An AI-driven dynamic body temperature monitoring system

Abstract

The present invention discloses an AI-driven dynamic body temperature monitoring system, which aims to solve the problem that existing body temperature detection equipment is difficult to adapt to individual differences and dynamic changes in the environment. The system includes a temperature sensing unit, an environment acquisition unit and a heart rate sensing unit, which are respectively used to obtain skin temperature, environmental parameters and heart rate variability characteristics. The system constructs a body temperature and environment correlation matrix, extracts nonlinear coupling areas, and fuses heart rate features to generate a three-modal state tensor. Abnormal sensitivity is evaluated through an attention mechanism and a graph neural network model, and personalized fever judgment thresholds are dynamically generated in combination with historical stable states, thereby achieving accurate health response decisions. The system further outputs behavioral recommendations based on the decision results, such as hydration or medical treatment reminders, thereby improving the intelligence and practicality of body temperature monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of body temperature monitoring, and in particular to an AI-driven dynamic body temperature monitoring system. Background Art

[0002] Currently, temperature monitoring technology is widely used in public health management, home care, and clinical diagnosis. Common temperature patches, infrared thermometers, or wearable temperature measurement devices typically rely on a single temperature sensor for periodic measurements and use fixed temperature thresholds to determine whether an individual is at risk of fever. Some devices integrate mobile applications for data recording and remote viewing, improving the user experience, but overall, these devices still rely primarily on static data collection and rule-based triggering.

[0003] However, existing technologies have several shortcomings. First, static temperature thresholds fail to account for individual physiological differences and environmental changes, making them prone to false positives or missed alerts. Second, traditional devices often ignore auxiliary physiological signals related to body temperature, such as heart rate changes, resulting in a lack of precision and contextual adaptability in fever identification. Furthermore, data analysis is often based on simple threshold logic or regression models, which struggle to handle the complex coupling relationships in time series and provide early warning of potential health risks.

[0004] In view of the above problems, there is an urgent need to provide a new body temperature monitoring solution that can integrate multimodal perception data and has the ability to identify dynamic anomalies. Summary of the Invention

[0005] The present application provides an AI-driven dynamic body temperature monitoring system to improve the accuracy of body temperature monitoring.

[0006] This application provides an AI-driven dynamic body temperature monitoring system, including:

[0007] The temperature sensing unit is used to collect the original data stream of skin surface temperature at a preset sampling frequency and construct a basal body temperature sequence; the acquisition unit is used to obtain the ambient temperature, humidity and airflow disturbance amplitude of the area where the temperature sensing unit is located to construct a dynamic environmental state vector; the collaborative sensing unit is used to continuously detect heart rate variability and extract short-term heart rate imbalance characteristics; the abnormality recognition engine is used to construct a body temperature environment response association matrix for a local period based on the basal body temperature sequence and the environmental state vector, extract the nonlinear sensitive area of ​​body temperature changes to environmental parameters, and generate a temperature-loop coupling map; according to the short-term heart rate imbalance characteristics, the temperature-loop coupling map is dynamically labeled and embedded to form a three-modal state. state fusion tensor; based on a multi-scale discriminant model that integrates the attention mechanism and the graph neural network structure, calculate the abnormal sensitivity score of each state unit in the three-modal state fusion tensor, and output a set of candidate abnormal fragments; perform comparative matching based on the historical individual stable window on the candidate abnormal fragment set, adjust the upper and lower limits of the current judgment threshold, and generate a dynamic fever judgment threshold for the current individual and scene; compare the abnormal sensitivity score with the dynamic fever judgment threshold to form a decision output on whether to trigger a health response; a suggestion generation unit is configured to output a response suggestion with human behavior guidance characteristics based on a built-in nursing response rule library after detecting the decision output that triggers a health response.

[0008] The beneficial effects of this application mainly include: (1) By integrating multimodal physiological data such as body temperature, environmental parameters, and heart rate variability, the system can dynamically identify the true abnormal state of individual body temperature changes, significantly improving the accuracy and individual adaptability of fever detection, and reducing the risk of false positives and missed reports. (2) By constructing a body temperature environment response correlation matrix and extracting nonlinear sensitive areas, the causal relationship between body temperature fluctuations and environmental disturbances can be effectively distinguished, making the system more resistant to external interference and suitable for a variety of complex usage scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic diagram of an AI-driven dynamic body temperature monitoring system provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0010] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0011] The first embodiment of this application provides an AI-driven dynamic body temperature monitoring system. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1A first embodiment of the present application provides an AI-driven dynamic body temperature monitoring system.

[0012] The AI-driven dynamic body temperature monitoring system includes a temperature sensing unit 101, a collection unit 102, a collaborative sensing unit 103, an anomaly recognition engine 104 and a suggestion generation unit 105.

[0013] The temperature sensing unit 101 is used to collect the raw data stream of skin surface temperature at a preset sampling frequency and construct a basal body temperature sequence.

[0014] The temperature sensing unit 101 is a key component in this system for collecting body temperature data. Its primary function is to stably collect raw temperature data from the human skin surface at a preset sampling frequency while the wearer is in the wearable state. This data is then used to construct a basic body temperature sequence for subsequent abnormality identification and health status assessment. This unit preferably adopts a flexible, adhesive design that can adapt to dynamic changes caused by skin surface deformation, bending, or movement, ensuring continuous good contact between the sensor and the skin, thereby improving temperature measurement accuracy and data stability.

[0015] In practice, temperature sensing unit 101 should include a set of flexible, highly sensitive temperature sensors, such as polymer thermistors, microthermocouples, or MEMS thermal sensors. These sensors should have a thermal response time of no more than 2 seconds and a measurement accuracy preferably better than ±0.1°C. This unit, through embedded control logic, samples temperature data at fixed intervals (e.g., 1 to 5 times per second). Each sample is recorded with the current timestamp and temperature value, and cached to form a data stream with a time-series structure.

[0016] In order to improve the accuracy and anti-interference ability of data collection, the temperature sensing unit 101 is preferably equipped with a data preprocessing function to perform filtering and outlier removal operations locally or through the main control chip. The processing method may include sliding average filtering, median filtering, or a noise reduction algorithm based on empirical mode decomposition to eliminate outliers caused by temperature drift in the initial stage of wearing, electrical interference, or unstable skin contact. In addition, in order to compensate for the interference of environmental factors on the temperature measurement results, a reference channel can be set in the temperature sensing unit to monitor the near-surface ambient temperature, and the measured body temperature data can be dynamically corrected by difference modeling.

[0017] In terms of wearable structure, the temperature sensing unit 101 should preferably use a medical-grade skin-compatible adhesive material, such as a silicone gel layer or hypoallergenic pressure-sensitive adhesive film, to ensure that the sensor can be stably attached to the user's skin surface for a long time without causing irritation or shedding. A buffer layer can also be provided to reduce microscopic peeling caused by human movement and improve data continuity.

[0018] The collected data is organized within the system into a basal temperature sequence: a time-ordered collection of temperature data that reflects the user's skin temperature fluctuations over a specific period of time. Each data point contains at least the acquisition time, temperature value, and a data quality label. This temperature sequence is not only used for real-time display or recording but also provides the raw input for subsequently constructing a temperature environment response matrix, identifying abnormal temperature segments, and generating personalized fever thresholds.

[0019] To sum up, the temperature sensing unit 101 in this system undertakes the task of collecting body temperature data with high precision and high continuity, provides the main line data information required by the abnormality recognition engine, and ensures the reliability and physiological significance of the data through a series of sensing, attachment and preprocessing methods. It is the prerequisite and basic link for realizing dynamic body temperature intelligent monitoring.

[0020] The acquisition unit 102 is used to obtain the ambient temperature, humidity and airflow disturbance amplitude of the area where the temperature sensing unit is located, so as to construct a dynamic environment state vector.

[0021] In this system, acquisition unit 102 forms the environmental sensing pathway. Its core purpose is to perceive and quantify the external environmental conditions surrounding the temperature sensing unit in real time, including three key parameters: ambient temperature, relative humidity, and airflow disturbance amplitude. This information supports subsequent modeling of the dynamic relationship between body temperature changes and environmental factors. The environmental state information acquired by this unit is directly used to construct the environmental state vector, a crucial basis for the system's personalized temperature anomaly identification and false alarm suppression.

[0022] In terms of implementation structure, the acquisition unit 102 should include multiple sensing components that work together. A digital temperature sensor, such as DS18B20 or similar high-resolution thermistors, can be used to collect ambient temperature. It is placed at a certain distance from the skin to sense the local air temperature and avoid being directly affected by skin thermal radiation. Relative humidity can be measured by an integrated capacitive humidity sensor, such as the SHT series components, which are characterized by miniaturization, fast response and long-term stability. The airflow disturbance amplitude is usually obtained by a thermal anemometer or a miniature differential pressure sensor, which is used to reflect the air flow intensity near the patch area, thereby inferring ventilation, sweating or external disturbance events that may cause skin temperature fluctuations.

[0023] The "airflow disturbance amplitude" refers to the intensity of the change in airflow velocity per unit time within a defined area around the temperature sensor. It is used to indicate the intensity of air flow in that area. Its purpose is to reflect the dynamic interference of air flow on the temperature transfer process on the skin surface, thereby assisting in determining whether body temperature fluctuations are affected by external airflow, such as fan blowing, open window ventilation, or air disturbance caused by strenuous exercise.

[0024] In actual measurements, the amplitude of airflow disturbances can be determined using a miniature hot-film wind velocity sensor or a pressure-differential airflow monitoring element. Hot-film wind velocity sensors are preferred. They place a heating element in the fluid and use a temperature sensor to detect the speed at which the airflow removes heat, thereby inferring the current wind speed. The sensor should be mounted near the edge of the temperature patch structure so that it can sense air flow outside the non-contact layer without being directly affected by skin thermal radiation.

[0025] The system sets a continuous monitoring window, for example, 2 to 5 seconds, and records a sequence of wind speed changes within this time window. The wind speed amplitude within that period is calculated by calculating the difference between the maximum and minimum values ​​in the sequence. This is then normalized against the average wind speed value to define the airflow disturbance amplitude for that section. For example, if the wind speed changes from 0.3 m / s to 1.2 m / s within 5 seconds, the disturbance amplitude is 0.9 m / s. After normalization, this can be converted into the airflow disturbance amplitude defining that section for use by the downstream identification module.

