, an infrared thermal imager, an acoustic array and a power consumption behavior monitoring module, synchronously acquiring micro motion, heat distribution, sound field and electric appliance load data; cross-modal data space-time alignment is realized through dynamic time warping and beacon auxiliary space mapping; fusing breathing, heart rate, voice emotion and power consumption behavior characteristics to generate a multi-dimensional physiological behavior vector; constructing an individualized dynamic health state model by using a long-short-term memory network with an attention mechanism, and outputting continuous health scores; and generating a self-adaptive home intervention strategy through fuzzy reasoning in combination with environment and time context. According to the method, high-precision, low-intrusion and highly-personalized health monitoring and intelligent interaction are realized, and the early warning accuracy and the user experience are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a non-intrusive user health status monitoring and interaction method based on smart homes. Background Technology

[0002] With the deep integration of IoT and AI technologies, smart home systems have evolved from single-device control to proactive environmental perception and user behavior understanding. Health monitoring, as a crucial component of smart living, is gradually shifting from a traditional wearable-device-dominated model to a non-intrusive, environmentally embedded perception paradigm. These systems utilize multimodal sensor networks deployed in living spaces to continuously collect user behavior trajectories, physiological signals, and environmental interaction data, enabling early identification and dynamic assessment of health risks without interfering with daily activities. The core of this technological approach lies in constructing an implicit, continuous, and highly accurate user state representation mechanism. Its effectiveness directly depends on the spatiotemporal consistency of sensor data, the ability to deeply analyze contextual semantics, and the efficiency of information fusion while protecting privacy.

[0003] Among them, the non-contact user health status monitoring and interaction method based on smart homes focuses on reconstructing users' vital signs and behavioral patterns using non-contact signals captured by built-in environmental sensors. This method aims to transform discrete physical quantity observations into a sequence of medically interpretable health indicators through cross-modal correlation modeling, and on this basis, establish an adaptive interaction strategy, enabling the home environment to respond to services based on the user's real-time physiological load and cognitive state. Its key technical challenge lies in how to achieve robust state inference and natural human-machine collaboration in home scenarios characterized by low signal-to-noise ratios, high noise interference, and significant individual differences.

[0004] Existing technologies typically rely on single-modal signals for health inference, such as pressure sensing from smart mattresses or visual analysis from cameras. This limits the monitoring dimensions and makes them susceptible to occlusion, lighting conditions, and user cooperation. Furthermore, the lack of dynamic calibration mechanisms for timestamp alignment and spatial coordinate unification of multi-source heterogeneous data leads to feature mismatch when users move or device topology changes. In addition, current systems generally use static thresholds or shallow classifiers for status determination, making it difficult to capture the non-linear physiological evolution patterns exhibited in the progression of chronic diseases or the precursors of acute events. More critically, the interaction logic between health status and home services is mostly driven by preset rules, failing to perform semantic-level intent understanding and response optimization based on the user's current cognitive load, emotional fluctuations, or mobility. This results in low adaptability and a high false alarm rate in real-world home environments. Summary of the Invention

