A health-care index prediction method based on a time series model and application thereof

Abstract

The present application relates to the fields of artificial intelligence, big data analysis and time series prediction, and discloses a health and wellness index prediction method and application based on a time series model, a multi-dimensional historical observation vector is constructed from a data source, a multi-dimensional state space model and a scenario classifier are trained, and a plurality of groups of multi-dimensional weight vectors corresponding to health and wellness weather scenarios defined by experts are stored. When predicting, a multi-dimensional factor prediction vector is output by driving the model, a dynamic weight vector is calculated in combination with the scenario probability distribution of the scenario classifier, and the health and wellness index prediction value is generated by the dot product of the two. An adaptive updating and evolution mechanism is introduced, a prediction residual is calculated through real observation values, a multi-dimensional statistical process control model is driven to perform Hotelling T square statistic monitoring. If an out-of-scenario anomaly triggers a new scenario discovery workflow, clustering of abnormal data identifies a digital signature of the new scenario, and experts inject new weights to activate the retraining of the scenario classifier, scenario adaptation and dynamic weighting are realized, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, big data analysis, and time series forecasting, specifically to a method and application for predicting a health and wellness index based on a time series model. Background Technology

[0002] With the popularization of health concepts and the increasing trend of population aging, the health and wellness industry is becoming increasingly integrated with diverse information such as meteorology and environment, leading to a growing demand for accurate predictions of the health and wellness index. The health and wellness index is usually derived from a comprehensive assessment of multiple environmental or physiological factors, and the weight and manner in which these factors affect human health and wellness are dynamically adjusted according to changes in the external environmental context.

[0003] However, existing methods for predicting health and wellness indices generally face challenges. On the one hand, many methods use fixed model parameters or static weights to evaluate different factors, making it difficult to effectively capture the dynamic changes in the impact of complex and ever-changing external environmental scenarios on health and wellness factors. When the environmental scenario changes, the prediction accuracy and adaptability of static models often decrease, leading to deviations between the prediction results and the actual situation. On the other hand, existing prediction systems typically lack real-time monitoring mechanisms for model operation, failing to promptly detect discrepancies between their prediction logic and the real environmental scenario. For example, when the system encounters out-of-context anomalies not present in its training samples, the model performance may quietly degrade.

[0004] Furthermore, existing technologies have shortcomings in the adaptive updating and evolution of model knowledge. Once a model is deployed, its underlying knowledge system (such as scenario definitions and weight parameters) is often fixed. When new environmental scenarios emerge or the characteristics of existing scenarios evolve, traditional methods struggle to effectively and automatically detect these changes and integrate them into the model. This prevents the model from continuously learning and optimizing, thus limiting the long-term stability and robustness of the prediction system and restricting its ability to capture emerging health and wellness influencing factors. Therefore, how to construct a health and wellness index prediction system that can dynamically adapt to changing scenarios and possess self-monitoring and adaptive evolution capabilities is a problem that needs to be solved in the current technological field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and application for predicting health and wellness index based on time series models. This method solves the problem that it is difficult to accurately capture complex and ever-changing external environmental scenarios and their dynamic changes in impact on health and wellness factors, which leads to limitations in the adaptability and robustness of the prediction model. Furthermore, it lacks an effective mechanism to adaptively discover and incorporate new environmental scenarios and update model parameters, thus limiting the long-term stable operation capability of the prediction system.

[0006] The first aspect of this invention provides a method for predicting a health and wellness index based on a time series model. This method first obtains historical time series data of heterogeneous health and wellness factors from one or more external data sources. Through data cleaning, missing value imputation, and standardization, and by setting a uniform time frequency, resampling and timestamp alignment are performed on multiple factor sequences to construct a multidimensional historical observation vector.

[0007] In the modeling phase, this method trains a multidimensional state-space model, which adopts the form of a structured time series model. It decomposes the state vector into a stack of local linear trend components, periodic seasonal components, and vector autoregressive components to capture the temporal evolution of the data. Simultaneously, a labeled training set is constructed based on historical observation vectors and their corresponding expert-annotated scenarios to train a scenario classifier. This scenario classifier is constructed using a Gaussian mixture model, estimating the Gaussian component parameters (including prior probabilities, multidimensional mean vectors, and multidimensional covariance matrices) corresponding to each initial health and wellness meteorological scenario using the expectation-maximization algorithm. This allows it to output the probability distribution of observed data belonging to each initial health and wellness meteorological scenario. Furthermore, the method utilizes the analytic hierarchy process (AHP) to process the judgment matrices defined by experts for each scenario. After passing a consistency check, the normalized eigenvector corresponding to the largest eigenvalue is calculated, thereby storing multiple sets of multidimensional weight vectors corresponding to multiple initial health and wellness meteorological scenarios.

[0008] In the prediction phase, this method drives a multidimensional state-space model to output a multidimensional factor prediction vector for the prediction time. This prediction vector is then input into a scenario classifier to obtain the probability distribution of various health and wellness meteorological scenarios at the current prediction time. Subsequently, the method performs weighted calculations on multiple stored multidimensional weight vectors based on this probability distribution to generate a dynamic weight vector for the prediction time. Finally, by performing a dot product operation between the multidimensional factor prediction vector and the dynamic weight vector, the predicted value of the health and wellness index is generated and output. This mechanism enables dynamic adjustment of weights based on the probability of the current environmental scenario, improving the adaptability of the prediction results to complex and ever-changing environments.

[0009] Furthermore, to endow the system with self-evolutionary capabilities, the method also includes a closed-loop monitoring and update process. The method acquires the current real observations, calculates the true prediction residuals of the multidimensional state-space model, and drives the multidimensional statistical process control model to update the statistical vector of the multidimensional exponentially weighted moving average model. Based on this, the Hotelling T-squared statistic is calculated and continuously compared with a set anomaly control upper limit. When the multidimensional state-space model is determined to encounter an out-of-scenario anomaly, a new scenario discovery workflow is triggered.