[0026] To improve noise immunity, a bandpass filter algorithm can be introduced into the raw wind speed signal to remove low-frequency environmental drift and high-frequency electronic noise. A disturbance identification threshold can also be set, identifying a significant disturbance only when the disturbance amplitude exceeds a set minimum change value (e.g., 0.2 m / s). The disturbance amplitude is ultimately encoded into the environmental state vector as a floating-point number or discrete level (e.g., 0 = stationary, 1 = mild disturbance, 2 = moderate disturbance, 3 = severe disturbance), which is then used to model the nonlinear coupling relationship between body temperature data and environmental factors.

[0027] Therefore, the airflow disturbance amplitude is not the absolute value of a physical quantity. Instead, it is a dynamic disturbance indicator constructed by analyzing the amplitude of air velocity changes per unit time. It is used to express the instability of the local air environment of the temperature monitoring patch. Using this indicator, the system can effectively identify the interference of body temperature changes caused by wind speed fluctuations, thereby improving the accuracy and robustness of anomaly detection.

[0028] To ensure temporal consistency among the three types of data, each sensor in acquisition unit 102 should be driven and scheduled by a unified control logic. Its sampling frequency can be set to be consistent with or a multiple of that of temperature sensing unit 101, for example, every 5 seconds or every 10 seconds, depending on the system's requirements for timing synchronization accuracy. All sensor output data must be timestamped after acquisition and digitized for transmission to the system processing module.

[0029] At the data processing level, acquisition unit 102 needs to filter and validate the raw environmental data. For temperature and humidity data, sliding window averaging or exponential smoothing algorithms can be used to eliminate instantaneous fluctuations. For airflow data, a disturbance threshold can be set to record a valid disturbance event only when the airflow change amplitude exceeds the background noise range. The disturbance intensity level is then output to facilitate subsequent determination of whether the temperature change is due to environmental disturbances.

[0030] Furthermore, to adapt to different usage environments, the acquisition unit 102 can support an automatic calibration mechanism. For example, during the initial startup of the device, in a static state, a period of stable background data can be recorded as a reference baseline for the current environment, which is then used for normalization or offset correction of all subsequent changes. In this way, the system can achieve environmental perception adaptation in different scenarios and improve the generalization ability of anomaly recognition.

[0031] The constructed environmental state vector should contain three standardized fields: ambient temperature, relative humidity, and airflow disturbance index. This vector is aligned with the basal body temperature sequence according to the time window and passed as one of the input data to the anomaly recognition engine 104. It participates in key processes such as local response association modeling, sensitivity score generation, and dynamic threshold calculation.

[0032] In general, the acquisition unit 102 achieves background compensation and contextual interpretation of body temperature data under non-ideal conditions through a refined, multi-dimensional external environment perception design, effectively improving the system's recognition accuracy of fever status and robustness to environmental interference.

[0033] The collaborative sensing unit 103 is used to continuously detect heart rate variability and extract short-term heart rate imbalance features.

[0034] In this system, the collaborative sensing unit 103 collects and analyzes heart rate variation information related to body temperature fluctuations. Its core function is to continuously monitor the user's heart rate data and, based on changes in the ECG signal during the sampling period, calculate heart rate variability parameters, which reflect the stress state of the autonomic nervous system. It then extracts short-term heart rate imbalance characteristics from these parameters to help determine whether temperature changes are due to true physiological fluctuations or merely surface temperature interference caused by environmental factors. By incorporating heart rate information into the body temperature recognition process, the collaborative sensing unit 103 effectively improves the anomaly recognition engine's accuracy and individualized adaptability to heat stress events.

[0035] In terms of specific structure, this unit preferably uses a heart rate sensor based on photoplethysmography (PPG) technology. A miniature electrocardiogram (ECG) patch module or implantable physiological sensor can also be used. The PPG sensor is installed near the patch and uses a near-infrared light source to illuminate the subcutaneous capillaries. The heart rate is estimated by the periodic fluctuations in reflected light intensity that vary with the pulse. To ensure data continuity, the system should set a constant sampling frequency, preferably 25 to 100 times per second, and ensure stable data collection even when the user is moving.

[0036] Heart rate variability (HRV) refers to the fluctuations in the intervals between consecutive heartbeats, reflecting the balance between the sympathetic and parasympathetic nervous systems. It has been widely used to assess an individual's physiological stress level, sleep state, and abnormal regulatory responses. In this system, HRV is calculated based on the sequence of RR intervals between heartbeats (i.e., the time interval between two consecutive heartbeats). This is done by analyzing the statistical characteristics of this sequence within a short time window (e.g., 5 or 1 minute), including metrics such as the standard deviation (SDNN), the mean RR interval, and the low-frequency / high-frequency power ratio (LF / HF). These parameters are calculated using offline or online algorithms, and the output is updated after each sampling window.

[0037] "Short-term heart rate imbalance" refers to abnormal patterns of heart rate regulation occurring within a given timeframe, manifested as abnormally widened RR interval fluctuations, short-term high-frequency imbalances, or significant shifts in the LF / HF ratio. The system quantitatively assesses these abnormalities by setting thresholds and combining them with an individual's historical heart rate fluctuation baseline.

[0038] For example, if the LF / HF ratio exceeds three standard deviations of the individual's stable period for two consecutive time windows, the system can flag it as sympathetic nervous system overactivation, generating a "short-term heart rate imbalance" label. This label is output as a quantitative vector or a graded label, and combined with body temperature changes and environmental disturbance information within the same time period to form a trimodal fusion input for state modeling and anomaly scoring in the anomaly recognition engine.

[0039] During one monitoring period, the user's consecutive RR intervals were 860ms, 905ms, 780ms, 935ms, 810ms, and 960ms. The difference between any two intervals in the sequence exceeded 50ms, with no clear stabilization trend. This resulted in the standard deviation of the sequence (SDNN) exceeding the three-fold fluctuation threshold for the user's resting state. This significant increase in RR interval fluctuations typically reflects a disharmony in the sympathetic-parasympathetic regulation mechanism and can be considered a typical short-term heart rate imbalance.

[0040] For example, in another monitoring scenario, the system performed a frequency-domain transformation on the RR interval sequence within the sampling time window. The results showed that the power of the high-frequency band (HF, 0.15–0.4 Hz) dropped sharply to below 30% of the baseline mean over two consecutive windows, while the power of the low-frequency band (LF, 0.04–0.15 Hz) remained stable, resulting in insufficient HF components to maintain a parasympathetic-dominant state. This short-term high-frequency imbalance often reflects autonomic nervous system disturbances prior to acute stress, anxiety, or a rise in body temperature, and constitutes a type of "short-term heart rate imbalance signature."

[0041] Furthermore, the collaborative sensing unit 103 must be capable of filtering abnormal data and assessing the signal-to-noise ratio to ensure accurate judgment of the validity of heart rate data during conditions such as exercise, hand lifting, or loosening of the wearer's wrist, thereby preventing the input of spurious signals into the subsequent discrimination system. To this end, accelerometer information can be integrated into the PPG or ECG signal processing chain to assist in the detection of motion artifacts, or a threshold for continuous heart rate jumps can be set to screen out abnormal heartbeat bands without physiological basis.

[0042] Through the short-term heart rate imbalance characteristics provided by the collaborative sensing unit 103, the system can not only identify the synchronous rise in body temperature caused by sympathetic activation, but also issue sensitive warnings in the early state where the body temperature change is small but the heart rate regulation is unstable, thereby achieving early identification of fever risks and physical stress states, and forming a key auxiliary sensing path in the AI-driven dynamic body temperature monitoring system.

[0043] Furthermore, the collaborative sensing unit is specifically configured to:

[0044] An individual heart rate fluctuation baseline template is constructed to set an adaptive recognition threshold for short-term heart rate imbalance characteristics. The template is based on historical heart rate variability sequences obtained from the target individual at rest, during low-intensity activity, and during sleep. The template extracts the mean, standard deviation, and ratio of low-frequency and high-frequency components of the RR interval under each behavioral state to form a heart rate variability reference under multiple states.

[0045] The RR interval sequence collected in the real-time sliding window is input into the frequency domain feature extraction process to obtain the low-frequency power, high-frequency power and their ratio in the current time window. At the same time, the change slope of the RR interval sequence is calculated to obtain the real-time heart rate variability characteristics of the current time window.

[0046] Matching the real-time heart rate variability feature with the individual heart rate fluctuation baseline template state by state, and outputting a matching deviation degree to reflect the degree of deviation between the current state and the individual baseline state;

[0047] Obtaining the ambient temperature, humidity, and airflow disturbance amplitude of the current time window in the area where the temperature sensing unit is located, combining the matching deviation with the current environmental state vector, and generating a state adjustment factor as an input for dynamically adjusting the sensitivity and tolerance range of short-term heart rate imbalance feature recognition;

[0048] Under the control of the state adjustment factor, dynamic sensitivity judgment is performed on the matching deviation to obtain a short-term heart rate imbalance feature.

[0049] In this embodiment, the main task of the collaborative perception unit is to identify the characteristics of short-term heart rate imbalance and, by comparing it with the individual's baseline heart rate state, determine whether the current heart rate fluctuates abnormally, especially whether there are signs of sympathetic or parasympathetic nervous system imbalance, thereby providing key physiological perception input for the dynamic label embedding of the subsequent abnormality recognition engine. To achieve this goal, the system must first establish a heart rate fluctuation baseline template for the target individual. This template is not a static single-value reference, but a multidimensional statistical structure covering multiple daily behavioral states, and is constructed as follows:

[0050] During initial use or long-term wear of the system, the system automatically collects heart rate variability (HRV) data during rest, low-intensity activity, and nighttime sleep. Resting state can be determined by the motion sensor when the individual has not exercised significantly for more than 10 minutes; low-intensity activity corresponds to periods of walking or standing; and nighttime sleep can be determined by a combination of time, illumination, and limb immobility. For each of these states, the system extracts the corresponding RR interval sequence and calculates its mean, standard deviation, and the frequency domain power of low-frequency (LF), high-frequency (HF), and their ratio (LF / HF). For example, during rest, the system may collect a 60-minute RR interval sequence. After wavelet decomposition and FFT calculation, this sequence yields LF = 420ms² and HF = 510ms², resulting in a LF / HF ratio of 0.82. The mean RR interval for this period is also recorded as 790ms, with a standard deviation of 42ms. The system uses the similar statistical values ​​under the above three states as a reference interval or center value to construct a multi-state baseline template to facilitate subsequent matching and comparison.