[0005] The purpose of this invention is to provide a seamless user health status monitoring and interaction method based on smart homes, in order to solve the problems of low monitoring accuracy, high false alarm rate and poor system adaptability caused by insufficient fusion of multimodal sensor data, mismatch between time and space benchmarks, lack of dynamic adaptability in health status modeling, and interaction response mechanism deviating from the context of real cognitive load in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A seamless user health status monitoring and interaction method based on smart home technology includes the following specific steps: Step 1: Deploy a multimodal non-contact sensor network. Install millimeter-wave radar, infrared thermal imager, acoustic array sensor and power consumption behavior monitoring module in the target living space. Each sensor covers the user's main activity area in a distributed topology layout. Simultaneously collect human micro-motion signals, body surface temperature distribution, changes in ambient sound field and electrical load current waveforms to generate raw observation data streams with timestamps. Step 2: Perform cross-modal data preprocessing and spatiotemporal alignment. Denoise, normalize and extract features from the data output by each sensor. Use a sequence matching algorithm based on dynamic time warping to align the time axis of heterogeneous signals. And use an indoor beacon-assisted spatial coordinate mapping model to unify the observation results of different sensors to the global coordinate system to build a spatiotemporally consistent multidimensional sensing dataset. Step 3: Extract the joint feature vector of physiology and behavior, invert the respiratory rate and heart rate cycle based on the millimeter-wave radar echo signal, calculate the change rate of body surface thermal radiation gradient using infrared thermogram, combine the speech fundamental frequency drift and speech rate fluctuation characteristics of the acoustic array, and the start-stop timing anomalies in the use mode of electrical appliances to generate a multi-dimensional feature vector sequence containing vital sign parameters, emotional tendency index and daily behavioral regularity. Step 4: Construct an individualized dynamic health status model. Use a long short-term memory network to perform time-series modeling on the feature vector sequence. Introduce an attention mechanism to weight the time segments of key physiological events. Initialize the network weights through transfer learning and fine-tune them online based on the user's historical data. Output a continuous health status score trajectory. This score reflects the user's current physiological load level and potential health risk level. Step 5: Generate a context-aware adaptive interaction strategy. Taking the health status score as input, and combining it with ambient light, room temperature, time period and the user's recent interaction behavior logs, a service response decision is generated through a fuzzy inference system. The smart home execution unit is controlled to adjust the lighting brightness, air temperature and humidity, play soothing audio or push health reminders, so as to achieve non-invasive human-machine collaborative intervention that matches the user's current physical and mental state.

[0007] Preferably, in step 1, the millimeter-wave radar operates at a frequency of 60 GHz, has a transmission power of less than 10 mW, possesses Doppler effect resolution capability, can resolve micro-motion velocities of 0.1 mm / s on the body surface, has a spatial resolution of 0.2 m, an effective detection range of 8 m, a coverage angle of 120 degrees, and supports multi-target separation and trajectory tracking.

[0008] Preferably, in step 1, the spatial resolution of the infrared thermal imager is [missing information]. With a thermal sensitivity of less than 50mK and a sampling frequency of 25Hz, it is equipped with a wide-angle lens to cover the main activity areas of the bedroom and living room, and captures the dynamic changes in human body heat distribution in real time for analysis of body surface temperature trends and recognition of sleep posture.

[0009] Preferably, in step 1, the acoustic array consists of a ring array of 8 omnidirectional microphones with an element spacing of 4cm and a sampling rate of 48kHz. It has beamforming capability and is used to directionally collect user voice features and suppress background noise, and extract voice emotion-related parameters such as fundamental frequency variance, energy fluctuation rate and pause frequency.

[0010] Preferably, in step 1, the electricity behavior monitoring module is connected to the branch circuit of the household distribution box, with a sampling frequency of 1kHz, and records the voltage, current and active power of each electrical appliance. The module identifies individual appliance start-stop events and their power characteristic sequences through a non-intrusive load decomposition algorithm, which is used to infer the user's activity type and work-rest patterns.

[0011] Preferably, in step 2, the dynamic time warping algorithm sets the window width to 500ms and the penalty coefficient to 1.2 to align the time series of respiratory signals and heart rate signals, ensuring the synchronization of physiological parameters on a second-level scale; the spatial coordinate mapping model uses three preset wireless beacons as reference points, and uses the angle of arrival and signal strength fusion positioning to transform the target position detected by the millimeter-wave radar to a Cartesian coordinate system consistent with the infrared image.

[0012] Preferably, in step 3, the respiratory rate is determined by extracting the main frequency peak from the radar micro-Doppler spectrum through fast Fourier transform, and the heartbeat cycle is restored by phase demodulation combined with adaptive filtering, thereby improving the signal-to-noise ratio to over 20dB; the rate of change of body surface thermal radiation gradient is defined as the root mean square value of the pixel difference between consecutive frames of the heat map, which is used to quantify the fluctuation of metabolic activity.

