AI traditional chinese medicine smart watch health early warning system fusing five elements and six qi and environment perception

Abstract

This invention relates to the intersection of smart wearable devices, digital TCM, and artificial intelligence technologies. It discloses an AI-powered TCM smartwatch health warning system integrating Five Elements and Six Qi theory with environmental perception. The system includes: a data acquisition module that collects real-time physiological data and temperature and humidity data from the user's wrist skin, outputting standardized physiological feature vectors and time-series temperature and humidity data; a Five Elements and Six Qi analysis module that generates constitution characteristics, long-term trend indicators, and comprehensive environmental Qi and Qi characteristics; an AI fusion warning module that integrates standardized physiological feature vectors, time-series temperature and humidity data, constitution characteristics, and comprehensive environmental Qi and Qi characteristics to form a fused feature sequence, which is input into a pre-trained lightweight improved temporal convolutional network model to output short-term, immediate warning results; and a human-computer interaction module that displays long-term trend indicators and short-term, immediate warning results, and pushes personalized conditioning suggestions. This invention effectively solves key problems in the modern application of TCM theory.

Description

Technical Field

[0001] This invention relates to the intersection of smart wearable devices, digital TCM, and artificial intelligence technologies, specifically to the fields of smart wearable devices and TCM health management technology. In particular, it relates to an AI smartwatch health early warning system that integrates the TCM Five Elements and Six Qi theory with multimodal environmental perception. This system is suitable for providing personalized and real-time TCM health risk early warning and conditioning guidance through a "wearable device data collection + mobile terminal display" model. Background Technology

[0002] With the popularization of smart wearable devices and the advancement of the modernization of traditional Chinese medicine (TCM), TCM health management devices have gradually become a research hotspot in the field of health technology. Existing TCM health management devices have several shortcomings in practical applications: First, they lack a systematic and quantitative application of the TCM theory of Five Elements and Six Qi. The calculation of constitution and Qi relies heavily on empirical judgment, lacking clear scientific basis and failing to dynamically reflect the interaction between constitution and environment. Second, most focus only on monitoring single physiological data, neglecting the synergistic influence of wrist microenvironment temperature and humidity with macro-meteorological and Qi parameters. Feature fusion logic is not rigorous, and weight settings lack reasonable support. Third, it is difficult to balance model complexity with the computing power and battery life of wearable devices, and the representativeness of training data is insufficient, leading to limitations in real-time performance and... The following issues hinder the development of existing devices: First, accuracy in early warning is difficult to achieve simultaneously. Second, the complex calendar calculations and deductions related to the Five Elements and Six Qi theory are mostly performed in real time on wearable devices, significantly increasing the device's workload and affecting system smoothness. Third, the warning hierarchy is ambiguous, with confusion between long-term trend prediction and short-term immediate warning concepts, and the expression lacks rigor, easily leading to user misunderstanding. Fourth, the generation mechanism of TCM conditioning suggestions is unclear, lacking a closed-loop design for user feedback and model optimization, resulting in insufficient dynamic adaptability between health assessment and early warning. Fifth, the limited display space of wearable devices makes it difficult to effectively display complex early warning information and conditioning suggestions, severely impacting user experience. These problems prevent existing devices from fully leveraging the advantages of TCM's "prevention of disease" approach, failing to provide users with accurate, convenient, and personalized health management services, and limiting the large-scale application and promotion of TCM health management technologies. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned deficiencies in existing TCM health management devices by providing an AI-powered TCM smartwatch health early warning system that integrates Five Elements and Six Qi theory with environmental perception. Specific objectives include: establishing a digital and quantitative application system for the Five Elements and Six Qi theory; clarifying the calculation methods for constitution and Qi; and realizing the dynamic representation of the interaction between constitution and environment. Specifically, it employs a lightweight improved TCN model and the HDPSO hyperparameter optimization algorithm proposed in this invention to optimize the multi-dimensional feature fusion logic, clarify the scientific basis for dimension alignment methods and weight coefficients, and improve the rigor and rationality of feature fusion. It designs a collaborative architecture of "watch-side data collection + mobile-side display," combining a lightweight model and hyperparameter optimization algorithm to balance model performance and terminal adaptability, ensuring system real-time performance and low power consumption. Finally, it constructs a two-layer early warning system of "long-term trend indication + short-term immediate warning," clarifying the mapping rules between warning levels and risk grades, and improving the rigor of warning statements.

[0004] To achieve the above objectives, the following technical solution is adopted: This invention provides an AI-powered TCM smartwatch health warning system that integrates Five Elements and Six Qi with environmental perception, comprising: a data acquisition module deployed on the smartwatch, a Five Elements and Six Qi analysis module deployed on the cloud and mobile phone, and an AI fusion warning module and a human-computer interaction module deployed on the mobile phone; The data acquisition module is used to collect the user's heart rate, blood oxygen saturation, skin conductivity physiological data, and wrist skin surface temperature and humidity data in real time. After data preprocessing, it outputs standardized physiological feature vectors and temperature and humidity time series data and uploads them to the mobile phone. The Five Elements and Six Qi Analysis Module includes an initialization calculation unit located in the cloud and a real-time fusion unit located on the mobile device. The initialization calculation unit performs Heavenly Stem-Branch conversion and constitution coding based on the user's input birth time, generates constitution characteristics, calculates annual luck parameters and basic constitution-luck interaction values, and generates long-term trend prompts. The real-time fusion unit receives temperature and humidity time-series data uploaded from the smartwatch and combines it with macro-meteorological data and annual luck parameters sent from the cloud to generate comprehensive environmental luck characteristics. The AI ​​fusion early warning module is used to receive and fuse the standardized physiological feature vector, temperature and humidity time series data, physical characteristics and environmental and weather comprehensive features to form a fusion feature sequence, and input it into a pre-trained lightweight improved temporal convolutional network model to output short-term real-time early warning results. The human-computer interaction module is used to display the long-term trend prompts and short-term real-time warning results, and push personalized conditioning suggestions based on the preset TCM conditioning rule engine.

[0005] Furthermore, the initialization calculation unit is configured to: The system converts the birth time into the Heavenly Stems and Earthly Branches based on the user's input, and generates physical characteristics represented by five-dimensional binary one-hot encoding based on the preset Five Elements mapping table. The physical imbalance index is calculated based on the physical characteristics. The physical imbalance index is used to characterize the degree of imbalance of the user in the five elements. The calculation method is as follows: assign weight coefficients obtained by optimizing through the attention mechanism to each of the five physical dimensions, and calculate the sum of the weighted absolute differences between the actual encoded value and the balance reference value of each dimension. Calculate the annual fortune parameters based on the sexagenary cycle of the current year; The basic interaction value between constitution and luck is calculated based on the physical characteristics and the annual luck parameters; and long-term trend prompts are generated based on a pre-set constitution-luck-seasonal mapping knowledge base.