[0051] During actual monitoring, the system uses a real-time sliding window mechanism to extract features from the current continuous RR interval sequence. The sliding window length can be set from 30 to 120 seconds, with each sliding interval being 10 seconds. The system first performs frequency domain conversion on the RR sequence within the window, extracting the current LF, HF, and their ratio, which reflect the spectral characteristics of the current autonomic nervous system regulation. Simultaneously, the system calculates the slope of the RR interval change within the window—the linear fit slope of the current sequence—to capture rapid increases or decreases in heart rate over short periods of time. For example, if the RR interval decreases from 780ms to 680ms within a given window, with a slope of -100ms / time unit, combined with a sudden increase in the LF / HF ratio (for example, from 0.8 to 1.5), this may indicate acute sympathetic activation and is a potential signal of a short-term heart rate imbalance.

[0052] The system compares the real-time heart rate variability features extracted in each time window with a previously constructed multi-state heart rate fluctuation baseline template. This comparison is not just a simple numerical comparison; rather, it calculates the degree of deviation of the current feature from the statistical interval within each state. For example, if the current LF / HF is 1.6, while it is 0.8±0.2 in the resting baseline template, the deviation is four times the standard deviation, and the system can determine that it is a serious deviation from the resting state. Similarly, if the standard deviation of the RR interval is significantly higher than the mean of the nighttime template, it can also indicate abnormal autonomic nervous system fluctuations. The system fuses all deviations using state weights and outputs a unified matching deviation value, which is used to represent the deviation amplitude of the current individual heart rate state relative to the baseline template.

[0053] Furthermore, to avoid misjudging normal physiological reactions caused by environmental changes, the system also collects environmental status information corresponding to the area where the temperature sensing unit is located, including ambient temperature, humidity, and airflow disturbance amplitude. This information is combined into a dynamic environmental state vector. Airflow disturbances can be provided in real time by a micro wind speed sensor, while humidity and temperature data are provided by the digital climate sensing component. The system combines the matching deviation and the current environmental state vector into a state adjustment factor calculation process to generate a state adjustment factor. This factor can be mathematically expressed as a weight or gain factor. Its function is to adjust the sensitivity and threshold range of subsequent imbalance feature recognition in real time.

[0054] This system uses a rule-based weight mapping approach to calculate the state adjustment factor. Specifically, the system first categorizes the current match deviation into several levels, such as "low deviation," "medium deviation," and "high deviation," based on the numerical range of the current match deviation, and assigns each a different basic scoring factor. Simultaneously, the system applies thresholds to various indicators in the environmental state vector—such as the magnitude of ambient temperature fluctuation, humidity fluctuation, and airflow disturbance intensity. If any indicator fluctuates significantly, a corresponding interference flag is set. Based on the number of interference flags and the combination of interference levels, the system assigns an environmental suppression weight to modify the aforementioned basic scoring factor. Ultimately, the state adjustment factor is derived by combining the basic scoring factor with the environmental suppression weight. Specifically, in situations where the match deviation is large but the environment is drastically changing, the adjustment factor is lowered to suppress false triggering. In contrast, in situations where the environment is stable but the deviation is high, the adjustment factor is increased to enhance the sensitivity of anomaly detection. For example, in situations with severe airflow disturbances or sudden temperature drops, even slight changes in heart rate may represent normal physiological reactions. The system will accordingly relax the "abnormal" criteria to avoid excessive triggering.

[0055] Ultimately, the system performs a dynamic sensitivity judgment on the matching deviation under the control of the state adjustment factor. This judgment process adjusts the originally statically set recognition threshold based on the adjustment factor and determines whether the recognition critical point is exceeded based on the current deviation. Once the deviation reaches or exceeds the threshold, the system marks the heart rate state corresponding to the time window as a short-term heart rate imbalance. This label not only serves as an important input for the abnormality recognition engine to embed dynamic labels in the temperature-environment coupling map, but also provides a key physiological reference for whether a health response is subsequently triggered.

[0056] After receiving the state adjustment factor and matching deviation, the system first adjusts the originally set recognition threshold range based on the adjustment factor. For example, when the adjustment factor is large, the system appropriately lowers the threshold to ensure sensitivity even to slight deviations. Conversely, when the adjustment factor is small, the system raises the threshold to avoid misidentifying anomalies under mild environmental interference. The system then compares the matching deviation in the current time window with the dynamically adjusted threshold. If the deviation exceeds the threshold, the system directly marks the time window as a short-term heart rate imbalance; if it does not exceed the threshold, it is determined to be a normal fluctuation. This determination result is attached to the current time segment as a label and pushed to the anomaly recognition engine in real time for dynamic label embedding in the subsequent temperature-environment coupling map. Through this process, the system implements a short-term heart rate anomaly detection mechanism that flexibly adapts to individual differences and environmental changes, effectively improving overall recognition accuracy while maintaining sensitivity.

[0057] For example, during a specific monitoring period, the system collected a 60-second series of RR intervals with a mean of 720ms, a standard deviation of 50ms, a LF ratio of 480ms², a HF ratio of 280ms², a LF / HF ratio of 1.71, and a slope of -60ms. Comparing this to the individual's daytime resting baseline template (LF / HF ratio of 0.85±0.15, standard deviation of 38ms, and a slope typically within ±15ms), the system determined that the current heart rate state deviated significantly. Simultaneously, the environmental conditions included a sudden drop in temperature of 4°C, a 15% increase in humidity, and a wind speed change rate of 0.3m / s. The state adjustment factor was assessed as 0.92, indicating that some of the deviation was attributable to environmental changes. Finally, even after adjustment, the dynamic threshold was still exceeded, confirming the presence of a short-term heart rate imbalance within this time window. The system then embedded this state into the temperature-environment coupling map for further processing by the anomaly recognition engine.

[0058] The abnormality recognition engine 104 is used to construct a body temperature environment response association matrix for a local time period based on the basal body temperature sequence and the environmental state vector, extract the nonlinear sensitive areas of body temperature changes to changes in environmental parameters, and generate a temperature-loop coupling map; dynamically embed labels into the temperature-loop coupling map based on the characteristics of short-term heart rate imbalance to form a three-modal state fusion tensor; calculate the abnormal sensitivity score of each state unit in the three-modal state fusion tensor based on a multi-scale discriminant model that integrates the attention mechanism and the graph neural network structure, and output a set of candidate abnormal fragments; perform comparative matching based on the historical individual stable window on the candidate abnormal fragment set, adjust the upper and lower limits of the current judgment threshold, and generate a dynamic fever judgment threshold for the current individual and scene; compare the abnormal sensitivity score with the dynamic fever judgment threshold to form a decision output on whether to trigger a health response.

[0059] The anomaly recognition engine 104 is the core component of the present invention's multimodal fusion intelligent temperature determination. It is responsible for modeling the complex interactions between body temperature changes, environmental disturbances, and heart rate responses from raw sensory data, and for identifying individual fever risks. To ensure high recognition accuracy and adaptability across diverse scenarios and individual conditions, the engine performs multi-stage operations in the following sequential flow. Each stage's input, processing mechanism, and output results have a clear data structure and repeatable implementation path, fully supporting the dynamic determination logic for abnormal body temperatures.

[0060] First, the abnormality recognition engine receives the basal body temperature sequence collected by the temperature sensing unit 101 and the environmental state vector obtained by the acquisition unit 102. The basal body temperature sequence is a one-dimensional time series structure, set as , where each represents the skin surface temperature value at the i-th sampling point with an accurate timestamp. The environmental state vector is a three-dimensional attribute set, set as , where e represents the ambient temperature, h represents the relative humidity, and f represents the airflow disturbance amplitude. These three correspond to each environmental sensing recording point aligned with time tᵢ in T. n represents the total number of data sampling points in the current time window, that is, the number of valid samples in the body temperature sequence and environmental state sequence obtained by the system from components such as the temperature sensing unit 101 and the acquisition unit 102 during a specific processing cycle.

[0061] Based on the above data, the system constructs the body temperature environment response correlation matrix of the local time period. This matrix is ​​recorded as , where each row corresponds to a sampling time point, and each column represents the synchronous observation value between body temperature, ambient temperature and airflow disturbance. To achieve nonlinear response modeling, the system performs normalization processing on each column in M ​​and extracts statistical features such as the response slope, residual sum of squares and offset direction of body temperature changes to ambient temperature changes within the window (e.g. 5 minutes) through a sliding window function to form a local response feature set. . All time windows By splicing, the temperature-ring coupling map can be generated.

[0062] The temperature-environment coupling graph is a structured temporal response graph, defined as G = (V, E), where nodes V represent response state units extracted within different time windows, each containing a local response feature vector. Edges E connect nodes that are temporally adjacent or have similar response patterns, and are accompanied by edge weights representing their coupling strength (e.g., cosine similarity or inverse Euclidean distance). The graph can naturally reveal whether a rapid temperature rise is associated with an abnormal environmental perturbation or is a typical physiological response, thereby supporting subsequent state fusion and scoring calculations.

[0063] Subsequently, the system introduces the short-term heart rate imbalance feature from the collaborative sensing unit 103 and dynamically embeds it into the temperature-loop coupling map. The embedding process is based on the node label mechanism: at each node In the , add the heart rate imbalance score corresponding to the time window As an additional feature dimension of the node, the graph structure is expanded into a three-modal state fusion tensor. This tensor is denoted as , where d is the fusion feature dimension of each node, including various attributes such as body temperature-environment response statistical characteristics, heart rate imbalance label, historical fluctuation amplitude, disturbance density, etc.