[0010] The new scenario discovery workflow includes capturing observational data within anomaly time windows, applying density-based clustering algorithms to identify new scenario clusters, and calculating their multidimensional digital signatures. The system stores the new scenarios and their signatures in an AHP weight library and marks them as pending review, subsequently triggering a human-machine collaborative interface. After receiving new AHP weights injected by domain experts for the new scenarios and confirming their activation, the system triggers retraining of the scenario classifier, enabling it to recognize all scenarios, including the new scenarios, thereby achieving continuous expansion and adaptive iteration of the model knowledge base.

[0011] A second aspect of this invention provides a health and wellness index prediction application based on a time series model. This application includes one or more processors configured to execute the method described in the first aspect. Through a combination of hardware and software, this application achieves fully automated processing from data acquisition, model training, dynamic prediction to anomaly monitoring and adaptive updates, making it suitable for health and wellness index monitoring and service scenarios requiring high accuracy and dynamic adaptation to environmental changes.

[0012] This invention provides a method and application for predicting the health and wellness index based on a time series model. It has the following beneficial effects: 1. This invention combines a multidimensional state-space model with a scenario classifier based on a Gaussian mixture model. Instead of relying on a single, fixed evaluation weight, it calculates the probability distribution of each health and wellness meteorological scenario based on the factor characteristics at the prediction time, and generates a dynamic weight vector accordingly. This mechanism can capture the nonlinear changes in the degree of impact of each evaluation factor on human health under different climatic or environmental backgrounds, ensuring that the prediction model maintains a high degree of matching with the current environmental characteristics under variable meteorological conditions, thus solving the problem of inaccurate predictions in traditional static weight models during sudden environmental changes.

[0013] 2. This invention introduces a multidimensional statistical process control model. By calculating the true prediction residual of the multidimensional state-space model and updating the multidimensional exponentially weighted moving average statistic, the Hotelling T-squared statistic is used to monitor the model status in real time. An anomaly control upper limit is set, which can make timely judgments and trigger subsequent processing procedures when the model encounters unknown out-of-context data or experiences performance drift. This effectively prevents the model from silently failing in unknown environments and improves the robustness and risk perception capability of the system.

[0014] 3. This invention designs a new scenario discovery workflow and a human-machine collaboration interface. It can automatically extract new scenario clusters from anomalous data using a density-based clustering algorithm, and inject new hierarchical analysis weights based on domain expert review, thereby triggering retraining of the classifier. This closed-loop design combines the discovery capabilities of data mining with the logical judgment of experts, enabling the system to continuously identify and incorporate new health and wellness meteorological patterns. This achieves autonomous updating and iterative evolution of the model knowledge base, extending the system's lifespan and scope of application. Attached Figure Description

[0015] Figure 1 This is a system framework diagram of the present invention.

[0016] The module includes: 10. Data preparation and management module; 20. Offline modeling and initialization module; 30. Online prediction and integration module; and 40. Adaptive update and evolution module. Detailed Implementation

[0017] The technical solutions in 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.

[0018] See attached document Figure 1 Embodiments of the present invention can be executed by one or more computing devices. A computing device includes one or more processors, a memory, a non-volatile storage medium for storing computer program instructions, and a network interface for communicating with an external data source or user terminal.

[0019] This invention provides a dynamic prediction and adaptive update system for health and wellness index. The system may include: a data preparation and management module 10; an offline modeling and initialization module 20; an online prediction and integration module 30; and an adaptive update and evolution module 40.

[0020] Data preparation and management module 10 is used to obtain data from one or more external data sources (such as meteorological databases and environmental monitoring databases). Historical time series data of heterogeneous health and wellness factors. These factors include, for example, temperature, relative humidity, air quality index, precipitation, and wind speed.

[0021] The data preparation and management module 10 is also used to perform preprocessing operations on historical time series data. These preprocessing operations include data cleaning, missing value imputation, and data standardization to generate data in a uniform format. dimensional historical observation vector : ; in: Indicates a point in time; The total number of factors; For the first Each factor in The observed value at time; This represents the transpose of a vector.

[0022] The offline modeling and initialization module 20 is connected to the data preparation and management module 10. The offline modeling and initialization module 20 is used to perform a series of initialization modeling operations when the system is first launched or during periodic calibration.

[0023] In one embodiment, the offline modeling and initialization module 20 is used to: receive An initial health and wellness meteorological scenario defined by domain experts. ,For example Hot and humid Cold and dry conditions, etc. (Targeting...) Each of the scenarios Receive the corresponding analytic hierarchy process (AHP) judgment matrix, calculate and store its unique value. dimensional weight vector : ; in: The initial total number of scenarios; For the context Next The weight coefficients of each factor. Group Scenario Weights It is stored in a weight library of the Analytic Hierarchy Process (AHP) for later use.

[0024] The offline modeling and initialization module 20 is also used for: using dimensional historical observation vector Train a multidimensional state-space model. The multidimensional state-space model includes observation equations and state equations: Observation equation: ; Equations of state: ; in: yes The unobservable state vector, The dimension of the state variable; yes The observation matrix; yes The state transition matrix; yes The observed noise vector follows distributed; yes The state noise vector follows distributed. and These are the covariance matrices of observation noise and state noise, respectively. The offline modeling and initialization module 20 initializes the model parameters using Kalman filtering and maximum likelihood estimation. Make an estimate.