[0006] Furthermore, the AI ​​fusion early warning module includes a feature fusion layer, which projects the standardized physiological feature vector, temperature and humidity time-series data, physical characteristics, and environmental and atmospheric comprehensive features onto a unified embedding space through a 1×1 convolutional layer and then splices them in time to generate the fused feature sequence.

[0007] Furthermore, the hyperparameters of the lightweight improved temporal convolutional network model are obtained through a hybrid discrete particle swarm optimization algorithm. The hybrid discrete particle swarm optimization algorithm achieves hyperparameter optimization through a three-stage process, including: a stage of generating an initial particle swarm based on the improved theorem, a stage of global exploration through discrete particle swarm optimization, and a stage of local development combined with hill climbing. The global exploration stage and the local development stage are repeated until the preset maximum number of iterations is reached, or the global optimal fitness is not improved for 5 consecutive rounds, and the final global optimal hyperparameters are output.

[0008] Furthermore, the objective function of the hybrid discrete particle swarm optimization algorithm is: Where X is the hyperparameter vector to be optimized. The health warning accuracy of the model on the validation set. This refers to the inference latency of the model on mobile devices.

[0009] Furthermore, in the stage of generating the initial particle swarm based on the improved theorem, the hybrid discrete particle swarm optimization algorithm is configured to: combine the hyperparameter search space of the lightweight improved temporal convolutional network model to generate at least one heuristic particle, the hyperparameter values ​​of which are determined based on preset empirical rules, and the remaining particles are generated randomly to form the initial particle swarm.

[0010] Furthermore, in the phase of global exploration via discrete particle swarm optimization, the hybrid discrete particle swarm optimization algorithm updates the particle velocity and position in the following manner: The velocity update is based on the particle's inertial velocity, the difference between the particle's own historical best position and its current position, and the difference between the group's global best position and its current position, and is fused according to a preset weight coefficient to determine the updated velocity vector. Position update: Adjust the particle's current position vector according to the updated velocity vector, and generate new hyperparameter candidate solutions based on the adjusted position vector, while ensuring that the values ​​of each hyperparameter in the newly generated hyperparameter candidate solutions conform to the predefined search space constraints.

[0011] Furthermore, in the stage of local development combined with the hill-climbing method, the hybrid discrete particle swarm optimization algorithm is configured to: randomly trigger a local search for the current hyperparameter candidate solution with a preset probability; when a local search is triggered, fix some hyperparameters in the current hyperparameter candidate solution, fine-tune the remaining hyperparameters, and calculate the fitness value of the new solution. If the new solution is better than the current solution, it is replaced; otherwise, the original solution is retained.

[0012] Furthermore, the short-term immediate warning result includes a risk level, which is determined by comparing the warning probability output by the lightweight improved temporal convolutional network model with a preset dynamic threshold; wherein, different disease types correspond to different dynamic thresholds, and the dynamic thresholds are determined separately for different disease types by performing receiver operating characteristic curve analysis on clinical data.

[0013] Furthermore, the human-computer interaction module also includes a user feedback closed-loop unit, which is configured to: Collect user feedback on the effectiveness of the personalized treatment suggestions pushed to them; The effect evaluation feedback is associated with the feature sequences and early warning results of the corresponding time period to construct a feedback dataset; Based on the feedback dataset, the artificial intelligence model in the AI ​​fusion early warning module is incrementally learned using an online gradient descent algorithm to dynamically optimize the model parameters.

[0014] Compared with the prior art, the present invention achieves the following beneficial effects: 1. This invention establishes for the first time a digital system of "cloud computing + mobile terminal access" for the Five Elements and Six Qi theory, unifies the encoding and quantification logic of physical constitution characteristics, clarifies the dimensional alignment method of multi-dimensional characteristics, effectively solves the key problems of the modern application of TCM theory and the lack of rigorous feature fusion logic, and achieves a consistency of 82% in physical constitution quantification, filling the gap in the existing technology for the quantitative application of TCM theory.

[0015] 2. This invention employs a lightweight improved TCN model and the HDPSO hyperparameter optimization algorithm proposed in this invention. Through a three-stage optimization process, it achieves efficient hyperparameter search. The inference latency of the model on the mobile device is ≤65ms, and the accuracy of short-term real-time warning is ≥85%. Meanwhile, the watch only undertakes the data collection and transmission functions, reducing battery life loss and perfectly balancing warning accuracy, real-time performance and terminal compatibility.

[0016] 3. This invention constructs a three-dimensional service architecture of "long-term trend prompts + short-term instant warnings + closed-loop feedback". Long-term trends can provide seasonal health guidance 3-6 months in advance, and short-term warning response time is ≤1 minute. Combined with mobile terminal pop-up windows, push notifications and other display methods, information is delivered more promptly and displayed more clearly. Incremental learning improves the adaptability of conditioning suggestions by 40%, significantly enhancing user participation in health management.

[0017] 4. This invention has been thoroughly clinically validated and is highly practical. Through clinical validation, the system's trend prediction accuracy for seasonal diseases reached 78%, and its real-time early warning accuracy for acute physiological abnormalities met the needs of clinical auxiliary judgment. It provides intelligent and quantifiable tool support for the TCM concept of "prevention of disease," and has broad application prospects.

[0018] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0019] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 This is a schematic diagram of the health early warning system architecture of an AI TCM smartwatch that integrates the Five Elements and Six Qi with environmental perception, provided by an embodiment of the present invention; Figure 2 This is a flowchart of the health early warning system for an AI-powered TCM smartwatch that integrates the Five Elements and Six Qi theory with environmental perception, provided in an embodiment of the present invention. Figure 3 This is a flowchart of the three-stage optimization process of the HDPSO algorithm according to an embodiment of the present invention; Figure 4 This is a schematic diagram showing the mobile terminal interface according to an embodiment of the present invention; Figure 5 This is a schematic diagram showing the performance comparison between the algorithm of this invention and traditional algorithms. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0022] Figure 1 This is a schematic diagram of the health early warning system architecture of an AI TCM smartwatch that integrates the Five Elements and Six Qi with environmental perception, provided by an embodiment of the present invention; Figure 2 This is a flowchart illustrating the workflow of the AI-powered TCM smartwatch health warning system, which integrates Five Elements and Six Qi with environmental perception, as provided in this embodiment of the invention. Figure 1 and Figure 2 As shown, an AI-powered TCM smartwatch health early warning system integrating Five Elements and Six Qi with environmental perception is presented. This system adopts a collaborative mode of "data collection on the smartwatch + display on the mobile phone." The smartwatch is responsible for data collection and preliminary processing, while all early warning results, trend alerts, and interactive functions are implemented through the mobile phone terminal. The system includes: a data collection module 110 deployed on the smartwatch; a Five Elements and Six Qi analysis module 120 deployed in the cloud and on the mobile phone; and an AI fusion early warning module 130 and a human-computer interaction module 140 deployed on the mobile phone. The specific technical solution is as follows: The data acquisition module 110 is used to collect the user's heart rate, blood oxygen saturation, skin conductivity physiological data, and wrist skin surface temperature and humidity data in real time. After data preprocessing, it outputs standardized physiological feature vectors and temperature and humidity time series data and uploads them to the mobile phone. Step S1: Construct data acquisition module 110: The data acquisition module 110 includes a multimodal physiological sensing module 111 and a wrist temperature and humidity sensing module 112. 1. Multimodal physiological sensing module 111: (1) Sensor integration: integrates heart rate sensor, blood oxygen sensor and skin conductivity sensor to collect user heart rate (HR), blood oxygen saturation (SpO2) and skin conductivity (GSR) data in real time. The sampling frequency is 1Hz and the data accuracy is ±1bpm, ±2% and ±0.1μS / cm, respectively.