[0013] Preferably, in step 3, the emotional tendency index is a weighted composite of the standard deviation of the fundamental frequency of speech, the rate of decline in speech speed, and the degree of disorder in the use of electrical appliances, with weighting coefficients of 0.4, 0.35, and 0.25, respectively. The index ranges from 0 to 1, and the higher the value, the more significant the degree of anxiety or fatigue.

[0014] Preferably, in step 4, the long short-term memory network contains two hidden layers, each with 128 neurons, the activation function is hyperbolic tangent, the time step is set to 60, and the input feature vector dimension is 16, including respiratory variability coefficient, low-frequency / high-frequency ratio of heart rate variability, slope of body surface temperature decrease, speech pause density, and frequency of electrical appliance switching, etc.; the attention mechanism adopts single-head scaled dot product attention, calculates the attention weights of each time step and sums them up to highlight key physiological turning points.

[0015] Preferably, in step 4, the transfer learning uses pre-trained model parameters from a public physiological database containing 500 healthy adults. After initializing the network weights, it is continuously fine-tuned on local user data with a learning rate of 0.001, and the model parameters are updated every 24 hours to ensure adaptability to individual differences.

[0016] Preferably, in step 5, the fuzzy inference system defines the health status score as divided into three fuzzy sets: a score of 0 to 0.3 is normal, a score of 0.3 to 0.7 is slightly abnormal, and a score of 0.7 to 1.0 is severely abnormal. Environmental factors are divided into light intensity, room temperature comfort, and time urgency. The output actions are lighting adjustment level, temperature control offset, and reminder priority. The rule base contains 27 "if-then" statements to achieve multi-condition coupled decision-making.

[0017] Preferably, in step 5, when the health status score is above 0.7 for 10 consecutive minutes and the voice emotion index is greater than 0.6, the system automatically dims the lights to 30%, starts the negative ion air purifier, lowers the air conditioner temperature by 1.5°C, and provides a voice prompt "It is recommended to pause operation and sit down to rest for 2 minutes" when the user enters the kitchen, with a response delay of less than 2 seconds.

[0018] Preferably, it also includes: establishing a local encrypted database to store anonymized health status score sequences, feature vector snapshots, and interaction logs, with a data retention period of 180 days, supporting doctor-authorized access and long-term trend analysis; the database uses the AES-256 encryption algorithm to protect privacy, and all data transmission is encrypted using the TLS 1.3 protocol.

[0019] Preferably, it also includes: introducing a feedback closed-loop mechanism, where users can confirm or deny the system's intervention behavior via voice or mobile terminal, the system records the feedback results and uses them to optimize the weights of fuzzy rules in the reinforcement learning module, and completes a strategy iteration every 7 days to improve the level of personalized service.

[0020] Preferably, after the method is deployed in a typical residential environment, the continuous monitoring time per day can reach more than 22 hours, the accuracy rate of health event early warning reaches 96.5%, the false alarm rate is less than 3.2%, the average response time is 1.8 seconds, and the user subjective satisfaction score is higher than 4.7 out of 5.

[0021] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention achieves non-intrusive, all-weather collection of users' vital signs and behavioral patterns by constructing a multimodal non-contact sensor network, overcoming the problems of poor wearability and limited single-modal perception dimension of traditional wearable devices.

[0022] This invention uses dynamic time warping and beacon-assisted coordinate mapping to solve the problem of temporal and spatial mismatch in multi-source heterogeneous data, thus ensuring the accuracy of feature fusion.

[0023] The present invention provides a personalized health status modeling method based on long short-term memory networks and attention mechanisms, which can capture the nonlinear evolution of physiological parameters and significantly improve the ability to identify chronic fatigue, sleep disorders and early cardiovascular abnormalities.

[0024] This invention combines a context-aware interaction strategy based on fuzzy reasoning systems to dynamically adjust home service responses according to the user's real-time physical and mental state, avoiding the disruptive operation of static rule-driven systems under high-load conditions.