[0025] The offline modeling and initialization module 20 is also used for: based on history and the corresponding expert annotation scenarios Train a context classifier (e.g., Gaussian mixture model). take over The factor vector is taken as input, and its corresponding element is output. Probability distribution of initial scenarios : ; in: This indicates that the input vector belongs to the scenario. The probability, and .

[0026] The offline modeling and initialization module 20 is also used to: predict residuals in one step on the training set using the M-SSM model. A multidimensional statistical process control model is initialized. In one embodiment, the multidimensional statistical process control model is a multidimensional exponentially weighted moving average model, used to calculate the statistical vector of the multidimensional exponentially weighted moving average model. : ; in: It is between The smoothing factor between them. And based on Based on the historical distribution, an upper limit for anomaly control is defined. .

[0027] The online prediction and integration module 30 and the offline modeling and initialization module 20 (specifically the M-SSM model and the scene classifier) It connects to the AHP weight library. The online prediction and integration module 30 is used to perform online, real-time prediction of the health and wellness index.

[0028] In one embodiment, the online prediction and integration module 30 is used to: receive Time-based prediction requests To predict the step size. The driving force of the M-SSM model is based on... The state at any given moment, output Moment Dimensional factor prediction vector .Will Input to the context classifier , obtain Probability distribution of a known scenario .according to and AHP weight library Group weights ,calculate Dynamic weight vector at time step : ; Will Dimensional factor prediction vector and 3D dynamic weight vector Perform a dot product operation to generate and output the final predicted value of the health and wellness index. : ; The module consists of 40 modules: Adaptive Update and Evolution, 10 modules: Data Preparation and Management, and 20 modules: Offline Modeling and Initialization (specifically, the M-SSM model, AHP weight library, and scenario classifier). Both the data and control connections are established in the multidimensional statistical process control model.

[0029] Adaptive Update and Evolution Module 40 is used for: Real observations are obtained from the data preparation and management module 10 at all times. Calculate the true prediction residuals of M-SSM. The multidimensional statistical process control model is updated to update the statistical vector of the multidimensional exponentially weighted moving average model. and Statistic. Continuous comparison. Statistics and When the M-SSM model is determined to encounter an out-of-context anomaly (e.g., continuous...), Time points , (The value is a preset integer), triggering a new scenario discovery workflow. The new scenario discovery workflow is used for: Capture an abnormal time window Observational data within .

[0030] right Data is applied using density-based clustering algorithms to identify... indivual( New scenario clusters .for Calculate one 3D digital signatures (e.g., centroid vectors) are defined as a new scenario. .

[0031] The adaptive update and evolution module 40 is also used to: adapt new scenarios The digital signature of the data is stored in the AHP weight library, and its status is marked as pending review. A human-computer collaboration interface (e.g., a management backend interface) is then triggered for review by domain experts. And inject the corresponding new AHP weights into it. .

[0032] exist Once confirmed and activated, the Adaptive Update and Evolution Module 40 triggers an update to the context classifier. Retraining, Able to identify including Including This allows for the adaptive evolution of the system model by considering various scenarios.

[0033] See attached document Figure 1 In this embodiment of the invention, the operations of Phase One (system initialization and offline modeling) are jointly executed by the data preparation and management module 10 and the offline modeling and initialization module 20. First, the data preparation and management module 10 performs data preparation and factor system construction, which provides the necessary, formatted historical data input for the subsequent offline modeling and initialization module 20.

[0034] The selection and acquisition of health and wellness factors. The data preparation and management module 10 acquires data from one or more heterogeneous data sources through a pre-defined data interface (e.g., database connection or network API). Historical time series data of various health and wellness factors. Data sources include, for example, the databases of the National Meteorological Information Center and urban environmental monitoring stations.

[0035] In one specific embodiment, The health and wellness factors are divided into at least two categories: Category 1, meteorological factors: including but not limited to temperature (degrees Celsius), relative humidity (%), wind speed (m / s), precipitation (mm), and sunshine duration (h). Category 2, environmental quality factors: including but not limited to air quality index and PM2.5 concentration (%). Nitrogen dioxide concentration (unit: ) and ozone concentration (unit: ).

[0036] The data preparation and management module 10 stores the acquired, timestamped raw data into the system's data storage area (e.g., a time-series database or relational database table) for subsequent processing. Data cleaning, imputation, and standardization are performed. Due to the heterogeneous source of the raw data and the existence of data quality issues, the data preparation and management module 10 executes a series of preprocessing operations.

[0037] Cleaning operation: For a single factor sequence, outliers are identified and processed. In one embodiment, a statistical thresholding method is used, for example... The three-standard-deviation criterion: Calculate the mean and standard deviation over a moving time window (e.g., 24 hours), mark data points that exceed the mean plus or minus three standard deviations as outliers, and treat them as missing values.

[0038] Imputation operation: Time series imputation is performed for missing values ​​generated during the cleaning operation or inherent missing values ​​in the original data. In one embodiment, a segmented imputation strategy is adopted according to the length of the missing time window: for short-term (e.g., no more than 3 consecutive time points missing) missing data, time-based cubic spline interpolation is used for filling; for long-term missing data, trend and seasonality terms based on seasonal decomposition are used for filling.

[0039] Standardization: To eliminate differences in units and numerical ranges among different factors, the following procedures are followed: Each of the factor sequences Perform Z-score normalization calculations individually. For any factor sequence Its standardized value The calculation is as follows: ; in: yes The original (or interpolated) observation at time; It is this factor The mean over the entire historical training set; It is this factor The standard deviation over the entire historical training set. and It is stored as a standardized parameter and used to perform the same standardized transformation on new data in the subsequent online prediction stage.

[0040] Construction of historical observation vectors. This step addresses the issue of inconsistent data collection frequencies for different factors (e.g., temperature data at the minute level, AQI data at the hour level).