[0023] (2) Data preprocessing: Outliers are removed by the 3σ rule, the data is smoothed by moving average filtering (window size is 5s), and dynamic threshold filtering algorithm is added to resist sweat and movement interference, and standardized physiological feature vectors are output.

[0024] (3) Data transmission: Standardized physiological feature vectors are uploaded to the paired mobile terminal in real time via Bluetooth Low Energy (BLE) technology. The watch does not store the original data, but only caches the data after preprocessing in the last 5 minutes to cope with transmission interruption.

[0025] (4) Accuracy description: The accuracy of ±2% blood oxygen meets the requirements of health early warning auxiliary judgment. Combining multimodal data fusion can further improve sensitivity and avoid misjudgment of a single indicator.

[0026] 2. Wrist temperature and humidity sensing module 112: (1) Sensor selection: A high-precision digital temperature and humidity sensor is used to monitor the micro-environment temperature (range: 0-50℃, accuracy ±0.2℃) and relative humidity (range: 0-100% RH, accuracy ±2% RH) of the user's wrist skin surface, with a sampling frequency of 0.5Hz.

[0027] (2) Data time series alignment: Real-time output of temperature and humidity time series data, after time series alignment with physiological data, is uploaded to the mobile terminal along with the physiological feature vector for subsequent feature fusion analysis.

[0028] The Five Elements and Six Qi Analysis Module 120 includes an initialization calculation unit 121 located in the cloud and a real-time fusion unit 122 located on the mobile device. The initialization calculation unit 121 performs the Heavenly Stems and Earthly Branches conversion and constitution coding based on the birth time input by the user, generates constitution characteristics, calculates the annual luck parameters and the basic interaction value between constitution and luck, and generates long-term trend prompts. The real-time fusion unit 122 receives the temperature and humidity time series data uploaded by the smart watch and combines it with the macro meteorological data and annual luck parameters sent from the cloud to generate comprehensive environmental luck characteristics. Step S2: Construct the Five Elements and Six Qi Analysis Module (Cloud + Mobile Collaboration): Step S21: Construct initialization computing unit 121: The initialization calculation unit 121 is configured to: convert the birth time into a Heavenly Stem and Earthly Branch based on the user's input, and generate a constitution characteristic represented by a five-dimensional binary one-hot encoding based on a preset Five Elements mapping table; calculate a constitution imbalance index based on the constitution characteristic, which characterizes the user's degree of imbalance in the Five Elements dimensions. The calculation method is as follows: assign weight coefficients optimized through an attention mechanism to each of the five constitution dimensions, and calculate the sum of the weighted absolute differences between the actual encoded values ​​and the balance reference values ​​of each dimension; calculate the annual luck parameter based on the Heavenly Stem and Earthly Branch of the current year; calculate the constitution-luck basic interaction value based on the constitution characteristic and the annual luck parameter; and generate long-term trend prompts based on a preset constitution-luck-seasonal mapping knowledge base. The specific process is as follows: 1. Heavenly Stems and Earthly Branches Conversion and Constitution Encoding: Upon first use, users input their birth time (year, month, day) via the mobile app. The cloud then converts this into Heavenly Stems and Earthly Branches, generating constitution characteristics based on the Five Elements mapping table (using five-dimensional binary one-hot encoding, such as [1,0,0,0,0] representing a dominant Wood constitution). The Five Elements mapping table is shown in Table 1 below: Table 1 Five Elements Mapping Table

[0029] 2. Quantification of Physical Characteristics: The physical imbalance index characterizes the degree of imbalance in an individual's physical constitution across the five elements, resulting in a physical constitution score. The value (0 or 1) corresponding to the five dimensions in the five-dimensional binary one-hot encoding is calculated as follows:

[0030] The Constitutional Imbalance Index is used to characterize the degree of imbalance in an individual's constitution across the five elements. Essentially, it is the weighted Hatton distance between the actual constitution's unique vector and the balance reference vector. The larger the distance value, the higher the degree of constitution imbalance. Physical fitness score is the value (0 or 1) of the five-dimensional binary one-hot encoding corresponding to the five elements, which are wood, fire, earth, metal, and water respectively. The weighting coefficients for the five elements were initially set based on expert experience and subsequently optimized on a training set of 3000 people using an attention mechanism. The optimal values ​​were ultimately selected. (Corresponding to wood) (Corresponding to fire) (Corresponding to Earth), (Corresponding to Metal), (Corresponding to Water), satisfying the requirements. . : Five Elements Balance Reference Vector, with values ​​of This represents an ideal physical state where the five elements are balanced. The Five Elements Dimension Index has a value range of 1-5, corresponding to the five dimensions of wood, fire, earth, metal, and water.

[0031] The formula essentially calculates the weighted Manhattan distance between the actual body constitution unique heat vector and the balance reference vector. The larger the distance value, the higher the degree of body constitution bias.

[0032] 3. Calculation of annual luck parameters: Based on the current year's Heavenly Stem and Earthly Branch, calculate the excess / deficiency coefficient of the Five Elements (range 0.8-1.2) and the intensity parameter of the Six Qi (range 0.7-1.3) to form the annual luck characteristic matrix.

[0033] 4. Calculation of the basic interaction value between constitution and luck: The basic interaction coefficient between the two is calculated by a preset algorithm and synchronized to the watch for storage along with the constitution characteristics and the annual luck characteristic matrix.

[0034] 5. Long-term trend alert generation: Long-term trend alerts are generated in the cloud based on a pre-set "Constitution-Luck-Season" mapping knowledge base. After calculation at the beginning of each year, the alerts are synchronized to the mobile app in text form.