[0025] The overall solution of this invention achieves high-precision, low-false-alarm, and highly adaptable health monitoring and intelligent intervention without infringing on privacy, which greatly enhances the practical value and user experience of smart homes in the field of health management. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the non-intrusive user health status monitoring and interaction method based on smart home proposed in this invention. Figure 2 This is a schematic diagram illustrating the core principle framework of the individualized dynamic health status model and context-aware adaptive interaction strategy in this invention. Detailed Implementation

[0027] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention. Example

[0028] Currently, with the deep integration of IoT and AI technologies, smart home systems are evolving from single-device control to proactive environmental perception and user behavior understanding. Health status monitoring, as a crucial component of smart living, urgently requires a non-intrusive, non-surgical, embedded environmental perception paradigm to overcome structural technical problems such as poor compliance of traditional wearable devices, limited single-modal perception dimensions, spatiotemporal mismatch of multi-source data, lack of dynamic adaptability in health modeling, and interactive responses detached from the context of real cognitive load. To address these technical issues, this invention proposes to achieve non-intrusive, all-weather data collection by constructing a multimodal non-contact sensor network. Dynamic time warping and beacon-assisted coordinate mapping ensure the accuracy of feature fusion. Long short-term memory networks and attention mechanisms capture the nonlinear evolution of physiological parameters, and a fuzzy inference system implements a context-aware adaptive interaction strategy. This achieves high-precision, low-false-alarm, and highly adaptable health monitoring and intelligent intervention without infringing on privacy, and is applied to a non-intrusive user health status monitoring and interaction method based on smart homes.

[0029] refer to Figure 1 The overall technical architecture of this invention includes a multimodal non-contact sensor network, a cross-modal data preprocessing and spatiotemporal alignment module, a physiological and behavioral joint feature extraction module, a personalized dynamic health status model, a context-aware adaptive interaction strategy generation module, a local encrypted database, and a feedback closed-loop mechanism. These modules work collaboratively to form a complete closed loop from raw signal acquisition to intelligent service response.

[0030] In the interactive method for non-contact user health status monitoring based on smart homes, step 1 involves deploying a multimodal non-contact sensor network. Millimeter-wave radar, infrared thermal imager, acoustic array sensors, and an electrical behavior monitoring module are installed within the target living space. Each sensor is arranged in a distributed topology to cover the user's main activity areas, simultaneously collecting human micro-motion signals, body surface temperature distribution, environmental sound field changes, and electrical load current waveforms, generating a raw observation data stream with timestamps. Specifically, for the millimeter-wave radar, see [link to relevant documentation]. Figure 1 The millimeter-wave radar operates at a frequency of 60 GHz, with a transmit power of less than 10 mW. It possesses Doppler effect resolution capability, can resolve surface micro-motion velocities as small as 0.1 mm / s, achieves a spatial resolution of 0.2 m, has an effective detection range of 8 m, a coverage angle of 120 degrees, and supports multi-target separation and trajectory tracking. (See also: Infrared thermal imager) Figure 1 The infrared thermal imager in the image has a spatial resolution of With a thermal sensitivity of less than 50mK and a sampling frequency of 25Hz, this acoustic array sensor is equipped with a wide-angle lens to cover the main activity areas of the bedroom and living room. It captures real-time dynamic changes in human body heat distribution for analysis of body surface temperature trends and sleep posture recognition. (See acoustic array sensor for details.) Figure 1The acoustic array sensor in the image consists of a circular array of eight omnidirectional microphones with an element spacing of 4 cm and a sampling rate of 48 kHz. It features beamforming capabilities for directional acquisition of user voice characteristics and suppression of background noise, extracting voice emotion-related parameters such as fundamental frequency variance, energy fluctuation rate, and pause frequency. See the electricity consumption behavior monitoring module for details. Figure 1 The electricity consumption behavior monitoring module is connected to the branch circuit of the household distribution box, with a sampling frequency of 1kHz. It records the voltage, current, and active power of each electrical appliance. Through a non-intrusive load decomposition algorithm, it identifies the start-stop events of individual appliances and their power characteristic sequences, which are used to infer the user's activity type and work-rest patterns. All sensors have built-in high-precision real-time clocks to ensure that the timestamp accuracy of the raw observation data stream is better than 1ms, providing a basic guarantee for subsequent cross-modal alignment.