[0041] The data preparation and management module 10 first sets a uniform time frequency (e.g., 1 hour). For data collected at a frequency higher than this uniform frequency (e.g., wind speed collected every 10 minutes), downsampling is performed within a time window (e.g., calculating the arithmetic mean of the data collected within the hour as the representative value for that hour). For data collected at a frequency lower than this uniform frequency (e.g., data collected daily), upsampling is performed (e.g., using forward padding or nearest neighbor interpolation to pad the daily data onto the 24-hour interval of that day).

[0042] In all After all factor sequences are resampled to the unified time frequency, the data preparation and management module 10 timestamps them. Align all Standardized factor values , build dimensional historical observation vector : ; in: This represents the transpose of a vector. The constructed historical time series is output to the offline modeling and initialization module 20 as a training source for the M-SSM model and the scene classifier. Input data.

[0043] See attached document Figure 1 The offline modeling and initialization module 20 is used to build an initial AHP scenario and weight library. This step aims to transform the differences in importance of different health and wellness factors under different meteorological conditions (i.e., nonlinear scenario dependence) in the health and wellness index assessment into a set of structured expert knowledge parameters that can be called by the model. This construction process is completed through human-computer interaction (e.g., through a dedicated configuration interface) and includes the following steps.

[0044] initial The definition of a health and wellness meteorological scenario. The offline modeling and initialization module 20 first receives the definition provided by domain experts (e.g., health and wellness or meteorology experts). indivual( , Initial, well-defined health and wellness meteorological scenarios (integer format) In one specific embodiment, This scenario is based on the selection and acquisition of health and wellness factors in the following steps. A semantic summary of different combinations of health and wellness factors. For example, At that time, the scenario can be defined as: Hot and humid; Cold and dry; Severe pollution; Warm and comfortable.

[0045] Construction and weighting of AHP decision matrices for each scenario Calculation. For Each of the scenarios The offline modeling and initialization modules 20 both guide the expert group to independently conduct hierarchical AHP evaluations. This evaluation process involves: scenario-based... , as the target layer, with Health and wellness factors This is the criterion level. The expert group used a 1-9 scale to evaluate... Each factor in The relative importance in each context is compared pairwise. For example, in In hot and humid conditions, the expert group determined the temperature factor. Compared to the AQI factor ( The relative importance of ) is 5 (significantly important); while Under severe pollution scenarios, the expert panel determined the AQI factor ( Relative to temperature factor ( The relative importance of ) is 7 (strongly important).

[0046] This step is... Each of the scenarios Generate a Judgment matrix : ; in: Indicates the situation Below, factor Relative factor The importance scale value, and , Subsequently, the offline modeling and initialization module 20 performs a check on each judgment matrix. Perform a consistency check. This check includes: calculate The largest eigenvalue Calculate the consistency index. Based on the number of factors Find the preset average random consistency index Calculate the consistency ratio. .

[0047] Offline modeling and initialization module 20 judgment Is it less than a preset consistency threshold (e.g., 0.1)? If the value is greater than this threshold, then the decision matrix is ​​determined. There is a lack of consistency, and the expert panel is advised to readjust its approach. The judgment matrix of the scenario; if If the value is less than or equal to the threshold, then the decision matrix is ​​determined. Passed the consistency test.

[0048] For matrices that pass the consistency test The offline modeling and initialization module 20 calculates its maximum eigenvalue. The corresponding normalized feature vector is defined as the context. Below dimensional static weight vector : ; in: This represents the transpose of a vector, and .

[0049] The database storage implementation of the AHP weight library. The offline modeling and initialization module 20 will handle all... Group passes the test: scenario-weighted pair The weights are stored in the system's data storage area (e.g., in an AHP weight table of a relational database). This weight library provides the baseline parameters for dynamic weight calculation for the subsequent online prediction and integration module 30.

[0050] See attached document Figure 1 The offline modeling and initialization module 20, after completing the AHP weight library construction, further executes the training of the M-SSM model. The purpose of this step is to build a model capable of capturing... The dynamic interaction relationships between various health and wellness factors, and the ability to simultaneously monitor them. A statistical model that uses multiple factors for joint prediction.

[0051] The offline modeling and initialization module 20 receives data from the data preparation and management module 10. dimensional historical observation vector The sequences are used as training data for M-SSM. In one specific embodiment, M-SSM is constructed using the form of a structured time series model. In this form, dimensional state vector It is explicitly decomposed into multiple unobservable components that have actual physical or statistical significance.

[0052] For example, state vector It can be configured as a stack of multiple components. .

[0053] in: It is a description A component of a common or individual local linear trend of factors; It is a description A seasonal component with a periodicity of one factor (e.g., a 24-hour daily cycle or a 7-day weekly cycle); It is a vector autoregressive component used to capture data after removing trends and seasonality. The dynamic correlation between the residuals of each factor. Correspondingly, the state transition matrix... and observation matrix It is also constructed in the form of a corresponding block matrix.

[0054] Training M-SSM involves training its parameter matrix (including...) and noise covariance matrix The unknown parameters (such as noise variance) are estimated. The offline modeling and initialization module 20 performs this estimation using a combination of Kalman filtering and maximum likelihood estimation. This training process specifically includes: initializing a set of parameters to be estimated for the M-SSM. Given this set of parameters, the offline modeling and initialization module 20 applies a Kalman filter to estimate the unknown parameters (such as noise variance). dimensional historical observation vector sequence( The total length of historical data is used for recursive processing.