[0064] To achieve efficient and structure-aware anomaly detection, the system uses a multi-scale discriminant model that integrates an attention mechanism with a graph neural network (GNN) structure to process the tensor X. The model structure consists of three main modules: a graph embedding extraction module, an attention enhancement module, and a multi-scale classifier module.

[0065] The graph embedding extraction module first performs information propagation and convolution calculation on node X based on the graph structure G, and uses the standard graph convolutional network (GCN) method to extract the context-aware representation vector of each node. The input of each layer of embedding is the node vector of the previous layer, and the output is the updated node feature after the weighted neighbor information is aggregated. , where h is the embedding dimension.

[0066] The attention enhancement module introduces a multi-head attention mechanism based on the graph embedding Z to learn the response sensitivity weights to abnormal changes in different feature subspaces. Each head outputs a set of weighted node representations, and the results of all heads are connected to form an enhanced embedding. , where k is the concatenated dimension. The attention weight calculation process can adopt a query-key-value structure, where the query vector is the current node representation and the key value comes from the adjacent node features.

[0067] The multi-scale classifier module uses multiple sets of graph convolution kernels with different receptive fields (i.e., time window spans) to perform classification prediction operations on the enhanced embedding Z' and finally outputs the abnormal sensitivity score of each node. , where each Indicates the confidence level of the current time window being judged as a temperature abnormality segment. The system sets the candidate segment screening threshold θ, all The set of time segments ≥ θ is recorded as the candidate abnormal segment set.

[0068] For this set, the system further compares and matches it with the historical individual stable window. The stable window consists of a long-term accumulated non-abnormal period body temperature sequence. By statistically analyzing the mean value, standard deviation, fluctuation frequency, and rhythm factor of this sequence, a stable reference template W is formed. The system calculates the degree of difference between the candidate segment and W using Euclidean distance or correlation function, and adjusts the upper and lower limits of the dynamic fever determination threshold accordingly. Finally, the system compares the sensitivity score of each time segment with the current threshold. If the score exceeds the threshold, a decision signal is output to trigger a health response.

[0069] Overall, the abnormality recognition engine 104 takes time series body temperature data as the main line, combines it with the multi-participant heart rate status of the environment, and performs in-depth calculations and processing in multiple dimensions such as graph modeling, multimodal fusion, contextual attention, and dynamic threshold control. This enables the system to have high sensitivity and low false alarms to changes in individual health status, and is the core technical link for realizing AI-driven intelligent body temperature recognition.

[0070] The following is a reference implementation code of the anomaly identification engine 104:

[0071] import numpy as np

[0072] import networkx as nx

[0073] import torch

[0074] import torch.nn as nn

[0075] import torch.nn.functional as F

[0076] # Construct body temperature-environment response matrix

[0077] def build_response_matrix(temperature_seq, environment_seq):

[0078] n = len(temperature_seq)

[0079] M = np.zeros((n, 3))

[0080] for i in range(n):

[0081] M[i, 0] = temperature_seq[i]

[0082] M[i, 1] = environment_seq[i][0]

[0083] M[i, 2] = environment_seq[i][2]

[0084] return M

[0085] # Sliding window feature extraction

[0086] def extract_response_features(M, window_size):

[0087] features = []

[0088] for i in range(len(M) - window_size + 1):

[0089] window = M[i:i+window_size]

[0090] temp = window[:, 0]

[0091] env = window[:, 1]

[0092] coef = np.polyfit(env, temp, 1)

[0093] slope = coef[0]

[0094] fit = np.polyval(coef, env)

[0095] residual = np.sum((fit - temp) 2)

[0096] offset = np.mean(temp - env)

[0097] F = [slope, residual, offset]

[0098] features.append(F)

[0099] return features

[0100] # Build the temperature-environment coupling graph

[0101] def build_graph(features):

[0102] G = nx.Graph()

[0103] n = len(features)

[0104] for i in range(n):

[0105] G.add_node(i, feature=np.array(features[i]))

[0106] for i in range(n - 1):

[0107] for j in range(i + 1, n):

[0108] vec1 = G.nodes[i]['feature']

[0109] vec2 = G.nodes[j]['feature']

[0110] sim = np.dot(vec1, vec2) / (np.linalg.norm(vec1) np.linalg.norm(vec2))

[0111] if sim > 0.8:

[0112] G.add_edge(i, j, weight = sim)

[0113] return G

[0114] # Heart rate imbalance score embedding

[0115] def embed_hrv_labels(G, hrv_scores):

[0116] for i, r in enumerate(hrv_scores):

[0117] original = G.nodes[i]['feature']

[0118] extended = np.append(original, r)

[0119] G.nodes[i]['feature'] = extended

[0120] return G

[0121] # GCN graph convolutional layer

[0122] class GCNLayer(nn.Module):

[0123] def __init__(self, in_feats, out_feats):

[0124] super(GCNLayer, self).__init__()

[0125] self.linear = nn.Linear(in_feats, out_feats)

[0126] def forward(self, x, adj):

[0127] D_inv = torch.diag(1.0 / torch.sum(adj, dim=1))

[0128] Ax = torch.matmul(adj, x)

[0129] norm_Ax = torch.matmul(D_inv, Ax)

[0130] return F.relu(self.linear(norm_Ax))

[0131] # Multi - head attention mechanism

[0132] class MultiHeadAttention(nn.Module):

[0133] def __init__(self, dim_in, dim_out, num_heads):

[0134] super(MultiHeadAttention, self).__init__()

[0135] self.num_heads = num_heads

[0136] self.qkv = nn.Linear(dim_in, dim_out * 3)

[0137] self.out_proj = nn.Linear(dim_out, dim_out)

[0138] def forward(self, x):

[0139] B, D = x.shape

[0140] qkv = self.qkv(x).reshape(B, 3, self.num_heads, D / / self.num_heads)

[0141] q, k, v = qkv[0], qkv[1], qkv[2] s

[0142] score = torch.matmul(q, k.transpose(-2, -1)) / np.sqrt(D)

[0143] Note: In the original text, there is a syntax error in line 27 where the closing parenthesis of `nn.Linear(dim_in, dim_out` is missing. It should be `nn.Linear(dim_in, dim_out * 3)` as corrected in the translation.attn = torch.softmax(score, dim=-1)

[0144] out = torch.matmul(attn, v)

[0145] out = out.reshape(B, D)

[0146] return self.out_proj(out)

[0147] # Multi-scale classifier module

[0148] class MultiScaleClassifier(nn.Module):

[0149] def __init__(self, in_dim):

[0150] super(MultiScaleClassifier, self).__init__()

[0151] self.scale1 = nn.Linear(in_dim, 1)

[0152] self.scale2 = nn.Sequential(nn.Linear(in_dim, in_dim), nn.ReLU(), nn.Linear(in_dim, 1))<000,0329>

[0153] def forward(self, x):

[0154] out1 = torch.sigmoid(self.scale1(x))

[0155] out2 = torch.sigmoid(self.scale2(x))

[0156] return (out1 + out2) / 2.0

[0157] # Threshold adjustment function

[0158] def adjust_threshold(candidate_scores, stable_template):

[0159] base = np.mean(stable_template)

[0160] std = np.std(stable_template)

[0161] lower = base + 1.5 std

[0162] upper = base + 3.0 std

[0163] return lower, upper

[0164] # Trigger judgment

[0165] def make_decision(scores, threshold):

[0166] return [s>= threshold for s in scores]

[0167] # Main recognition process (calling path)

[0168] def anomaly_detection_pipeline(temperature_seq, environment_seq, hrv_scores, stable_template, window_size=10):

[0169] M = build_response_matrix(temperature_seq, environment_seq)

[0170] features = extract_response_features(M, window_size)

[0171] G = build_graph(features)

[0172] G = embed_hrv_labels(G, hrv_scores)

[0173] # Construct node feature matrix X and adjacency matrix Adj

[0174] X_np = np.array([G.nodes[i]['feature'] for i in G.nodes()])

[0175] X = torch.tensor(X_np, dtype=torch.float32)

[0176] A = nx.adjacency_matrix(G).toarray()

[0177] A = torch.tensor(A, dtype=torch.float32)

[0178] # Model construction

[0179] gcn1 = GCNLayer(X.shape[1], 32)

[0180] gcn2 = GCNLayer(32, 64)

[0181] attn = MultiHeadAttention(64, 64, num_heads=4)

[0182] clf = MultiScaleClassifier(64)

[0183] # Model forward inference

[0184] Z1 = gcn1(X, A)

[0185] Z2 = gcn2(Z1, A)

[0186] Z_attn = attn(Z2)

[0187] scores = clf(Z_attn).detach().numpy().flatten()

[0188] # Threshold and response determination

[0189] _, dynamic_threshold = adjust_threshold(scores, stable_template)

[0190] decisions = make_decision(scores, dynamic_threshold)

[0191] return scores, decisions

[0192] The anomaly detection engine's code simulates the core judgment process of an intelligent temperature monitoring system. It first continuously collects data such as body temperature, ambient temperature, and airflow changes from the user, organizing them into a matrix mapped over time. The system then uses a sliding window approach to extract the characteristics of body temperature responses to environmental changes during each time period. These time periods are then constructed into a graph structure, with each node representing the state of a time period and labelled with heart rate variability indicators. The system then uses a graph neural network to analyze the relationships between these nodes and, through an attention mechanism, enhances sensitivity to anomalies. Ultimately, it outputs a fever risk score for each time period. The system then automatically adjusts the judgment criteria based on the user's past temperature fluctuations when healthy and stable, determining whether the current score is abnormal and triggering a health alert. This entire process demonstrates how fever diagnosis can be determined by comprehensively considering body temperature, environmental factors, heart rate, and individual differences.