[0035] Step S22: Construct real-time fusion unit 122: Static data storage: The mobile device stores static physical characteristics, annual luck characteristics, and long-term trend prompts, while the watch does not store any static characteristic data.

[0036] Luck-Environment Interaction Quantification: The mobile device receives temperature and humidity data uploaded by the watch, and simultaneously obtains the daily weather luck adaptation value (based on macro meteorological data conversion) sent from the cloud. The comprehensive environmental luck characteristics are generated through the following formula:

[0037] : Comprehensive characteristics of environmental luck, used to quantify the synergistic impact of annual luck and macro-meteorology. The normalized score of the luck parameter, ranging from 0 to 1, is obtained by normalizing the annual excess / deficiency coefficient of the five lucks and the six qi transformation intensity parameter. Meteorological data normalization score, ranging from 0 to 1, is obtained by normalizing the wrist temperature and humidity data uploaded by the watch and the daily macro meteorological data sent from the cloud. The luck parameter weight coefficient is set to 0.6. The initial value is based on the theory of "correspondence between man and nature" and is optimized and verified in 300 clinical cases through ridge regression. The meteorological data weighting coefficient is set to 0.4. The initial value is based on the theory of "correspondence between man and nature" and is optimized and verified in 300 clinical cases through ridge regression.

[0038] AI fusion early warning module 130 is used to receive and fuse the standardized physiological feature vector, temperature and humidity time series data, physical characteristics and environmental and weather comprehensive features to form a fusion feature sequence, and input it into a pre-trained lightweight improved temporal convolutional network model to output short-term real-time early warning results. Step S3: Construct AI fusion early warning module 130: Step S31, Model Architecture Design: The core model employs a lightweight improved temporal convolutional neural network (TCN), built upon one-dimensional depthwise separable convolution (Dv1D), and uses the TensorFlow Lite framework for INT8 quantization and 30% pruning to ensure compatibility with embedded devices.

[0039] To efficiently optimize the hyperparameters of the TCN model, this invention proposes a Hybrid Discrete Particle Swarm Optimization (HDPSO) algorithm, which achieves hyperparameter optimization through a three-stage process: first, an initial particle swarm is generated based on an improved theorem; then, global exploration is performed through Discrete Particle Swarm Optimization (DPSO); and finally, local development is achieved by randomly combining hill climbing, balancing exploration and development capabilities and avoiding getting trapped in local optima.

[0040] Step S32: Feature Fusion Layer Design Furthermore, the AI ​​fusion early warning module includes a feature fusion layer, which projects standardized physiological feature vectors, temperature and humidity time-series data, physical characteristics, and comprehensive environmental and atmospheric features onto a unified embedding space through a 1×1 convolutional layer, and then performs temporal concatenation to generate the fused feature sequence. Specifically: (1) Dimensional alignment method: The mobile phone receives the physiological feature vector uploaded by the watch, and projects each original feature vector (physiological features, temperature and humidity data, physical characteristics, and environmental and luck comprehensive features) into a unified 64-dimensional embedding space through a 1×1 convolutional layer (e.g., the original 4-hour physiological data sequence is 3-dimensional × 16 time points = 48-dimensional, which is projected to 64-dimensional through 1×1 convolution; low-dimensional vectors are expanded to 64-dimensional through a fully connected layer), and then performs temporal splicing to form a 128-dimensional fused feature sequence.

[0041] (2) Time window setting: The fusion time window is 4 hours (16 15-minute intervals).

[0042] Step S33, Model Training Process: 1. Data Collection: Obtain physiological data, wrist temperature and humidity data, birth information, and 1-year health records (including disease occurrence time and symptom type) from 3,000 users of different age groups. Simultaneously collect meteorological data and luck parameters for the corresponding time periods. After data anonymization, a training dataset is formed, which meets the requirements for medical data compliance.

[0043] 2. Data preprocessing: The time series data is normalized (normalization range [0,1]). The physical constitution and luck features are encoded using one-hot encoding. The ratio of training set, validation set and test set is 8:1:1.

[0044] 3. Hyperparameter Optimization: The key hyperparameters of TCN are optimized using the HDPSO algorithm. The optimization objective is to maximize the warning accuracy on the validation set. The objective function is as follows:

[0045] The objective function value of the Hybrid Discrete Particle Swarm Optimization (HDPSO) algorithm is used to comprehensively evaluate the performance of the hyperparameter combination of a lightweight temporal convolutional network (TCN). The larger the function value, the better the overall performance of the corresponding hyperparameter combination. The hyperparameter vector to be optimized contains the core hyperparameters of the Lightweight Temporal Convolutional Network (TCN), specifically the number of kernels, batch size, and learning rate, with corresponding search spaces of [16, 32, 64] and [8, 16, 32], respectively. . Based on hyperparameter vectors The health warning accuracy of the constructed lightweight temporal convolutional network (TCN) on the validation set is a core metric for measuring the model's risk identification precision. Latency Based on hyperparameter vectors The inference latency of the constructed lightweight temporal convolutional network (TCN), expressed in milliseconds (ms), reflects the model's real-time inference efficiency on mobile devices. 0.8: The weight coefficient for the warning accuracy term, set based on the principle of "accuracy first, real-time performance second," highlighting warning accuracy as the core optimization objective. 0.2: The weight coefficient for the inference latency constraint term, set based on the principle of "accuracy first, real-time performance second," used to constrain model real-time performance during optimization, balancing warning accuracy and operational efficiency. 100: The baseline constraint threshold for model inference latency, expressed in milliseconds (ms), is the maximum allowable inference latency set in this invention, used to normalize the inference latency data to... The interval is used in the comprehensive calculation of the objective function.

[0046] 4. Model Training: The fused feature sequence is input into the lightweight TCN model, and the cross-entropy loss function is used. Iterative training is performed based on the Adam optimizer for 200 iterations. Training stops when the loss on the validation set does not decrease for 10 consecutive iterations. After training, redundant parameters are removed by model pruning to reduce computational complexity.

[0047] 5. Model Validation: The model performance was validated on the test set provided by the partner hospital. The accuracy of short-term real-time early warning was ≥85%, and the false alarm rate was ≤5%. The average inference latency on mainstream Android / iOS mobile terminals was 65ms, which meets the requirements for real-time display.

[0048] Step S34: Set key parameters: (1) TCN structural parameters: kernel size is 7, number of hidden channels is [16,32,64], inflation factor is [1,2,4], dropout rate is 0.2; after lightweighting, the model size is ≤4.5MB and the inference latency is ≤100ms.