[0055] At each step of the recursive process Kalman filter is based on State estimation at time t and Observations at time Calculate the one-step prediction error as well as covariance matrix The offline modeling and initialization module 20 utilizes a Kalman filter in... The prediction error of each step output at time 1 and its covariance matrix Construct a (log) likelihood function This function The value is used to quantify the goodness of fit of the current parameter set to historical observation data.

[0056] The offline modeling and initialization module 20 employs a numerical optimization algorithm based on the log-likelihood function. Maximizing the value of the parameter is the optimization objective. The optimal estimate of the parameter to be estimated in M-SSM is iteratively searched and solved.

[0057] When the numerical optimization algorithm converges (e.g., the log-likelihood function in two consecutive iterations), The iteration terminates when the gain of the parameter value is less than a preset convergence threshold, or the change in the parameter estimate is less than a preset threshold. The parameter set obtained at this point is determined as the final training result of M-SSM. The offline modeling and initialization module 20 then uses this trained parameter matrix containing the optimal estimates ( It is stored in the system's data storage area as a trained M-SSM model instance for use by the online prediction and integration module 30.

[0058] See attached document Figure 1 The offline modeling and initialization module 20 executes the scene classifier after the AHP weight library is built and the M-SSM model is trained. The purpose of this step is to create a model that can learn from the input. The health and wellness factor vector automatically and probabilistically determines the health and wellness scenario to which the vector belongs, providing a basis for dynamic weight soft switching in the online prediction phase.

[0059] Training set construction. The offline modeling and initialization module 20 first constructs a training set for supervised learning. The data in this training set... This comes from the data preparation and management module 10. dimensional historical observation vector sequence( The labels of this training set This originates from the offline modeling and initialization module 20 in the initial... The steps in defining a health and wellness weather scenario receive... An initial scenario A quantitative definition. In one embodiment, each scenario It not only has semantic names (e.g., hot and humid), but also includes a set of quantitative boundary rules (e.g., temperature). And humidity ).

[0060] The offline modeling and initialization module 20 iterates through all historical observation vectors. ,application Group quantitative boundary rules, for each Assign a unique context label Thus, a system containing... A labeled training set of samples. Classifier training. The offline modeling and initialization module 20 uses the training set to train a Gaussian mixture model as a context classifier. .

[0061] In one specific embodiment, the classifier is constructed as follows: A mixture of Gaussian components, in which Values ​​and the number of initial scenarios in the AHP weight library Strict consistency. The offline modeling and initialization module 20 uses the expectation-maximization algorithm based on the training set. Estimate the parameters of the classifier. More specifically, for each of the K scenarios... The offline modeling and initialization module 20 uses all tags for Data subset Calculate the parameters of the Gaussian component for this scenario: The prior probability of this scenario (Right now ), based on this scenario The sample frequencies in the training set are determined. This scenario... dimensional mean vector ( (Vector). The scenario covariance matrix After training, the context classifier Depend on Group parameters Full definition. Context classifier The function is: when receiving a new dimensional factor vector (For example, prediction vectors from the online prediction module) When ), it calculates the vector using Bayes' theorem. Belongs to each scenario posterior probability .

[0062] ; in: That is, prior probability ; yes In the context The likelihood of the Gaussian component is calculated using the following formula: The probability density function of a 2D multivariate Gaussian distribution: ; in: It is the covariance matrix The determinant of; It is its inverse matrix; This represents the transpose of a vector.

[0063] Storage. The offline modeling and initialization module 20 stores this set of trained data, containing all... Group parameters Context classifier Instances are stored in the system's data storage area (e.g., a model file or database entry) for use by the online prediction and integration module 30 when performing online predictions. This classifier... Output It is the direct input for achieving dynamic weight synthesis.

[0064] See attached document Figure 1 The offline modeling and initialization module 20, after completing the training of the M-SSM model, further performs the initialization of the residual monitoring system. The purpose of this step is to establish a statistical benchmark for detecting when the predictions of the M-SSM model systematically deviate from its historical behavior pattern during subsequent online operation.

[0065] The offline modeling and initialization module 20 first obtains the data based on historical observation vectors during the M-SSM model training process. sequence( Generated from (total length of historical data) One-step prediction of residual sequence The residual The calculation is as follows: ; in: Is The actual observation vector at any given time; The M-SSM model is based on Time information The predicted vector at time step [time], under a well-fitted steady state, is [the value of the vector]. The sequence should be approximately Gaussian white noise.

[0066] The offline modeling and initialization module 20 uses this residual sequence. To establish a multidimensional statistical process control model. In one specific embodiment, the multidimensional statistical process control model adopts a multidimensional exponentially weighted moving average model. The offline modeling and initialization module 20 first sets a smoothing factor. ( ), and then based on Calculate the statistical vector of the multidimensional exponentially weighted moving average model for sequence calculation : ; in: Initialized to Zero-dimensional vector.

[0067] In order to Vie The vector is transformed into a single monitoring statistic, which is first calculated by the offline modeling and initialization module 20. Vector in steady state covariance matrix In one embodiment, The calculation method is as follows: ; in: yes Dimensional residual The covariance matrix at steady state. For ease of calculation, when When it is large enough, the covariance matrix It can be approximated as: ; It can be trained by the M-SSM model or by The sample covariance matrix of the sequence is estimated. Subsequently, the offline modeling and initialization module 20 is based on... and its covariance matrix ,calculate Hotelling's Moments Statistic : ; in: Indicates vector transpose; yes The inverse matrix.

[0068] The ultimate goal of this step is to Set a statistical control upper limit for the statistic. Exceeding this threshold indicates that the model is out of control. The offline modeling and initialization module 20 is based on... The theoretical distribution under steady state (e.g., when) When it is large, its distribution can be derived from Distribution or (Approximate distribution), set a predetermined Type I error rate (e.g.) ).