[0193] Furthermore, the anomaly identification engine is specifically used to:

[0194] The matching deviation corresponding to the short-term heart rate imbalance feature is used as the dynamic label value and embedded into each graph node in the temperature-environment coupling map. The graph node is composed of a time segment in the body temperature environment response association matrix of a local time period; the temperature-environment coupling map after embedding the dynamic label is adjusted for edge weights based on the slight change factor of the airflow disturbance amplitude between the graph nodes, and the connection relationship between the nodes is established to enhance the diffusion consistency of the dynamic label in areas with similar environmental disturbance characteristics; after completing the edge weight adjustment, continuous node groups are extracted according to the sliding time window, and the skin temperature change trend, environmental state vector and dynamic label value contained in each group of nodes are aggregated to generate a multimodal segment vector; all multimodal segment vectors are mapped into a tensor representation with a unified structural dimension to construct a three-modal state fusion tensor.

[0195] In this embodiment, the functional logic involved in the anomaly recognition engine has multi-level data organization and state fusion capabilities. The core goal is to achieve an organic combination of short-term heart rate imbalance characteristics and body temperature-environment response structure, and to construct a tensor expression structure with multi-modal linkage characteristics through time series sliding processing. In the specific implementation, the system first obtains the short-term heart rate imbalance characteristics identified by the previous module, and compares it with the individual heart rate fluctuation baseline template in combination with its corresponding time window, and calculates the matching deviation of the time period. The matching deviation reflects the degree of deviation between the current heart rate fluctuation and the historical baseline state, and therefore has a strong physiological abnormality indication significance. In order to accurately integrate this feature information into the time series structure, the system uses the matching deviation as a dynamic label value and embeds it into each graph node in the temperature-loop coupling graph.

[0196] During the graph construction process, the system constructs graph nodes based on the time segments in the body temperature-environment response association matrix. Each node corresponds to a specific time window, which includes the current skin temperature change rate, ambient temperature, humidity, and airflow disturbance amplitude. When constructing the graph edge structure, the system does not use a unified edge weight processing, but introduces the airflow disturbance amplitude micro-change factor as the basis for edge weight adjustment. Specifically, for any two adjacent or neighboring time window nodes, the system analyzes the airflow disturbance amplitude change. If the change trend is similar, that is, the disturbance micro-change factor is small, a higher edge weight is assigned to strengthen the connection density between such nodes, thereby enhancing the consistency of the propagation of dynamic label values ​​within similar environmental intervals, reflecting the continuity and correlation of body temperature changes in response to the environment.

[0197] Once the edge weight adjustment is completed, the system starts the sliding time window strategy, setting the time span to 5 seconds, 10 seconds or 15 seconds, and extracting continuous node groups in the entire graph structure. For example, for a 30-second period of collected data, setting the sliding window step size to 5 seconds can construct multiple overlapping continuous node groups. For each node group, the system extracts the multimodal elements of the nodes it contains, including the skin temperature change trend, the environmental state vector and the dynamic label value. The skin temperature change trend can be extracted by extracting its rising or falling direction and rate through the differential sequence of temperature values ​​within the node; the environmental state vector includes the original values ​​of temperature, humidity and disturbance amplitude and their rate of change; the dynamic label value is the aforementioned heart rate matching deviation. The three types of information together constitute the multimodal segment vector of the segment, reflecting the coupling relationship between physiology and the external environment in a specific time period.

[0198] After normalization and dimension alignment, all multimodal segment vectors are uniformly mapped into a tensor representation. This tensor is a trimodal state fusion tensor with fixed dimensions. The first dimension corresponds to the time window index, the second dimension represents the modality type (e.g., temperature trend, environmental state, heart rate label), and the third dimension represents the sub-feature dimension within each modality. In practice, this tensor can be implemented using frameworks such as NumPy and PyTorch as a three-dimensional array structure, such as Tensor[time index][modality index][feature dimension], where each tensor unit corresponds to a set of linked states for a specific time segment. This tensor not only provides a unified input interface for subsequent anomaly sensitivity scoring but also lays the structural foundation for cross-temporal and cross-modal contextual discrimination.

[0199] For example, in a specific application scenario, the system continuously collects 30 seconds of skin temperature, environmental data, and heart rate data with every 5 seconds as a time segment. It is divided into 6 nodes in total, and the body temperature environment response correlation matrix is ​​constructed in turn. The sensitivity between the skin temperature change rate and the environmental parameter change rate is used as the graph node element; the heart rate deviation is calculated and embedded as the dynamic label value of each node; by comparing the small changes in the airflow disturbance amplitude between adjacent nodes, the edge weight is adjusted to form a graph structure reflecting the continuity of the disturbance; then, 3 consecutive node segments are extracted by sliding, and their body temperature trends, environmental states, and dynamic labels are merged to generate three sets of multimodal segment vectors; finally, they are uniformly encoded into a three-modal state fusion tensor in the form of Tensor[3][3][n] to support the subsequent spatiotemporal anomaly recognition process.

[0200] The above process not only ensures the close coupling between different modal data, but also utilizes the edge weight mechanism and dynamic label fusion mechanism of the graph structure to establish a correlated propagation structure between abnormal heart rate and body temperature environment response.

[0201] Furthermore, the anomaly identification engine is specifically used to:

[0202] Extract the initial embedding representation of each state unit in the trimodal state fusion tensor, the initial embedding representation includes the dynamic label value, the skin temperature change trend and the environmental state vector, and performs position encoding according to the connection relationship between nodes in the graph structure; based on the initial embedding representation and the node position encoding, the importance weights of the three types of modal data in each state unit are calculated through the fusion attention mechanism to form a multimodal fusion attention matrix, which is used to enhance the expression contribution of different modal data in the temporal adjacent state; the multimodal fusion attention matrix and the initial embedding representation of each state unit in the trimodal state fusion tensor are jointly input into the graph neural network structure, and a cross-node feature aggregation operation is performed to obtain the aggregated feature vector of each state unit under its adjacent structure; based on the aggregated feature vector, its cross-scale neighborhood distribution characteristics are extracted through graph convolution kernels of different scales, and the convolution output results of each scale are fused to calculate the abnormal sensitivity score of each state unit in the trimodal state fusion tensor.

[0203] In this embodiment, the anomaly recognition engine is used to perform high-dimensional expression and intelligent identification of each state unit in the trimodal state fusion tensor. The entire processing process runs through multiple interconnected steps such as data extraction, attention weighting, graph structure information propagation and multi-scale discrimination.

[0204] First, the system extracts the initial embedding representation of each state unit from the constructed trimodal state fusion tensor. In this embodiment, a state unit represents the coupling of the physiological state and its external environmental state within a fixed time window. It is specifically composed of three parts: the first is a dynamic label value, which is derived from the short-term heart rate imbalance feature analysis process and reflects the degree of deviation of the current heart rate from the individual baseline; the second is the skin temperature change trend, which is modeled as a sequence of temperature differences over several time slices in the past, reflecting local temperature fluctuations; and the third is the environmental state vector, which contains key parameters such as ambient temperature, humidity, and airflow disturbances within the current time window. These three modal data are encoded as vectors of uniform length and positionally encoded according to the connection relationship between graph nodes in the temperature-environment coupling graph. This ensures that each state unit contains contextual information of the spatial adjacency structure, laying the foundation for subsequent graph structure propagation operations.

[0205] After completing the initial embedding representation and positional encoding construction, the system enters the fused attention mechanism phase. During this phase, the system comprehensively considers the importance of the three modalities to the anomaly recognition task. To this end, the system designs a multi-head attention calculation process, calculating the attention score for the dynamic label value, temperature trend, and environmental state vector within each state unit. For example, if skin temperature fluctuates significantly while heart rate fluctuates slightly within a specific time window, the system will increase the attention weight of the temperature modality and reduce the influence of the dynamic label modality. This weight is calculated independently across multiple channels using a dot-product attention mechanism, and ultimately outputs a three-dimensional attention matrix as a weighted sum, known as the multimodal fused attention matrix. This matrix not only reflects the differences in importance between modalities but also enhances the consistency of representation between adjacent states in different time periods. It is particularly suitable for processing vital sign data that coexists with high-frequency fluctuations and low-frequency trends.

[0206] After the fused attention matrix is ​​generated, it is fed into the constructed graph neural network structure along with the initial embedding representation. In this example, the graph neural network employs a propagation mechanism based on adjacency weight control. This means that each state unit not only uses its own features in computation but also gathers information from its neighboring nodes, forming a cross-node feature aggregation. The system performs one or more layers of graph convolution operations within the graph structure, integrating the features of neighboring nodes into the current node representation via a weighted summation, thereby obtaining a new aggregated feature vector. This vector reflects the contextual relevance of the state unit within the entire graph, helping to identify persistent physiological abnormalities under a specific perturbation propagation path. For example, when certain node clusters exhibit both high heart rate deviations and high airflow fluctuations, and their neighboring nodes in the graph also exhibit abnormal temperature trends, this joint distribution feature is retained in the aggregate vector, providing a basis for subsequent anomaly detection.

[0207] After completing the calculation of the aggregated feature vector, the system will send it to the multi-scale discriminant module to extract deeper pattern information. This module performs convolution operations on the aggregated feature vector by setting multiple graph convolution kernels of different sizes, such as 3-node, 5-node, and 7-node neighborhood scales. Each convolution kernel extracts local pattern information within its corresponding adjacent range and then outputs the convolution results separately. These results represent the neighborhood structure response at different granularities. The system then splices or weightedly fuses these multi-scale convolution outputs to obtain a final discriminant vector. This vector is the abnormal sensitivity score of each state unit after fusing the trimodal information, structural adjacency information, and attention weights, which is used to measure whether the current state is at abnormal physiological risk.

[0208] For example, in a continuous monitoring sequence, an individual's body temperature rises rapidly in a high-temperature environment, and heart rate fluctuations show an abnormal increase during the nighttime rest period. Airflow disturbance signals also continue to increase during this period. At this time, the temperature mode and heart rate mode identified by the attention mechanism have high weights. After propagation through the graph neural network, the adjacent nodes aggregated all show similar temperature rise trends and heart rate imbalance signals. Multi-scale convolution further reveals that this abnormal feature is stable at multiple time granularities. Ultimately, the system outputs a high abnormal sensitivity score for this state unit.

[0209] In summary, the present invention establishes an accurate and highly robust three-modal fusion anomaly recognition path through steps such as multi-layer embedding extraction, attention fusion, graph structure aggregation and multi-scale discrimination.