[0049] (2) HDPSO parameters: particle swarm size is 20, maximum number of iterations is 30, inertia weight ω=0.7, cognitive coefficient c1=1.5, social coefficient c2=1.5.

[0050] (3) Early warning threshold: Furthermore, the short-term immediate warning results include a risk level, which is determined by comparing the warning probability output by the lightweight improved temporal convolutional network model with a preset dynamic threshold. Different disease types correspond to different dynamic thresholds, which are determined separately for each disease type through receiver operating characteristic (ROC) curve analysis of clinical data. Specifically: Based on clinical data from partner hospitals, dynamic thresholds for different disease types (respiratory diseases: 0.65, digestive diseases: 0.70, cardiovascular diseases: 0.68, etc.) were determined using receiver operating characteristic (ROC) curves. The risk level classification rules for short-term immediate warnings are as follows: a warning probability ∈ [threshold, threshold + 0.1) is defined as "low risk", a warning probability ∈ [threshold + 0.1, threshold + 0.2) is defined as "medium risk", and a warning probability ≥ threshold + 0.2 is defined as "high risk".

[0051] Step S35: Complete Implementation Example of HDPSO Algorithm (I) Complete Implementation Details of the HDPSO Algorithm Step S351, Core Concepts and Particle Definition: 1. Definition of particle position: Let the particle... Let be the candidate solution vector for the hyperparameters, where For a certain hyperparameter value corresponding to TCN (e.g.) The number of convolution kernels, For batch size, (learning rate); accompanying binary vector Used to mark modifiable hyperparameter bits. =0 indicates that the corresponding hyperparameter can be adjusted. =1 indicates that the hyperparameters are fixed.

[0052] : No. Each particle represents a candidate hyperparameter solution vector, which is composed of multiple key hyperparameter values ​​of TCN; : No. The specific value of the th hyperparameter in each particle corresponds to a core hyperparameter of TCN (such as the number of convolutional kernels, batch size, learning rate, etc.). : The index identifier of a particle, used to distinguish different particles in a particle swarm; : Index identifier for hyperparameters, used to distinguish different hyperparameter bits in the candidate solution vector; : No. Each particle is assigned a binary tag vector, which is used to mark the modifiable hyperparameter bits in the candidate hyperparameter solution vector; : No. The first particle in the binary tag vector The digit value indicates that the corresponding hyperparameter is adjustable, while 1 indicates that the corresponding hyperparameter is fixed.

[0053] 2. Definition of particle velocity: velocity For an n-bit binary vector, by comparing it with the position vector The direction of interactive control hyperparameter update.

[0054] : No. The velocity vector of each particle is... A bit binary vector, used to interact with the position vector to control the direction of hyperparameter updates; : No. The th particle velocity vector in the th particle velocity vector The bit value is the core binary bit that controls the direction of hyperparameter updates; The number of dimensions of the hyperparameters should be consistent with the number of bits in the TCN's optimized hyperparameters, particle position vector, and velocity vector. Step S352, Definition of core operators: 1. Subtraction operator ( ): Calculates the difference between two vectors. If corresponding elements are the same, the result bit is 1; otherwise, it is 0. It is used to extract the difference between candidate hyperparameter solutions and the optimal solution. Example: .

[0055] in, The subtraction operator is used to calculate the difference between two vectors. If the corresponding elements are the same, the result bit is 1; otherwise, it is 0. Its core function is to extract the difference between the candidate solution and the optimal solution of hyperparameters. 2. Addition Operator ( ): Integrating the effects of inertia, local optima, and global optima on velocity, the formula is as follows: ( The uncertainty of binary vector fusion is resolved by selecting the velocity bit corresponding to the weight using random numbers.

[0056] in, The addition operator is used to fuse the effects of inertia, local optima, and global optima on velocity, achieving weighted fusion of binary velocity vectors. The final velocity vector after fusion is obtained by weighted fusion of multiple basic velocity vectors using an addition operator; : No. The weight coefficients of each basic velocity vector satisfy the condition that the sum of all weight coefficients is 1, and are used to allocate the proportion of influence of different velocity vectors on the final velocity. : No. The basic velocity vector participating in the fusion can be the particle inertial velocity, the local optimal velocity, the global optimal velocity, etc. : Index identifier of the underlying velocity vector participating in the fusion.

[0057] 3. Multiplication operators ( ): Updates the position vector, adjusting the corresponding position only when the velocity bit is 0. Example: .

[0058] in, The multiplication operator is used to update the position vector of a particle. It adjusts the corresponding position only when the velocity is 0. It is the core operator for updating candidate hyperparameter solutions.

[0059] Step S353, HDPSO three-stage execution process: The hyperparameters of the lightweight improved temporal convolutional network model are obtained through a hybrid discrete particle swarm optimization (DPSO) algorithm. The DPSO algorithm optimizes hyperparameters through a three-stage process: generating an initial particle swarm based on an improved theorem (Stage 1), performing global exploration using DPSO (Stage 2), and performing local development using a hill-climbing method (Stage 3). The global exploration and local development stages are repeated until a preset maximum number of iterations is reached, or five consecutive rounds show no improvement in global fitness, at which point the final globally optimal hyperparameters are output. Figure 3 The diagram shown is a three-stage optimization flowchart of the HDPSO algorithm according to an embodiment of the present invention, as detailed below: 1. Phase One: Initial Particle Swarm Generation In Phase 1, the hybrid discrete particle swarm optimization algorithm is configured as follows: It combines the hyperparameter search space of a lightweight improved temporal convolutional network model to generate at least one heuristic particle. The hyperparameter values ​​of this heuristic particle are determined based on preset empirical rules. The remaining particles are generated randomly to form the initial particle swarm. Specifically: Based on the improved theorems (Theorem 1: A subset of candidate hyperparameter solutions is still a valid solution; Theorem 2: A solution that satisfies the hyperparameter value constraints is a valid solution), combined with the TCN hyperparameter search space, three heuristic particles are generated (e.g., number of convolution kernels = 32, batch size = 16, learning rate = 0.005), and the remaining 17 particles are generated through random permutation to ensure that the initial population covers the key regions of the search space, with a total particle swarm size of 20.

[0060] 2. Phase Two: Global Exploration of DPSO In Phase Two, the hybrid discrete particle swarm optimization algorithm updates the velocity and position of particles in the following ways: Velocity update: Based on the particle's inertial velocity, the difference between the particle's historical best position and its current position, and the difference between the swarm's global best position and its current position, these are fused according to preset weighting coefficients to determine the updated velocity vector; Position update: Based on the updated velocity vector, the particle's current position vector is adjusted, and new hyperparameter candidate solutions are generated based on the adjusted position vector, while ensuring that the hyperparameter values ​​in the newly generated hyperparameter candidate solutions all conform to predefined search space constraints. Specifically: Speed ​​update: via formula Update speed.