[0069] In one specific embodiment, this control upper limit according to The asymptotic distribution is used to determine this. For example, based on a chosen Type I error rate. , Set as this distribution Quantiles. In another embodiment, this can be achieved through Monte Carlo simulation or based on... indivual The empirical quantiles of the sample are used to calibrate the The offline modeling and initialization module 20 will complete the calibration of the smoothing factor. Covariance matrix (or its inverse matrix) and control limits The values ​​are stored in the system's data storage area. These parameters constitute a complete, initialized residual monitoring system for use by the adaptive update and evolution module 40 during the online phase.

[0070] See attached document Figure 1 This stage is the core workflow of the health and wellness index dynamic prediction and adaptive update system, which provides prediction services to users or downstream applications. This process is executed by the online prediction and integration module 30, which calls the model and parameter library built by the offline modeling and initialization module 20.

[0071] Multidimensional factor vector prediction. This step is used to generate the future. Moment The predicted value of the Vicon nutrition factor vector. The online prediction and integration module 30 first receives a prediction request that specifies the target prediction time. , To predict step size (e.g.) This indicates a prediction for the next 24 hours. To perform the prediction, the online prediction and integration module 30 first needs to obtain the current... The system state at time t. The online prediction and ensemble module 30 utilizes the M-SSM model parameters trained in the M-SSM model training step, as well as the Kalman filter initialized in the M-SSM model training step, based on the up to t. Time (including) All available observation vectors at time (time) Perform a filtering operation to calculate the current... Filtered state estimation vector at time 1 .

[0072] In obtaining the current State estimation vector at time 1 Then, the online prediction and integration module 30 executes. Step-by-step prediction. This prediction process is achieved by recursively applying the state equations of the M-SSM. This recursive process does not depend on For observations after time step 1, only the trained state transition matrix is ​​used. : ; ; ... ; The above recursive calculation yields State prediction vector at time step . It is the state transition matrix estimated and stored offline during the M-SSM model training step. After calculating... State prediction vector at time step Subsequently, the online prediction and integration module 30 applies the observation equations of M-SSM to calculate... Moment Dimension factor observation prediction vector This calculation is based on observation equations and sets future... Observation noise at any time The expected value is 0: ; in: It is the observation matrix that is estimated and stored offline during the training step of the M-SSM model.

[0073] Dimensional factor prediction vector As the output of this step, it is transmitted to the next processing unit of the online prediction and integration module 30, namely dynamic scenario matching and weight generation.

[0074] See attached document Figure 1 Online prediction and integration module 30, in generating Factor prediction vector at time step Next, dynamic scenario matching and weight generation steps are performed. This step aims to solve the problem that fixed weights in traditional methods cannot adapt to changing weather environments. By calculating the membership degree of the predicted state in the preset scenario space, a smooth and non-linear dynamic configuration of the weight vector is achieved.

[0075] The prediction vector is input and its feature mapping is performed. The online prediction and ensemble module 30 calls the scenario classifier trained and stored in the offline phase. Example. The online prediction and integration module 30 will output the multidimensional factor vector prediction step. Dimensional factor prediction vector As input data, it is transmitted to .

[0076] In one specific embodiment, when the scenario classifier When using a Gaussian mixture model, this input process actually converts the predicted vector... Mapped to by In the feature space defined by Gaussian components. Module 30 is based on storage. Group Gaussian mixture model parameters (mean vector) Covariance Matrix First calculate Relative to each scenario The likelihood of ).

[0077] Calculation of scenario probability distribution. Scenario classifier. Based on the calculated likelihood and the stored prior probability Applying Bayesian inference computation Belongs to each preset scenario posterior probability This calculation process generates a dimensional probability distribution vector : ; in: In the future The weather conditions at that time belong to the first The probability confidence level for a typical scenario (e.g., hot and humid). This vector satisfies the normalization constraint: ; This probability distribution vector objectively quantifies the degree of transition or fuzzy boundary between various typical scenarios of the weather state at the prediction time. For example, a state transitioning from hot and humid to warm and comfortable produces a probability distribution of [0.6, 0.4, 0, 0][0.6, 0.4, 0, 0]. The dynamic weight vector is synthesized. The online prediction and integration module 30 retrieves the corresponding weights from the AHP weight library based on the probability distribution vector. Each base weight vector , ,…, Each All A dimensional vector represents a vector in a specific context. Down The importance of each health and wellness factor is ranked. Subsequently, the online prediction and integration module 30 performs a weighted linear combination operation to generate... The final dynamic weight vector at time step This operation enables soft switching between different scenario weights, avoiding sudden weight changes caused by hard thresholds for scenario determination. ; Expanded representation: ; The generated It is Dimensional vector: ; in: Indicates in Under the predicted weather conditions at that time, the first The dynamic contribution weights of each health and wellness factor to the final health and wellness index. Because... And each benchmark The sum of the elements is 1, therefore the synthesized element is... Automatically satisfying normalization conditions This dynamic weight vector is then passed to the integration and output steps of the wellness index.

[0078] See attached document Figure 1 The online prediction and integration module 30 completes the dynamic weight vector... After the data is generated, the integration and output of the wellness index is immediately performed. This step is the final stage of online prediction of the wellness index, aiming to combine multi-dimensional factor prediction values ​​with dynamic evaluation weights to generate a single wellness index with comprehensive indicative significance.