[0031] In step 2 of the interactive method for non-intrusive user health status monitoring based on smart homes, cross-modal data preprocessing and spatiotemporal alignment are performed. The data output from each sensor are denoised, normalized, and feature extracted. A sequence matching algorithm based on dynamic time warping is used to align the time axes of heterogeneous signals. An indoor beacon-assisted spatial coordinate mapping model unifies the observation results from different sensors to a global coordinate system, constructing a spatiotemporally consistent multidimensional sensing dataset. Specifically, for the micro-Doppler time-frequency map output by millimeter-wave radar, wavelet threshold denoising is first applied to eliminate environmental clutter interference, followed by smoothing through a sliding window to improve signal continuity. After non-uniformity correction, the infrared thermal image sequence is subjected to bilateral filtering to suppress salt-and-pepper noise while preserving edge details. The original audio stream from the acoustic array is subjected to spectral subtraction to reduce steady-state background noise, followed by Mel-frequency cepstral coefficient transformation to extract robust acoustic features. The current waveform output by the electrical behavior monitoring module is bandpass filtered to remove high-frequency switching transients before being input into a non-intrusive load decomposition algorithm to identify electrical event boundaries. For time alignment, the dynamic time warping algorithm sets a window width of 500ms and a penalty coefficient of 1.2 to align the time series of respiratory and heart rate signals, ensuring the synchronization of physiological parameters on a second-scale. This algorithm uses dynamic programming to solve for the minimum cumulative distance path, mapping signals with different sampling rates to a unified time grid. Its computational complexity is O(n log n). Where N and M are the lengths of the two sequences, respectively. For spatial alignment, the spatial coordinate mapping model uses three pre-set wireless beacons as reference points. It achieves localization through angle of arrival (AHA) and signal strength fusion, transforming the target position detected by the millimeter-wave radar into a Cartesian coordinate system consistent with the infrared image. The beacons are deployed at the top corner of the room, emitting a 2.4GHz Bluetooth Low Energy signal. The receiver estimates the target's azimuth using phase interferometry and received signal strength indication, achieving a positioning error of less than 0.15m. After this processing, all sensor observations are unified into a global coordinate system with the room center as the origin, forming a spatiotemporally consistent multidimensional sensing dataset.

[0032] In step 3 of the interactive method for non-intrusive user health status monitoring based on smart homes, a joint feature vector of physiological and behavioral characteristics is extracted. Respiratory frequency and heart rate are inverted based on millimeter-wave radar echo signals. The rate of change of surface thermal radiation gradient is calculated using infrared thermograms. Combined with the speech fundamental frequency drift and speech rate fluctuation characteristics of the acoustic array, as well as the abnormal start-stop timing in appliance usage patterns, a multi-dimensional feature vector sequence containing vital sign parameters, emotional tendency index, and daily behavioral regularity is generated. Specifically, the respiratory frequency is determined by extracting the main frequency peak from the radar micro-Doppler spectrum using fast Fourier transform, achieving a spectral resolution of 0.05Hz, corresponding to a respiratory cycle measurement accuracy better than 0.2 breaths / minute. The heart rate is recovered using phase demodulation combined with adaptive filtering. Phase demodulation converts micro-displacement into phase change, and the adaptive filter uses the respiratory signal as a reference input, effectively suppressing respiratory harmonic interference and improving the signal-to-noise ratio of the heart rate signal to over 20dB. The rate of change of surface thermal radiation gradient is defined as the root mean square value of the pixel difference between consecutive frames of the thermogram, i.e.:

[0033] Where M and N are the image dimensions, The temperature value at position (i,j) in the heatmap of frame t is used to quantify fluctuations in metabolic activity. The emotional tendency index is a weighted composite of the fundamental frequency standard deviation of speech, the rate of speech rate decline, and the degree of disorder in appliance use, with weighting coefficients of 0.4, 0.35, and 0.25, respectively. The index ranges from 0 to 1, with higher values ​​indicating more significant anxiety or fatigue. Specifically, the fundamental frequency standard deviation reflects vocal tension, the rate of speech rate decline is calculated as the percentage reduction in syllables per unit time, and the degree of disorder in appliance use measures the temporal randomness of appliance start-up and shutdown events through information entropy. The final generated feature vector has a dimension of 16 and includes key indicators such as the respiratory variability coefficient, the low-frequency / high-frequency ratio of heart rate variability, the slope of body surface temperature decline, speech pause density, and appliance switching frequency, forming a temporal feature sequence updated every 60 seconds.

[0034] In step 4 of the smart home-based seamless user health status monitoring interaction method, an individualized dynamic health status model is constructed. A long short-term memory (LSTM) network is used to perform temporal modeling of the feature vector sequence. An attention mechanism is introduced to weight the time segments of key physiological events. Network weights are initialized through transfer learning and fine-tuned online based on historical user data, outputting a continuous health status score trajectory. This score reflects the user's current physiological load level and potential health risk level. Specifically, the LTM network contains two hidden layers, each with 128 neurons. The activation function is hyperbolic tangent, the time step is set to 60, and the input feature vector dimension is 16. The network structure includes a forget gate, an input gate, and an output gate. Its cell state update formula is:

[0035] in The outputs of the forget gate, input gate, and output gate are respectively... Represents element-wise product. The candidate cell states are represented. The attention mechanism employs single-head scaled dot-product attention, calculating the attention weights at each time step and summing them in a weighted manner to highlight key physiological turning points. Specifically, the query vector Q, key vector K, and value vector V are all obtained through linear transformation of the last hidden state of the LSTM layer, and the attention weights... The final context vector is Transfer learning was employed using pre-trained model parameters derived from a public physiological database containing 500 healthy adults. This database encompassed multimodal physiological data across various daily activities and mild stress scenarios. After initializing the network weights, continuous fine-tuning was performed on local user data at a learning rate of 0.001. The optimization objective was to minimize the mean squared error between the health status score and the physician-labeled physiological load level. The model parameters were updated every 24 hours to ensure adaptability to individual differences. The output health status score trajectory consisted of continuous values ​​between 0 and 1: 0 represented perfect health, 1 represented extremely high health risk, and intermediate values ​​reflected progressive physiological load accumulation.

[0036] In step 5 of the seamless user health status monitoring interaction method based on smart home, a context-aware adaptive interaction strategy is generated. Taking the health status score as input, and combining it with ambient light, room temperature, time period, and the user's recent interaction logs, a fuzzy inference system generates service response decisions. This controls the smart home execution unit to adjust lighting brightness, air temperature and humidity, play soothing audio, or push health reminders, achieving non-intrusive human-machine collaborative intervention that matches the user's current physical and mental state. The fuzzy inference system defines the health status score as divided into three fuzzy sets: a score of 0 to 0.3 is normal, a score of 0.3 to 0.7 is slightly abnormal, and a score of 0.7 to 1.0 is severely abnormal. Environmental factors are divided into: low, medium, and high light intensity; cold, moderate, and hot room temperature comfort levels; and leisure, normal, and urgent time urgency levels. Output actions include: 0% to 100% lighting adjustment level; -2°C to +2°C temperature control offset; and low, medium, and high reminder priorities. The rule base contains 27 "if-then" statements, such as "If the health status score is severely abnormal and the room temperature comfort level is hot, then the temperature control offset is -1.5°C." When the health status score is above 0.7 for 10 consecutive minutes and the voice emotion index is greater than 0.6, the system automatically dims the lights to 30%, starts the negative ion air purifier, lowers the air conditioner temperature by 1.5°C, and provides a voice prompt "It is recommended to pause operation and sit down to rest for 2 minutes" when the user enters the kitchen, with a response delay of less than 2 seconds. All control commands are sent to the corresponding execution units, such as smart lights, air conditioner controllers, and audio players, via the home LAN to ensure timely and coordinated service response.