[0061] Position update: via formula Adjust the position vector and generate new hyperparameter candidate solutions based on the adjusted position vector, ensuring that the hyperparameter values ​​meet the search space constraints (e.g., the number of convolution kernels is only selected from [16, 32, 64]).

[0062] Particle number The velocity vector of the next iteration is the new velocity after the iteration update; Particle number The inertial velocity vector of the next iteration reflects the velocity inertia of the particle in the previous iteration; : No. The local optimal position vector of a particle represents the optimal hyperparameter candidate solution vector searched during the iteration process of that particle; Particle number The current position vector of the next iteration corresponds to the current hyperparameter candidate solution vector; Xgb: the global optimal position vector of the particle swarm, representing the optimal hyperparameter candidate solution vector searched during the entire particle swarm iteration process; : The weighting coefficient of the inertial velocity vector, with a value of 0.3, is used to allocate the proportion of the influence of the particle's inertial velocity on the new velocity; : The weighting coefficient of the local optimal difference velocity vector, with a value of 0.35, is used to allocate the proportion of the influence of the difference between the particle's local optimum and its current position on the new velocity; : The weight coefficient of the global optimal difference velocity vector, with a value of 0.35, is used to allocate the proportion of the influence of the difference between the global optimal particle swarm and the current position on the new velocity.

[0063] 3. Phase Three: Partial Development of Hill Climbing In Phase 3, the hybrid discrete particle swarm optimization algorithm is configured to: randomly trigger a local search for the current hyperparameter candidate solution with a preset probability; when a local search is triggered, some hyperparameters in the current hyperparameter candidate solution are fixed, and the remaining hyperparameters are fine-tuned; the fitness value of the new solution is calculated, and if it is better than the current solution, it is replaced; otherwise, the original solution is retained. Specifically: Randomly generated ,like A local search is performed on the candidate hyperparameter solutions corresponding to the current particle.

[0064] Local search logic: Fix two hyperparameters, fine-tune the third hyperparameter (e.g., fix batch size = 16, learning rate = 0.005, and adjust the number of convolutional kernels from 32 to 64), calculate the fitness of the new solution (early warning accuracy + inference delay constraint), if it is better than the current solution, replace it; otherwise, retain the original solution.

[0065] : A randomly generated probability value, with a range of values. This is used to determine whether to perform a local search on candidate solutions for hyperparameters; The threshold for determining local search is 0.5, when the random probability value... A local search operation is triggered when the value exceeds this threshold.

[0066] 4. Iteration Termination: Repeat Phase 2 and Phase 3 until the maximum number of iterations of 30 is reached, or the global optimal fitness has not improved for 5 consecutive rounds, and output the final global optimal hyperparameter.

[0067] (II) Detailed Process of TCN Model Hyperparameter Update Step S354, Hyperparameter encoding mapping: 1. Map the hyperparameters to be optimized in the TCN model to the dimension of the particle position vector, as follows: Dimension 1: Number of convolution kernels, with values ​​{16, 32, 64}, corresponding to particle position values ​​{0, 1, 2}; Dimension 2: Batch size, with values ​​{8, 16, 32}, corresponding to particle position values ​​{0, 1, 2}; Dimension 3: Learning rate, with values ​​{0.001, 0.005, 0.01}, corresponding to particle position values ​​{0, 1, 2}.

[0068] 2. Example: The hyperparameter combination corresponding to the particle position vector (1,1,1) is (number of convolution kernels = 32, batch size = 16, learning rate = 0.005).

[0069] Step S355, Fitness Function Calculation: 1. For each combination of hyperparameters corresponding to a particle, construct a temporary TCN model and train it on the training set for 10 rounds.

[0070] 2. Calculate the early warning accuracy on the validation set. and reasoning delay Through formula Calculate the fitness value and comprehensively evaluate the performance of the hyperparameter combination.

[0071] Step S356, Hyperparameter Iterative Update: 1. Iterations 1-30: Each round executes DPSO global exploration, updates particle position and velocity, and generates new hyperparameter combinations; each round randomly triggers Hill Climbing local development to optimize the current optimal hyperparameters.

[0072] 2. Example of an iterative process: Initial globally optimal hyperparameters: (16, 8, 0.001), fitness = 0.78; 10th iteration: (32,16,0.005) is obtained through DPSO update, fitness = 0.85, and replaced with the new global optimum; 20th iteration: The result obtained through Hill Climbing optimization is (32, 16, 0.003), with a fitness of 0.87, which is then replaced with the new global optimum; 30th iteration: No better solution found, iteration terminated.

[0073] Step S357, Optimal Hyperparameter Output and Model Update: 1. After the iteration terminates, output the globally optimal combination of hyperparameters (e.g., number of convolution kernels = 32, batch size = 16, learning rate = 0.003).

[0074] 2. Configure the optimal hyperparameters into the TCN model, complete model training and lightweight processing, and push the model to the user terminal through the mobile app background to ensure the real-time performance and accuracy of the model on the mobile device.

[0075] The human-computer interaction module 140 is used to display the long-term trend prompts and short-term real-time warning results, and push personalized conditioning suggestions based on the preset TCM conditioning rule engine.

[0076] Step S4: Construct human-computer interaction module 140: Step S41, Dual-layer output display design: (1) Long-term trend prompts: The annually updated long-term trend prompts (such as "In 2025, the metal element will be strong, and those with a wood-type constitution should pay special attention to liver and gallbladder health") are displayed in a prominent position on the homepage of the mobile app, allowing users to click to view detailed interpretations and providing seasonal health guidance 3-6 months in advance.

[0077] (2) Short-term real-time early warning display: When the AI ​​fusion early warning module 130 identifies an abnormal real-time feature sequence, it will prompt the risk level (low, medium, high), associated disease type and early warning probability through mobile app pop-up window + push notification + vibration. The corresponding minute / hour level early warning can be clicked to enter the App to view risk details and cause analysis.

[0078] Step S42, Conditioning Suggestion Generation Mechanism: (1) Rule Engine Construction: A TCM conditioning rule engine is constructed to generate personalized suggestions based on the four-dimensional mapping relationship of "constitution-disease-luck-real-time status". The rule engine includes a dietary taboo database, an acupoint massage database, and a rest guidance database. Example rule: If the user has a wood type constitution + high cardiovascular risk + high humidity on the day, the recommendation is "Avoid cold and raw foods, massage the Neiguan acupoint for 5 minutes every day, and reduce prolonged sitting".

[0079] (2) Suggested Push Method: Treatment suggestions are displayed simultaneously with the alert results in a pop-up window of the mobile app, and also stored in the "Health Suggestions" section, allowing users to view and save them at any time. Some key suggestions can be set with timed reminders (such as acupressure massage time reminders). Figure 4 The image shown is a schematic diagram of the mobile terminal interface according to an embodiment of the present invention.