[0079] The online prediction and integration module 30 obtains the results from the multidimensional factor vector prediction steps of this module. Dimensional factor prediction vector And obtain from the dynamic scenario matching and weight generation steps 3D dynamic weight vector The online prediction and integration module has 30 pairs of these two. Performing a dot product operation, also known as an inner product operation, on a 3D vector. This operation will... Predicted values ​​of each health and wellness factor Its dynamic importance weight under this prediction scenario Multiply, then all The summation of product terms is performed. The formula for this integrated operation is as follows: ; in: Is The final health and wellness index prediction at any given time (a scalar). It is calculated in the dynamic scenario matching and weight generation steps. Transpose of a 3D dynamic weight vector; It is calculated in the multidimensional factor vector prediction step. Dimension factor prediction vector.

[0080] Expanding the formula into a summation form, the specific calculation process is as follows: ; in: Traverse from 1 to ( (Total number of factors); yes The first in Elements (i.e., factors) (Dynamic weights); yes The first in Each element.

[0081] The online prediction and integration module 30 will calculate the scalar values. This serves as the final result of the health and wellness index prediction. The results can be output to downstream applications via the system's data interface or stored in a prediction result table in the data storage area.

[0082] Phase 3: Adaptive Updates to the AHP Scenario Library See attached document Figure 1 This stage is one of the core functions of the health and wellness index dynamic prediction and adaptive update system, and is executed by the adaptive update and evolution module 40. This stage is continuously executed during the system's online operation, with the aim of monitoring the effectiveness of the prediction model and automatically detecting when the model encounters new scenarios that it has not learned in its offline training phase.

[0083] Real-time monitoring of residual flow. This step is the starting point for adaptive updates, used to quantify the instantaneous prediction bias of the M-SSM model in a real-world operating environment. At any given time... (For example, at the end of an hourly cycle), the adaptive update and evolution module 40 first obtains data from the data preparation and management module 10. Real-time, latest 3D true observation vector .Should It is a vector processed using the same standardized process as the offline training data.

[0084] Simultaneously, the adaptive update and evolution module 40 calls the M-SSM model trained in the M-SSM model training step. Specifically, it utilizes a Kalman filter in... State estimation at time 1 and the stored state transition matrix ,calculate One-step prediction vector at time step Subsequently, the adaptive update and evolution module 40 calculates... Moment 3D true prediction residual vector .Should It is the vector difference between the actual observed value and the model's one-step prediction: ; The residual vector The M-SSM model was objectively quantified in... Always The joint prediction bias of each factor. Adaptive Update and Evolution Module 40 will... Time calculated 3D residual vector The data is input into the residual monitoring system initialized in the residual monitoring system initialization step.

[0085] In one specific embodiment, the multidimensional statistical process control model is a multidimensional exponentially weighted moving average model. The adaptive update and evolution module 40 uses the smoothing factor stored in the residual monitoring system initialization step. as well as Statistical vector of the multidimensional exponentially weighted moving average model at time t. ( (as a zero vector), recursively updated Statistical vector of the multidimensional exponentially weighted moving average model at time t. : ; this Vector synthesis Instantaneous deviation of time and Historical deviation information prior to the time point (reflected in) (in Chinese). To make Vie The vector is transformed into a single, easily monitored scalar statistic, which is then further computed by the adaptive update and evolution module 40. Hotelling's Moments Statistic The adaptive update and evolution module 40 calls the data calculated and stored during the residual monitoring system initialization step. The inverse of the steady-state covariance matrix of a vector Perform the following quadratic calculations: ; in: This represents the transpose of a vector.

[0086] Calculated scalar Statistics are measures of... A multidimensional statistical measure of the cumulative prediction bias of the M-SSM model up to time 40. The value is output as this step and transmitted to the anomaly mode detection and triggering step, where it is compared with the control upper limit specified in the residual monitoring system initialization step. Conduct continuous comparisons.

[0087] See attached document Figure 1 The adaptive update and evolution module 40 calculates and outputs the results during the real-time monitoring step of the residual flow. Moment After the statistics were collected, then at each time point... The abnormal mode detection and triggering steps are executed. In this step, the adaptive update and evolution module 40 runs in a continuous monitoring loop. The adaptive update and evolution module 40 invokes the abnormal control upper limit calibrated and stored during the offline initialization phase. .

[0088] Adaptive Update and Evolution Module 40 Perform a comparison operation at any time: The current value of the statistic and The values ​​are compared. Based on the comparison result, the adaptive update and evolution module 40 applies a preset out-of-context anomaly detection rule. The purpose of this rule is to distinguish transient, isolated data point anomalies (e.g., single...). Value greater than This is related to persistent, systematic model prediction bias. Persistent, systematic model prediction bias is considered a signal of emerging new health and wellness scenarios that were not defined during the offline modeling phase.

[0089] In one specific embodiment, the anomaly determination rule is defined as: when The value of the statistic is continuous a point in time (i.e.) , (Time) are all greater than At that time, The trigger condition is determined at a certain moment. It is a pre-defined value that is greater than or equal to... Positive integers (e.g.) or ), and store it as system configuration parameters. The adaptive update and evolution module 40 maintains an internal counter. If at any time , The value is less than or equal to If the counter is reset to 0, then the counter is reset to 0.

[0090] When the anomaly detection rule (e.g., continuous) Points )exist When the condition is met, the adaptive update and evolution module 40 determines that the M-SSM model has entered a runaway state and attributes this state to the emergence of a potential new health and wellness scenario. Upon confirmation, the adaptive update and evolution module 40 immediately generates a new scenario discovery trigger signal. This trigger signal activates the extraction and definition step of new scenario features in this phase and instructs this step to begin capturing and analyzing the anomalous data that led to this runaway state.

[0091] See attached document Figure 1The adaptive update and evolution module 40, after generating a new scenario discovery trigger signal in the abnormal mode detection and triggering step, immediately executes the step of extracting and defining new scenario features. The purpose of this step is to automatically and quantitatively separate and solidify new health and wellness meteorological patterns from the observation data that caused the model to go out of control.