[0210] Furthermore, the anomaly identification engine is specifically used to:

[0211] Based on the aggregated feature vectors of each state unit in the candidate abnormal segment set in the trimodal state fusion tensor, the joint fluctuation amplitude including the skin temperature change trend and the environmental state vector is calculated to characterize the comprehensive change level of the current temperature-environment response; the joint fluctuation amplitude is used as input and the overlapping interval is matched with the historical individual stable window stored in the system segment by segment to extract the historical reference temperature sequence and corresponding heart rate deviation of each segment under the same behavioral state and similar environmental disturbance; for each state unit in the current candidate abnormal segment set, the deviation score is calculated based on the difference between it and the historical individual stable window to obtain the temperature deviation confidence index of the state unit to reflect the relative abnormality degree in the current state; the temperature deviation confidence indexes of all state units are aggregated to construct a sensitivity distribution curve of the current candidate abnormal segment set, and the sensitivity distribution curve is compared with the threshold distribution in the historical individual stable window to extract the sensitivity deviation interval; based on the sensitivity deviation interval, the upper and lower limits of the current judgment threshold are dynamically adjusted to generate a dynamic fever judgment threshold for the current individual and scene.

[0212] In the AI-driven dynamic body temperature monitoring system of the present invention, the anomaly recognition engine not only has the traditional static rule recognition function, but also introduces a highly personalized and dynamically adaptable judgment mechanism. In particular, in the process of generating the dynamic fever judgment threshold, a fusion calculation is performed based on multi-source state information and historical individual stable data to ensure that the system has accurate and stable response capabilities in different individuals and environmental scenarios.

[0213] First, the system extracts the portion of the trimodal state fusion tensor obtained from the preceding module that has been labeled as a set of candidate abnormal segments. Each candidate abnormal segment is composed of several state units, each of which contains a feature vector aggregated within a time window. These vectors combine key physiological and environmental variables such as skin temperature trends, environmental state vectors, and dynamic label values. The system uses these aggregated feature vectors to calculate the joint fluctuation amplitude, which is used to comprehensively assess the response strength between skin temperature changes and environmental perturbations during the current time period. For example, if the slope of skin temperature rise in a state unit increases significantly while environmental perturbations (such as airflow rate and light intensity) remain stable, it indicates that the individual has a body temperature response that is not synchronized with the external environment. This will result in a large joint fluctuation amplitude, which serves as a significant signal of potential abnormality.

[0214] Next, the system uses this joint fluctuation amplitude as a retrieval condition and matches it segment by segment with the historical individual stable window. The so-called historical individual stable window refers to the reference data set consisting of the system's long-term records of the individual's body temperature sequence and heart rate response data when there are no obvious health abnormalities. The matching process needs to consider the similarity of behavioral states (such as sitting still, walking, sleeping) and the similarity of environmental disturbance levels to ensure the rationality and pertinence of the comparison process. In each match, the system selects historical segments with high similarity and extracts their temperature change trajectories and heart rate deviations to form a stable reference set for comparative analysis.

[0215] Based on the difference between the current state unit and the matched historical segment, the system further calculates the temperature deviation confidence index. Specifically, the system overlaps and compares the temperature change trend curve of the current state unit with the temperature change trend of the corresponding interval in the historical segment, quantifies the degree of deviation in multiple dimensions such as slope, fluctuation range and turning point, and combines it with the heart rate deviation index of the period for joint evaluation, and outputs the temperature deviation confidence. For example, if the current body temperature change shows a steeper rising trend than the same behavioral state in the historical stable window, and is accompanied by a slight increase in heart rate, the system can assign it a higher confidence score to characterize the probability that there may be an abnormality in the current state.

[0216] The system then aggregates the temperature deviation confidence indicators of all state units to construct an overall sensitivity distribution curve for the set of candidate abnormal fragments. This distribution curve reflects the relative ranking and aggregation trend of the temperature deviation degree of each state unit in the current fragment. To enhance the guidance of individualized features, the system compares this distribution curve with the corresponding sensitivity threshold distribution in the historical individual stability window, thereby extracting the sensitivity deviation range of the current fragment relative to the historical benchmark. This deviation range statistically reflects the systematic increase or decrease in the current abnormality level, and can provide a basis for dynamic adjustment of the sensitivity level on top of the original recognition framework.

[0217] After completing the extraction of the offset interval, the system dynamically adjusts the upper and lower limits of the current threshold used to determine fever risk, thereby generating a dynamic fever determination threshold that is unique to the current individual and the current environmental scenario. This threshold can be fine-tuned according to the specific application, such as setting the adjustment ratio boundary, judgment accuracy tolerance, etc., to ensure that the system output achieves a reasonable balance between responsiveness and robustness. For example, for a sudden rise in body temperature at rest at night, the system can set a higher sensitivity to ensure timely recognition; for a rise in body temperature after exercise during the day, the system will consider environmental factors and behavioral patterns to appropriately relax the threshold range to avoid misjudgment.

[0218] In summary, this embodiment provides a method that fully integrates current data and historical stable windows to dynamically adjust the fever determination threshold, and has a high degree of personalization, self-adaptation and nonlinear response capabilities.

[0219] The suggestion generating unit 105 is configured to output a response suggestion with human behavior guiding characteristics according to the built-in nursing response rule library after detecting the decision output that triggers the health response.

[0220] The suggestion generation unit 105 plays a key role in connecting the upper and lower levels and closed-loop linkage in the AI-driven dynamic body temperature monitoring system of the present invention. The input of this unit is the "health response determination result" output by the abnormality recognition engine 104, which is specifically in the form of a Boolean signal (such as true or false) indicating whether a health response is required for the current time segment, and may be accompanied by auxiliary information such as the abnormal sensitivity score value corresponding to the response, the triggering time, the score duration, and the slope of the score change curve. The generation of this Boolean signal is based on the score value obtained by the aforementioned multimodal fusion analysis path. , compared with the dynamic fever determination threshold T generated by the system according to the historical stable window W, when the score exceeds T, it is determined to be a trigger state.

[0221] On this basis, the suggestion generation unit 105 will call the built-in nursing response rule library, which is maintained in the form of a structured rule set. Each rule contains several condition fields, matching thresholds, and corresponding response suggestion entries. For example, the rule entry can be set with the following logic: "If the abnormality score is greater than 0.85, and the score lasts for more than 3 consecutive windows, and the airflow disturbance amplitude is less than a certain threshold (excluding exercise induction) in the past 15 minutes, then generate the suggestion of 'recommending immediate drinking of water and entering a resting state'; if the score is between 0.75 and 0.85, but accompanied by an increase in the heart rate imbalance index exceeding two times the standard deviation of the steady-state average, then generate a moderate intensity prompt of 'recommending remote medical consultation'."

[0222] The suggestion generation unit 105 must be implemented with a rule-matching engine capable of combining and determining the multi-dimensional features of the anomaly determination information (including anomaly score level, duration, environmental disturbance context, historical matching degree, heart rate imbalance index, etc.), and selecting the recommended output path that best suits the current situation based on a priority or weighting mechanism. This suggestion is output as structured data, including the response type (e.g., mild, moderate, severe), the recommended content (e.g., drink water, rest, seek medical attention, contact relatives), the recommended execution timeframe (e.g., immediately, within 15 minutes, or continuous observation for 1 hour), and an optional personalized justification (e.g., "Since changes in body temperature are highly synchronized with abnormal heart rate, it is recommended to contact a family doctor").

[0223] The output results of the suggestion generation unit can be fed back to the user through multiple channels, including but not limited to: sending to the corresponding mobile terminal application APP via Bluetooth to display graphic and text suggestion content; triggering voice prompts through the voice broadcast unit (such as infant care scenarios); sending to the bound remote medical management platform through the system's built-in push mechanism to realize remote monitoring linkage.

[0224] It is worth noting that the suggestion generation unit 105 has a certain strategic memory capability, that is, after the system has been running for a period of time, it can record the user's past acceptance behavior of the response suggestions, and adjust the rule output weight according to the user's response behavior to the suggestions (such as execution or ignoring), thereby realizing long-term fine-tuning and learning of the personalized suggestion generation path.

[0225] In a specific implementation, the suggestion generation unit 105 can be implemented by an inference engine based on a decision tree rule set, combined with a lightweight logistic regression model to perform suggestion weight sorting, and can also be expanded to a hierarchical intervention mechanism based on a rule-nested control chart to ensure that the suggestion content always provides personalized behavioral support to users on the premise of being credible, practical, and explainable.

[0226] Therefore, the suggestion generation unit 105 is not only a single prompt module, but a multi-dimensional linkage suggestion generation system that integrates the recognition engine judgment, individual physiological state, environmental context and historical response behavior feedback. It aims to convert AI recognition results into actual behavioral operations that can be perceived and executed by users, thereby significantly improving the practicality, guidance and long-term health management value of the dynamic body temperature monitoring system.

[0227] Furthermore, the response control unit is configured to perform the following steps before the suggestion generation unit outputs the human behavior guidance suggestion:

[0228] Based on the set of candidate abnormal segments output by the abnormality recognition engine, the response suggestion history corresponding to each trigger segment, the time deviation value between the response suggestion output time and the actual body temperature stabilization interval are extracted, and the corresponding response delay score is calculated. The response delay score is used to measure the response time efficiency of the individual body temperature recovery after the suggestion is issued; the response delay score is integrated with the user interaction behavior index in the adjacent time window, where the user interaction behavior index includes whether the user views the terminal prompt after the trigger moment, whether the interactive instruction is executed on the terminal, and the instruction response time, to construct the response cooperation state feature under this cycle. The response cooperation state feature is used to characterize the user's active participation level after receiving the suggestion; according to the degree of correlation between the response cooperation state feature and the stability of the abnormality score, the response intervention intensity adjustment parameter of the current cycle is calculated. The anomaly score stability is the variance between the anomaly sensitivity scores of adjacent time segments, which is used to reflect whether the abnormal state is in a continuous and stable stage. The response intervention intensity adjustment parameter is used to determine whether to delay the output of the suggestion, whether it is necessary to increase the clarity of the suggestion content, or whether to call a backup reminder method; combined with the response intervention intensity adjustment parameter, the current anomaly score trend change slope and the execution effectiveness results of the suggestions in the previous cycle, the output method of the suggestion generation unit is dynamically adjusted. The anomaly score trend change slope is the average slope value of the anomaly score change in multiple consecutive time windows, which is used to reflect whether the anomaly risk is rapidly increasing. The execution effectiveness result is judged based on whether the individual's body temperature drops significantly after the suggestion is issued. The output method includes controlling the text length, prompt rhythm and semantic clarity level of the suggestion to improve the suggestion adoption rate and intervention effect.