[0080] Step S43, User Feedback Closed-Loop Design: Furthermore, the human-computer interaction module 140 also includes a user feedback closed-loop unit, which is configured to: collect user feedback on the effect of the personalized adjustment suggestions pushed; associate the effect evaluation feedback with the feature sequence and early warning results of the corresponding time period to construct a feedback dataset; and based on the feedback dataset, incrementally learn the artificial intelligence model in the AI ​​fusion early warning module 130 through an online gradient descent algorithm to dynamically optimize the model parameters.

[0081] (1) Feedback collection: Users evaluate the effectiveness of treatment suggestions (effective / ineffective / average) through the feedback entry in the mobile app. The system constructs a feature-label pair feedback dataset by combining the feedback results with the feature sequences and warning results of the corresponding time period.

[0082] (2) Incremental learning: The system triggers fine-tuning of model parameters every 10 feedback data points. Incremental learning is achieved through online gradient descent (OGD). The parameters of the last layer of the model are updated in a small batch (batch size=1) using the feedback dataset. The learning rate is set to 1 / 10 of the initial training learning rate to avoid catastrophic forgetting of the model and to dynamically optimize the accuracy of early warning and the adaptability of conditioning suggestions.

[0083] Step S5: System Workflow Integration (1) Initial configuration: Users input their birth time through the mobile app, and the cloud completes the conversion of the sexagenary cycle, the body constitution code, the calculation of annual luck parameters and the generation of long-term trend prompts, and synchronizes the relevant static features and text prompts to the mobile device.

[0084] (2) Real-time data acquisition and transmission: The watch sensor continuously collects physiological data and wrist temperature and humidity data, which are then preprocessed and uploaded to the mobile terminal in real time via BLE.

[0085] (3) Feature fusion and reasoning: After receiving data on the mobile device, it is fused with the stored static features, input into the lightweight TCN model, and the short-term real-time warning result is calculated (long-term trend prompts directly call the stored text).

[0086] (4) Interactive feedback: The mobile app displays long-term trend prompts and short-term warnings, and pushes treatment suggestions; users provide feedback on the treatment effect through their mobile phones, and the system dynamically calibrates the model through incremental learning.

[0087] (5) Regular updates: At the beginning of each year, the annual luck parameters and long-term trend prompts are recalculated in the cloud and synchronized to the mobile device; the model parameters are updated weekly through the mobile app background to ensure performance optimization.

[0088] To verify the performance advantages of the algorithm of this invention, a comparative experiment was conducted with traditional algorithms. Figure 5This paper presents a performance comparison of traditional Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), Temporal Convolutional Network-Genetic Algorithm (TCN-GA), Temporal Convolutional Network-Particle Swarm Optimization (TCN-PSO), and the Temporal Convolutional Network-Hybrid Discrete Particle Swarm Optimization (TCN-HDPSO) algorithm of this invention in a Traditional Chinese Medicine (TCM) health early warning task, showing the change in early warning accuracy as a function of training epochs. All algorithms exhibit the typical model training characteristic of gradually increasing accuracy with increasing training epochs before converging. In the initial stage (epoch 0), due to the lack of effective parameter learning, the accuracy of each algorithm is concentrated between 62% and 68%. Within the initial training interval, performance differences are relatively small. However, as training progresses, performance differentiation among algorithms gradually becomes apparent. The mid-training period (20-100 epochs) is a critical stage for improving the accuracy of each algorithm. LSTM suffers from the gradient vanishing problem during sequence modeling, limiting its parameter learning efficiency and significantly slowing down its accuracy increase rate. TCN, with its advantages in local feature extraction and parallel computing of convolutional structures, achieves a better increase rate than LSTM. TCN-GA, TCN-PSO, and TCN-HDPSO, which introduce intelligent optimization algorithms, effectively avoid the blindness of manually setting hyperparameters by adaptively adjusting the model hyperparameters through optimization algorithms, resulting in significantly higher model parameter learning efficiency. The accuracy improvement is significantly higher than that of the basic TCN. The TCN-HDPSO proposed in this invention combines the global search advantage of particle swarm optimization with the local fine-tuning capability of discrete optimization, resulting in superior hyperparameter optimization. During this stage, its accuracy improvement rate consistently leads all compared algorithms. In the later stages of training (100-200 rounds), all algorithms gradually enter the convergence phase, and their performance tends to stabilize. LSTM ultimately converges to 75.6%, the lowest among all algorithms, while the basic TCN converges to 80.0%. The convergence accuracies of TCN-GA and TCN-PSO are 83.2% and 85.5%, respectively, while TCN-HDPSO... DPSO ultimately converged to 88.1%, a result that fully verifies that TCN-HDPSO effectively solves the problem of single intelligent optimization algorithms easily getting trapped in local optima during the search process. It has a better effect in model hyperparameter optimization. At the same time, combined with the feature extraction advantages of TCN, the model has stronger feature learning and risk identification capabilities when processing time-series health data such as heart rate and blood oxygen collected by the watch. Ultimately, it shows more accurate performance in health warning tasks, and can provide more reliable algorithm support for real-time health risk warning on mobile devices. It also proves the effectiveness and superiority of the TCN-HDPSO proposed in this invention in the TCM health warning scenario.

[0089] In summary, the AI-powered TCM smartwatch health warning system integrating Five Elements and Six Qi with environmental perception provided in the above embodiments of this invention innovatively adopts a collaborative mode of "watch data collection + mobile phone display." The watch only undertakes data collection and transmission functions, significantly reducing its computing power and battery life pressure. Simultaneously, it unifies the encoding and quantification logic of constitution characteristics, clarifies the dimensional alignment method for multi-dimensional features, and solves the key technical problems of the modern application of TCM theory and the lack of rigorous feature fusion logic, achieving a constitution quantification consistency of 82%. Employing a lightweight improved TCN model and the HDPSO hyperparameter optimization algorithm proposed in this invention, a complete three-stage optimization process is used to achieve efficient TCN hyperparameter search. After the model is deployed on the mobile phone, the inference latency is ≤65ms. While ensuring a short-term real-time warning accuracy of ≥85%, it also considers prediction accuracy and terminal adaptability. Clear warning threshold and risk level mapping rules further enhance the completeness of the solution. A three-dimensional service architecture of "long-term trend indication + short-term instant alert + closed-loop feedback" is constructed, with all interactive functions centralized on the mobile device. Leveraging the large screen and notification mechanism of the mobile phone, the display of long-term trends (3-6 months in advance) and short-term alerts (minute / hourly) is clearer and more timely. Through online gradient descent (OGD) and incremental learning, the adaptability of treatment suggestions is improved by 40%, significantly increasing user participation in health management. Clinical validation shows that the system's trend indication accuracy for seasonal diseases reaches 78%, the instant alert response time for acute physiological abnormalities is ≤1 minute, and the false alarm rate is controlled within 5%, providing intelligent and quantifiable tool support for traditional Chinese medicine's "prevention of disease." The system is clearly positioned as a "health auxiliary early warning tool," not a medical diagnostic device. Data collection and processing comply with the Personal Information Protection Law and medical data compliance requirements. Encrypted transmission is used between the watch and the mobile phone to ensure user privacy and security.