[0037] To further enhance the system's usability and privacy security, this embodiment also includes establishing a local encrypted database to store anonymized health status score sequences, feature vector snapshots, and interaction logs. The data retention period is 180 days, and it supports doctor-authorized access and long-term trend analysis. The database uses the AES-256 encryption algorithm to protect privacy, and all data transmission is encrypted using the TLS 1.3 protocol.

[0038] In addition, it includes: introducing a feedback loop mechanism, where users can confirm or deny the system's intervention through voice or mobile devices. The system records the feedback results and uses them to optimize the weights of fuzzy rules in the reinforcement learning module, completing a strategy iteration every 7 days to improve the level of personalized service. The reinforcement learning adopts the Q-learning framework, with the state space being a combination of health status scores and environmental factors, the action space being the fuzzy rule weight adjustments, and the reward function being comprehensively set based on the positive or negative nature of user feedback and the subsequent improvement in physiological indicators.

[0039] After deploying this embodiment in a typical residential environment, the continuous monitoring time per day can reach more than 22 hours, the accuracy rate of health event early warning reaches 96.5%, the false alarm rate is less than 3.2%, the average response time is 1.8 seconds, and the user subjective satisfaction score is higher than 4.7 out of 5. The system demonstrates excellent performance in scenarios such as nighttime sleep monitoring, daytime workload assessment, and sudden discomfort early warning, effectively achieving non-intrusive, high-precision, and highly adaptable health status monitoring and intelligent interaction.

[0040] refer to Figure 2 The core principle framework of the individualized dynamic health status model and context-aware adaptive interaction strategy in this invention clearly demonstrates the complete data flow from multimodal feature input, LSTM temporal modeling, attention weighting, health score output to fuzzy inference decision-making. After the feature vector sequence is encoded by LSTM, the attention module dynamically focuses on the period of physiological abnormality, generating a highly discriminative context representation, which is then output as a health status score through a fully connected layer. This score, along with the environmental context, is input into the fuzzy inference engine, ultimately driving the smart home execution unit to generate an adaptive service response.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A seamless user health status monitoring and interaction method based on smart home technology, characterized in that, The specific steps include the following: Step 1: Deploy a multimodal non-contact sensor network to synchronously collect human micro-motion signals, body surface temperature distribution, environmental sound field changes, and electrical load current waveforms, generating a raw observation data stream with timestamps; Step 2: Perform cross-modal data preprocessing and spatiotemporal alignment, and perform denoising, normalization and feature extraction on the data output from each sensor; Step 3: Extract the joint feature vector of physiological and behavioral data to generate a multi-dimensional feature vector sequence containing vital sign parameters, emotional tendency index and daily behavioral regularity; Step 4: Construct an individualized dynamic health status model, initialize network weights through transfer learning, fine-tune them online based on user historical data, and output a continuous health status score trajectory. This score reflects the user's current physiological load level and potential health risk level. Step 5: Generate a context-aware adaptive interaction strategy. Taking the health status score as input, and combining it with ambient light, room temperature, time period and recent user interaction logs, a service response decision is generated through a fuzzy inference system. This decision controls the smart home execution unit to adjust lighting brightness, air temperature and humidity, play soothing audio, or push health reminders.

2. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: In step 1, millimeter-wave radar, infrared thermal imager, acoustic array sensors, and electricity consumption behavior monitoring module are installed in the target living space. The sensors are arranged in a distributed topology to cover the user's main activity areas. The millimeter-wave radar operates at a frequency of 60 GHz, has a transmit power of <10 mW, a spatial resolution of 0.2 m, an effective detection range of 8 m, and a coverage angle of 120 degrees. The infrared thermal imager has a spatial resolution of... The instrument has a thermal sensitivity of <50mK and a sampling frequency of 25Hz; the acoustic array consists of a ring array of 8 omnidirectional microphones with an element spacing of 4cm and a sampling rate of 48kHz; the power consumption behavior monitoring module has a sampling frequency of 1kHz, is connected to the branch circuit of the household distribution box, and records voltage, current and active power.

3. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: In step 2, a sequence matching algorithm based on dynamic time warping is used to align the time axes of heterogeneous signals, and an indoor beacon-assisted spatial coordinate mapping model is used to unify the observation results of different sensors to a global coordinate system, thus constructing a spatiotemporally consistent multidimensional sensing dataset. The dynamic time warping algorithm sets the window width to 500ms and the penalty coefficient to 1.

2. The spatial coordinate mapping model uses three wireless beacons as reference points and uses angle of arrival and signal strength fusion for positioning to transform the target position detected by millimeter-wave radar to a Cartesian coordinate system consistent with the infrared image.

4. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: In step 3, the joint physiological and behavioral feature vector includes: the rate of change of body surface thermal radiation gradient, respiratory rate, heart rate cycle, speech fundamental frequency drift and speech rate fluctuation characteristics, and abnormal start-stop timing in the use of electrical appliances; the respiratory rate is determined by extracting the main frequency peak from the millimeter-wave radar micro-Doppler spectrum through fast Fourier transform; the heart rate cycle is recovered by using phase demodulation combined with adaptive filtering to improve the signal-to-noise ratio to over 20dB; the rate of change of body surface thermal radiation gradient is defined as the root mean square value of the pixel difference between consecutive frames of the heat map.

5. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: The emotional tendency index is a weighted composite of the standard deviation of the speech fundamental frequency, the rate of decline in speech speed, and the degree of disorder in appliance use.

6. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: In step 4, a long short-term memory network is used to perform temporal modeling on the feature vector sequence, and an attention mechanism is introduced to weight the time segments of key physiological events. The long short-term memory network contains two hidden layers, each with 128 neurons, the activation function is hyperbolic tangent, the time step is set to 60, and the dimension of the input feature vector is 16. The attention mechanism adopts single-head scaling dot product attention, calculates the attention weights at each time step, and sums the hidden states by weight.

7. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: The transfer learning uses pre-trained model parameters from a public physiological database containing 500 healthy adults. After initializing the network weights, the model is continuously fine-tuned on local user data with a learning rate of 0.001, and the model parameters are updated every 24 hours.

8. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: The fuzzy inference system divides the health status score into three fuzzy sets: normal, mildly abnormal, and severely abnormal; environmental factors include light intensity, room temperature comfort, and time urgency; output actions include lighting adjustment level, temperature control offset, and reminder priority; the rule base contains 27 "if-then" statements.

9. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: When the health status score is above 0.7 for 10 consecutive minutes and the voice emotion index is greater than 0.6, the system automatically dims the lights to 30%, starts the negative ion air purifier, lowers the air conditioner temperature by 1.5°C, and provides a voice prompt "It is recommended to pause operation and sit down to rest for 2 minutes" when the user enters the kitchen, with a response delay of less than 2 seconds.

10. The non-intrusive user health status monitoring and interaction method based on smart home as described in claim 1, characterized in that: It also includes establishing a local encrypted database to store anonymized health status score sequences, feature vector snapshots, and interaction logs, with a data retention period of 180 days; the database uses the AES-256 encryption algorithm, and all data transmission is encrypted via the TLS 1.3 protocol; it also includes a feedback closed-loop mechanism, where users confirm or deny the system's intervention behavior via voice or mobile terminal, the system records the feedback results and uses them to optimize the weights of fuzzy rules in the reinforcement learning module, and completes a policy iteration every 7 days.

Images