[0229] First, the system receives a set of candidate anomaly segments from the anomaly identification engine. Each segment represents a time segment identified as exhibiting significant sensitivity to abnormal body temperature. For each candidate segment, the system searches the historical suggestion database for corresponding response suggestions and extracts the following information: first, the time the suggestion corresponding to the segment was issued; second, the time the user actually took action after the suggestion was issued (such as clicking to confirm or executing the suggestion); and third, the time period after the suggestion was issued when the user's temperature stabilized and returned to the normal range. Using this data, the system calculates a "response delay score" for the suggestion. This score measures whether the time interval between the suggestion issuance and the user's actual temperature stabilization is reasonable or excessive. This score is expressed as a ratio: the numerator is the time the temperature first dropped after the suggestion was issued minus the time the suggestion was issued, and the denominator is the time from the start of the temperature drop to the completion of the temperature stabilization. This is used to normalize the score to prevent extreme values ​​from influencing the evaluation results.

[0230] For example, if a suggestion is issued at 8:00, the user's temperature begins to drop at 8:05, and returns to normal at 8:10, the response delay score can be expressed as a ratio of 5 minutes to 5 minutes, which is 1. If the temperature starts to drop much later, or the drop takes a long time, the score will be greater than 1, indicating a low response efficiency.

[0231] Subsequently, the system integrates the above-mentioned scores with the user interaction behavior indicators in the current cycle to construct a "response cooperation state feature". This feature is used to describe whether the user actively cooperates with the system behavior guidance after the suggestion is issued. It includes the following aspects: first, whether the user has viewed the terminal prompt information, which can be judged by whether the system detects screen wake-up or application foreground activity; second, whether the user has performed an operation response, such as clicking a button, voice response, etc.; third, the time interval between the suggestion issuance and the first response, that is, the response delay time. The system quantifies these three contents into Boolean values ​​or normalized real values, and together with the above-mentioned response delay score, it constitutes the behavior cooperation state vector of the current cycle. Specifically, the first item in the state vector indicates whether the user viewed the prompt, the second item indicates whether the operation response was performed, the third item is the normalized value of the response time, and the fourth item is the response delay score, which is used to comprehensively evaluate the user's response cooperation level.

[0232] After obtaining the aforementioned behavioral state characteristics, the system further evaluates whether the current abnormal state is in a "continuously stable" phase. To this end, the system calculates the variance between the abnormality sensitivity scores output by the anomaly identification engine in several consecutive time segments as a measure of "anomaly score stability." The smaller the variance of the score, the more consistent the current abnormal state remains across multiple time segments, demonstrating continuity and stability. Conversely, if the variance is large, it indicates that the abnormal state may not yet have formed a stable trend. The system sets a score stability threshold. If the variance is below this threshold, the abnormal state is considered stable; otherwise, it is considered unstable.

[0233] At this point, the system inputs the aforementioned response coordination state characteristics and the stability of the anomaly score into the control strategy mapping logic to calculate the "response intervention intensity adjustment parameter." This parameter is a real-valued value that represents the degree to which the system should strengthen or weaken the content of the suggestion. If the user responds positively but the anomaly status continues to worsen, the suggestions should be more guiding. If the user responds negatively or the anomaly status remains uncertain, the suggestion output can be appropriately delayed or a lighter prompt can be used. This parameter can range from 0 to 1, with higher values ​​indicating a more proactive intervention strategy when outputting suggestions.

[0234] In addition, the system introduces two important parameters for assessing current risk trends: First, the slope of the anomaly score trend change, which is used to measure whether the anomaly score is rapidly rising, falling, or remaining flat across multiple time segments. This slope can be obtained by calculating the average of the score changes in each time segment. For example, if the score is consecutively 0.6, 0.7, 0.9, and 1.1, the trend is clearly rising. Second, the effectiveness of the execution of the recommendations in the previous cycle is used to determine whether the previous round of recommendations successfully triggered a significant drop in body temperature. If the body temperature drops by more than a preset temperature threshold (such as 0.5°C) within a certain period of time after the recommendation is issued, it is considered valid; otherwise, it is invalid.

[0235] Ultimately, the system makes a comprehensive assessment of the intervention intensity adjustment parameter, the slope of the anomaly score trend, and the effectiveness of the previous cycle's recommendations. This assessment then controls the output strategy of the recommendation generation unit. This strategy can include the following adjustments: first, controlling the length and detail of the recommendation text; second, adjusting the prompt cadence, such as whether to repeat the text or add voice prompts; third, improving semantic clarity, such as adding specific action instructions or warnings to the recommendations; and fourth, invoking alternative reminder methods, such as vibration, lighting, or external device prompts.

[0236] The following will systematically illustrate the entire workflow of the above-mentioned response control unit with the help of a complete specific example, covering the entire process from receiving candidate abnormal fragments to adjusting the recommended output strategy.

[0237] Taking a user's monitoring period from 8:00 AM to 8:15 AM as an example, the system generates a frame of data every 30 seconds, for a total of 30 frames. The temperature sensing unit collects skin temperature, the environmental collection unit acquires humidity, air temperature, and airflow disturbance amplitude, and the collaborative sensing unit acquires heart rate variability. After preliminary analysis of the data, the anomaly recognition engine marks candidate anomaly segments in the time period from 8:06 AM to 8:08 AM (i.e., frames 13 to 16), with the following anomaly scores:

[0238] = 0.65 ;

[0239] = 0.72 ;

[0240] = 0.78 ;

[0241] = 0.69;

[0242] At this time, the system records the current dynamic fever determination threshold as:

[0243] θ = 0.70;

[0244] 1. Calculate the response delay score:

[0245] Retrieving the user's response records for similar segments during the same time period on the previous day yields the following historical data:

[0246] Previous cycle response suggestion time: 8:04:00;

[0247] The time when the user's body temperature first dropped: 8:08:00;

[0248] Time for body temperature to stabilize and return to normal: 8:12:00;

[0249] The delay time for the first drop in body temperature is:

[0250] = 8:08 - 8:04 = 4 minutes

[0251] The delay time for body temperature stabilization is:

[0252] = 8:12 - 8:04 = 8 minutes;

[0253] When calculating the response delay score, the system uses a dual-indicator mean normalization scoring strategy:

[0254] Delayed Scoring =

[0255] ;

[0256] 2. Build response coordination status features:

[0257] The user behavior log for the current cycle shows:

[0258] Tip time: 8:06:30;

[0259] User viewed the prompt time: 8:06:42, viewing delay = 12 seconds;

[0260] User operation time: 8:06:50, response delay = 20 seconds;

[0261] The system sets the maximum viewing window to 60 seconds, and normalization yields:

[0262] Check: 1;

[0263] Whether to operate: 1;

[0264] Operation Delay Score =

[0265] ;

[0266] The response matching state eigenvector is:

[0267] R = [1, 1, 0.667];

[0268] Incorporate the delayed scores to form the complete matching feature vector:

[0269] F = [1, 1, 0.667, 0.333];

[0270] 3. Calculating the stability of anomaly scores:

[0271] The candidate anomaly score sequence is:

[0272] S = [0.65, 0.72, 0.78, 0.69];

[0273] Calculate the mean:

[0274] ;

[0275] Calculate the variance:

[0276] ;

[0277] The anomaly score stability is:

[0278] Stab = 0.00225 (lower than the preset threshold of 0.005, indicating abnormal stability);

[0279] 4. Calculate the slope of the abnormal score trend change:

[0280] Define the mean slope as the average of the differences between adjacent ratings:

[0281] Δ1=0.72−0.65=0.07;

[0282] Δ2=0.78−0.72=0.06;

[0283] Δ3=0.69−0.78=-0.09;

[0284] Average slope:

[0285] ;

[0286] The system preset significant rising threshold is 0.02, which has not been reached at present, so the upward trend is considered general.

[0287] 5. Integrate and adjust the recommended output:

[0288] Refer to the effectiveness of the previous cycle's recommendations (e.g., a 0.6°C drop in body temperature is considered effective), and combine the above parameters:

[0289] The response delay score is low (slow recovery), the user response is fast (good cooperation), the abnormal status is stable (small score fluctuation), the trend growth is slight (the upward trend is not obvious), and the previous suggestion is valid.

[0290] The system calculates the response intervention intensity adjustment parameter as 0.78 (based on the strategy mapping function) and determines that the current output should enhance the suggestion expression, including the following strategies:

[0291] The recommended text is adjusted from the simplified "Please pay attention to rising body temperature" to the detailed "Please enter a ventilated environment and replenish water, avoid strenuous activities"; the recommended prompt method is adjusted to voice + text + vibration combined output; the prompt frequency is adjusted from the original 1 time / 30 seconds to 2 times / 30 seconds.

[0292] Before generating behavioral guidance recommendations, the system comprehensively assesses multiple factors to determine the strength and presentation of the recommendations. This assessment is implemented by a rule engine called a "strategy mapping function." This function is not an abstract mathematical expression, but rather an executable rule function with specific computational logic. It accepts multiple input parameters and outputs a rating (low, medium, or high) that controls the presentation of the recommendations.

[0293] The input to this policy mapping function includes the following parts:

[0294] Response delay score: represents how quickly the user's body temperature stabilizes after the last suggestion was issued. A smaller value means a slower recovery. User response cooperation score: includes whether the user clicks to view the prompt, whether the terminal is operated, the time of response, etc. A higher score indicates a more active user. Abnormal score stability: indicates whether the current abnormality exists continuously. If the abnormality score does not change much within multiple time windows, it is considered to have high stability. Abnormal score trend slope: indicates whether the abnormality score continues to rise, which is a judgment on the trend of risk escalation. Whether the previous round of suggestions is effective: determines whether the last suggestion led to an improvement in body temperature, expressed as a Boolean value (true / false).