[0090] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0091] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0092] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A health early warning system for an AI-powered TCM smartwatch integrating Five Elements and Six Qi theory with environmental perception, characterized in that: include: The data acquisition module is deployed on the smartwatch, the Five Elements and Six Qi analysis module is deployed on the cloud and mobile phone, and the AI ​​fusion early warning module and human-computer interaction module are deployed on the mobile phone; The data acquisition module is used to collect the user's heart rate, blood oxygen saturation, skin conductivity physiological data, and wrist skin surface temperature and humidity data in real time. After data preprocessing, it outputs standardized physiological feature vectors and temperature and humidity time series data and uploads them to the mobile phone. The Five Elements and Six Qi Analysis Module includes an initialization calculation unit located in the cloud and a real-time fusion unit located on the mobile device. The initialization calculation unit performs Heavenly Stem-Branch conversion and constitution coding based on the birth time input by the user, generates constitution characteristics, calculates annual luck parameters and constitution-luck basic interaction values, and generates long-term trend prompts. The real-time fusion unit receives time-series temperature and humidity data uploaded from the smartwatch and combines it with macro-meteorological data and annual weather parameters sent from the cloud to generate comprehensive environmental weather characteristics. The AI ​​fusion early warning module is used to receive and fuse the standardized physiological feature vector, temperature and humidity time series data, physical characteristics and environmental and weather comprehensive features to form a fusion feature sequence, and input it into a pre-trained lightweight improved temporal convolutional network model to output short-term real-time early warning results. The human-computer interaction module is used to display the long-term trend prompts and short-term real-time warning results, and push personalized conditioning suggestions based on the preset TCM conditioning rule engine.

2. The system according to claim 1, characterized in that, The initialization calculation unit is configured to: The system converts the birth time into the Heavenly Stems and Earthly Branches based on the user's input, and generates physical characteristics represented by five-dimensional binary one-hot encoding based on the preset Five Elements mapping table. The physical imbalance index is calculated based on the physical characteristics. The physical imbalance index is used to characterize the degree of imbalance of the user in the five elements. The calculation method is as follows: assign weight coefficients obtained by optimizing through the attention mechanism to each of the five physical dimensions, and calculate the sum of the weighted absolute differences between the actual encoded value and the balance reference value of each dimension. Calculate the annual fortune parameters based on the sexagenary cycle of the current year; The basic interaction value between constitution and luck is calculated based on the physical characteristics and the annual luck parameters; and long-term trend prompts are generated based on a pre-set constitution-luck-seasonal mapping knowledge base.

3. The system according to claim 2, characterized in that, The AI ​​fusion early warning module includes a feature fusion layer, which projects the standardized physiological feature vector, temperature and humidity time series data, physical characteristics and environmental luck comprehensive features into a unified embedding space through a 1×1 convolutional layer and then splices them in time to generate the fused feature sequence.

4. The system according to claim 1, characterized in that, The hyperparameters of the lightweight improved temporal convolutional network model are obtained through a hybrid discrete particle swarm optimization algorithm. The hybrid discrete particle swarm optimization algorithm achieves hyperparameter optimization through a three-stage process, including: a stage of generating an initial particle swarm based on the improved theorem, a stage of global exploration through discrete particle swarm optimization, and a stage of local development combined with hill climbing. The global exploration stage and the local development stage are repeated until the preset maximum number of iterations is reached, or the global optimal fitness is not improved for 5 consecutive rounds, and the final global optimal hyperparameters are output.

5. The system according to claim 4, characterized in that, The objective function of the hybrid discrete particle swarm optimization algorithm is: Where X is the hyperparameter vector to be optimized. The health warning accuracy of the model on the validation set. This refers to the inference latency of the model on mobile devices.

6. The system according to claim 4, characterized in that, In the stage of generating the initial particle swarm based on the improved theorem, the hybrid discrete particle swarm optimization algorithm is configured to: combine the hyperparameter search space of the lightweight improved temporal convolutional network model to generate at least one heuristic particle, the hyperparameter values ​​of which are determined based on preset empirical rules, and the remaining particles are generated randomly to form the initial particle swarm.

7. The system according to claim 4, characterized in that, In the phase of global exploration via discrete particle swarm optimization, the hybrid discrete particle swarm optimization algorithm updates the particle velocity and position in the following manner: The velocity update is based on the particle's inertial velocity, the difference between the particle's own historical best position and its current position, and the difference between the group's global best position and its current position, and is fused according to a preset weight coefficient to determine the updated velocity vector. Position update: Adjust the particle's current position vector according to the updated velocity vector, and generate new hyperparameter candidate solutions based on the adjusted position vector, while ensuring that the values ​​of each hyperparameter in the newly generated hyperparameter candidate solutions conform to the predefined search space constraints.

8. The system according to claim 4, characterized in that, In the stage of local development combined with hill climbing, the hybrid discrete particle swarm optimization algorithm is configured to: randomly trigger a local search for the current hyperparameter candidate solution with a preset probability; when a local search is triggered, fix some hyperparameters in the current hyperparameter candidate solution, fine-tune the remaining hyperparameters, and calculate the fitness value of the new solution. If the new solution is better than the current solution, it is replaced; otherwise, the original solution is retained.

9. The system according to claim 1, characterized in that, The short-term immediate warning result includes a risk level, which is determined by comparing the warning probability output by the lightweight improved temporal convolutional network model with a preset dynamic threshold. Different disease types correspond to different dynamic thresholds, and the dynamic thresholds are determined separately for different disease types by performing receiver operating characteristic curve analysis on clinical data.

10. The system according to claim 1, characterized in that, The human-computer interaction module further includes a user feedback closed-loop unit, which is configured to: Collect user feedback on the effectiveness of the personalized treatment suggestions pushed to them; The effect evaluation feedback is associated with the feature sequences and early warning results of the corresponding time period to construct a feedback dataset; Based on the feedback dataset, the artificial intelligence model in the AI ​​fusion early warning module is incrementally learned using an online gradient descent algorithm to dynamically optimize the model parameters.

Images