[0295] Based on these parameters, the system uses a conditional tree structure to output a suggestion level label (low / mid / high), thereby determining whether the behavioral suggestion is output in the form of a silent prompt, a small text reminder, or a full-width interactive method with voice and operation guidance.

[0296] The following is an example code implementation of the policy mapping function:

[0297] def strategy_mapping_function(delay_score, user_score, stability_score, trend_slope, last_effective):

[0298] # Initialize the recommended level to low

[0299] suggestion_level = 'low'

[0300] # Conditional judgment logic

[0301] if trend_slope>0.05 or stability_score<0.02:

[0302] # If the abnormal score increases rapidly or the abnormality persists, the intervention level needs to be increased.

[0303] if user_score>= 0.7:

[0304] if delay_score>0.4 or not last_effective:

[0305] suggestion_level = 'high'

[0306] else:

[0307] suggestion_level = 'mid'

[0308] else:

[0309] suggestion_level = 'mid'

[0310] else:

[0311] if user_score<0.4 and delay_score<0.3:

[0312] suggestion_level = 'low'

[0313] elif user_score>0.5 or last_effective:

[0314] suggestion_level = 'mid'

[0315] return suggestion_level

[0316] The suggestion_level returned by this function will be used to control the output of the suggestion generation unit:

[0317] 'low': Displays an icon or silent flashing reminder without any text content.

[0318] 'mid': Displays concise text suggestions, such as "Drink some water to cool down," with accompanying icon prompts.

[0319] 'high': Provides advice in the form of voice, text, and images, and adds clear behavioral guidance instructions, such as "Immediately stop outdoor activities and measure your temperature."

[0320] The above process completely covers all the calculation logic of the response control unit, including historical data callback, behavior quantification, stability analysis, trend judgment, parameter fusion and recommendation output control.

[0321] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An AI-driven dynamic body temperature monitoring system, characterized in that: include: The temperature sensing unit is used to collect the raw data stream of skin surface temperature at a preset sampling frequency and construct a basal body temperature sequence; A collection unit, configured to obtain the ambient temperature, humidity, and airflow disturbance amplitude of the area where the temperature sensing unit is located, so as to construct a dynamic environmental state vector; Collaborative sensing unit, used to continuously detect heart rate variability and extract short-term heart rate imbalance features; The abnormality recognition engine is used to construct a body temperature environment response association matrix for a local period based on the basal body temperature sequence and the environmental state vector, extract the nonlinear sensitive areas of body temperature changes to changes in environmental parameters, and generate a temperature-loop coupling map; based on the characteristics of short-term heart rate imbalance, the temperature-loop coupling map is dynamically labeled and embedded to form a three-modal state fusion tensor; based on a multi-scale discriminant model that integrates the attention mechanism and the graph neural network structure, the abnormal sensitivity score of each state unit in the three-modal state fusion tensor is calculated, and a set of candidate abnormal fragments is output; the candidate abnormal fragment set is subjected to comparative matching based on the historical individual stable window, the upper and lower limits of the current judgment threshold are adjusted, and a dynamic fever judgment threshold for the current individual and scene is generated; the abnormal sensitivity score is compared with the dynamic fever judgment threshold to form a decision output on whether to trigger a health response; a suggestion generating unit configured to output a response suggestion having human behavior guiding characteristics according to a built-in nursing response rule base after detecting a decision output triggering a health response; The anomaly recognition engine is specifically used to: The matching deviation corresponding to the short-term heart rate imbalance feature is used as a dynamic label value and embedded into each graph node in the temperature-environment coupling graph. The graph node is composed of a time segment in the body temperature environment response correlation matrix of the local period; The temperature-environment coupling graph after embedding dynamic labels is weighted according to the airflow disturbance amplitude micro-variation factor between the graph nodes to establish the connection relationship between the nodes, which is used to enhance the diffusion consistency of dynamic labels in areas with similar environmental disturbance characteristics. After completing the edge weight adjustment, continuous node groups are extracted based on the sliding time window. For each group of nodes, the skin temperature change trend, environmental state vector, and dynamic label value contained in the nodes are aggregated and combined to generate a multimodal segment vector. All multimodal segment vectors are mapped into tensor representations with unified structural dimensions to construct a trimodal state fusion tensor.

2. The AI-driven dynamic body temperature monitoring system according to claim 1, characterized in that: The system further includes a response control unit, wherein the response control unit is configured to perform the following steps before the suggestion generating unit outputs the human behavior guidance suggestion: Based on the set of candidate abnormal segments output by the abnormality recognition engine, the response suggestion history corresponding to each trigger segment is extracted, the time deviation value between the response suggestion output time and the actual body temperature stabilization interval is calculated, and the corresponding response delay score is calculated. The response delay score is used to measure the response time efficiency of the individual's body temperature recovery after the suggestion is issued; The response delay score is integrated with user interaction behavior indicators within adjacent time windows. The user interaction behavior indicators include whether the user views the terminal prompt after the trigger moment, whether the interactive command is executed on the terminal, and the command response time. This is used to construct the response cooperation status feature for this period. The response cooperation status feature is used to characterize the user's active participation after receiving the suggestion. Calculate the response intervention intensity adjustment parameter for the current cycle based on the degree of correlation between the response coordination state characteristics and the anomaly score stability. The anomaly score stability is the variance between the anomaly sensitivity scores of adjacent time segments, which is used to reflect whether the anomaly state is in a sustained and stable stage. The response intervention intensity adjustment parameter is used to determine whether to delay the output of the suggestion, whether to increase the clarity of the suggestion content, or whether to invoke a backup reminder method. The output mode of the suggestion generation unit is dynamically adjusted in combination with the response intervention intensity adjustment parameter, the current abnormal score trend change slope, and the execution effectiveness results of the suggestions in the previous cycle. The abnormal score trend change slope is the average slope value of the abnormal score change in multiple consecutive time windows, which is used to reflect whether the abnormal risk is rising rapidly. The execution effectiveness result is judged based on whether the individual's body temperature drops significantly after the suggestion is issued. The output mode includes controlling the text length, prompt rhythm, and semantic clarity level of the suggestion to improve the suggestion adoption rate and intervention effect.

3. The AI-driven dynamic body temperature monitoring system according to claim 1, characterized in that: The collaborative sensing unit is specifically configured to: An individual heart rate fluctuation baseline template is constructed to set an adaptive recognition threshold for short-term heart rate imbalance characteristics. The template is based on historical heart rate variability sequences obtained from the target individual at rest, during low-intensity activity, and during sleep. The template extracts the mean, standard deviation, and ratio of low-frequency and high-frequency components of the RR interval under each behavioral state to form a heart rate variability reference under multiple states. The RR interval sequence collected in the real-time sliding window is input into the frequency domain feature extraction process to obtain the low-frequency power, high-frequency power and their ratio in the current time window. At the same time, the change slope of the RR interval sequence is calculated to obtain the real-time heart rate variability characteristics of the current time window. Matching the real-time heart rate variability feature with the individual heart rate fluctuation baseline template state by state, and outputting a matching deviation degree to reflect the degree of deviation between the current state and the individual baseline state; Obtaining the ambient temperature, humidity, and airflow disturbance amplitude of the current time window in the area where the temperature sensing unit is located, combining the matching deviation with the current environmental state vector, and generating a state adjustment factor as an input for dynamically adjusting the sensitivity and tolerance range of short-term heart rate imbalance feature recognition; Under the control of the state adjustment factor, dynamic sensitivity judgment is performed on the matching deviation to obtain a short-term heart rate imbalance feature.

4. The AI-driven dynamic body temperature monitoring system according to claim 1, characterized in that: The anomaly recognition engine is specifically used to: Extracting an initial embedding representation of each state unit in the trimodal state fusion tensor, the initial embedding representation including a dynamic label value, a skin temperature change trend, and an environmental state vector, and performing position encoding according to the connection relationship between nodes in a graph structure; Based on the initial embedding representation and node position encoding, the importance weights of the three types of modal data in each state unit are calculated through a fusion attention mechanism to form a multimodal fusion attention matrix, which is used to enhance the expression contribution of different modal data in temporally adjacent states; Inputting the multimodal fusion attention matrix and the initial embedding representation of each state unit in the trimodal state fusion tensor into the graph neural network structure, performing a cross-node feature aggregation operation, and obtaining an aggregated feature vector of each state unit under its adjacency structure; Based on the aggregated feature vector, its cross-scale neighborhood distribution features are extracted through graph convolution kernels of different scales, and the convolution output results of each scale are fused to calculate the abnormal sensitivity score of each state unit in the trimodal state fusion tensor.

5. The AI-driven dynamic body temperature monitoring system according to claim 1, characterized in that: The anomaly recognition engine is specifically used to: Based on the aggregated feature vectors of each state unit in the candidate abnormal segment set in the trimodal state fusion tensor, the combined fluctuation amplitude including the skin temperature change trend and the environmental state vector is calculated to characterize the comprehensive change level of the current temperature-environment response; The combined fluctuation amplitude is used as input and the overlapping intervals are matched segment by segment with the historical individual stable windows stored in the system, and the historical reference temperature sequence and corresponding heart rate deviation of each segment under the same behavioral state and similar environmental disturbance are extracted; For each state unit in the current candidate abnormal fragment set, an offset score is calculated based on the difference between the state unit and the historical individual stable window, and a temperature deviation confidence index of the state unit is obtained to reflect the relative abnormality degree in the current state; Aggregating the temperature deviation confidence indicators of all state units, constructing a sensitivity distribution curve of the current candidate abnormal fragment set, and comparing the sensitivity distribution curve with the threshold distribution in the historical individual stable window to extract the sensitivity offset interval; Based on the sensitivity offset range, the upper and lower limits of the current determination threshold are dynamically adjusted to generate a dynamic fever determination threshold for the current individual and scene.

Images