[0092] Capturing anomaly data windows. In response to a trigger signal, the adaptive update and evolution module 40 first retrieves an anomaly time window from the system's data storage area (e.g., a cache or database table storing recent observation data). within 3D true observation vector .

[0093] In one specific embodiment, the The window is defined as the time interval that triggers the exception determination, that is, it is determined to be an exception in the exception mode detection and triggering step. Continuous Time points In another embodiment, to obtain more comprehensive data features, the window can be extended to a longer time period that includes the trigger interval (e.g., from...). arrive , For a preset, greater than (Integers). Captured It is a by A set of dimensional vectors .

[0094] Clustering identification of new scenario clusters. Adaptive update and evolution module 40 will capture the clusters. The dataset is used as input, and a clustering algorithm is applied to identify whether there are one or more clusters in the dataset that have not been clustered. The initial scenario covers a high-density data cluster.

[0095] In one specific embodiment, the clustering algorithm employs a density-based spatial clustering algorithm. This algorithm was chosen because it does not require pre-specifying the number of new scenarios. It can identify clusters of arbitrary shapes and recognize sparse data points that do not belong to any new scenario as noise. The adaptive update and evolution module 40 uses a preset neighborhood radius. and minimum points The parameters are used to execute the spatial clustering algorithm. This algorithm will... In the dataset dimensional vector partitioning indivual Clusters with core density , and a set of noise points (if it exists).

[0096] Definition of digital signatures in the new scenario. For the output of spatial clustering algorithms... A new scenario cluster ( For each of these, the adaptive update and evolution module 40 defines it as a new, pending health and wellness scenario and assigns it a unique identifier. ( (This represents the total number of scenarios in the current AHP weight library). This is to quantitatively describe the new scenario. Module 40 calculates the scenario cluster. One A digital signature. In one specific embodiment, the digital signature is defined as the cluster. of 3D centroid vector (i.e., all within this cluster) 3D observation vector (The arithmetic mean vector). Adaptive update and evolution module 40 will... New scenario identifiers and its corresponding 3D digital signature (centroid vector) Stored in the system's data storage area. This output will trigger the weight injection and model evolution steps of human-machine collaboration.

[0097] See attached document Figure 1 The adaptive update and evolution module 40 incorporates the new scenario into the extraction and definition of new scenario features. and its digital signature After being stored in the database and marked as pending review, a human-machine collaborative knowledge base closed-loop process is executed. This step is a crucial link in realizing the evolution of system knowledge, ensuring that new patterns discovered through data-driven approaches can be confirmed by expert knowledge and ultimately fed back into the online prediction model.

[0098] The new scenario undergoes manual review and semantic injection. The adaptive update and evolution module 40, on a dedicated system maintenance and expert configuration interface (e.g., a web management backend), pushes a notification of a new scenario awaiting review to one or more authorized domain experts (e.g., experts in health and wellness or meteorology). This interface displays the new scenario to the experts. Digital signature (Right now (dimensional centroid vector), and can optionally provide anomaly data windows for clustering the scenario. Visualization of raw data.

[0099] Experts, based on their expertise, assess the practical significance of this new scenario. For example, experts identify this new scenario... The characteristics are moderate temperature, low wind speed, and high ozone concentration, which are initially... The scenario was not defined. The expert then performed two actions through the interface: Confirm: Confirm the scenario. It has practical health and wellness significance, and we agree to include it in the knowledge base. Semantic injection: for this scenario Input a human-readable semantic name, such as: Summer afternoon stable ozone pollution. Inject AHP weights into the new scenario. After expert confirmation and approval of the new scenario... After semantic injection, the adaptive update and evolution module 40 immediately invokes the AHP judgment matrix construction and weighting for each scenario on the configuration interface. The calculation steps (offline AHP construction) follow the same process, guiding experts to this new scenario. Inject weighted knowledge.

[0100] Adaptive Update and Evolution Module 40 requires the expert group to consider this new scenario. (For example, stable ozone pollution during summer afternoons) as the target layer, for Each health and wellness factor was compared pairwise. For example, in this scenario, the expert panel determined that the relative importance of ozone concentration was far greater than that of wind speed or humidity. The adaptive update and evolution module 40 receives input from the expert panel... Constructed Judgment Matrix And immediately perform a consistency check. If If the value exceeds a threshold (e.g., 0.1), the expert is prompted to adjust their judgment until the consistency test is passed. If the test is passed, the adaptive update and evolution module 40 calculates... corresponding dimensional static weight vector .

[0101] The AHP weight library and context classifier are updated in a closed loop. This step is the final execution stage of the knowledge loop. The adaptive update and evolution module 40 receives... Then, perform the following two key update operations: AHP Weight Library Expansion: Adaptive Update and Evolution Module 40 applies this set of new, approved scenario weights to... It will be formally written into the AHP weight library and will The status changed from pending review to activated. At this point, the total number of baseline scenarios in the AHP weight library increased from [previous status]. Become .

[0102] Context classifier Evolution of AHP weights: Simply updating the AHP weight library is insufficient because in the context classifier The classifier trained in the training steps Only know The adaptive update and evolution module 40 must handle new scenarios. Knowledge Injection Classifier In one specific embodiment, the adaptive update and evolution module 40 performs incremental updates or model retraining: Incremental update: ... Treat it as a new Gaussian component. Use the anomalous data window captured in the new scenario feature extraction and definition steps. (Right now (the sample set), and the centroid calculated in this step. As the initial mean of this new component, for Parameters (including) Perform incremental updates to increase the number of components from [previous number]. Become Model retraining: Dataset (labeled as) ) and the original training set construction steps Historical datasets (labeled as) Merge and retrain a Each component Classifier.

[0103] After this closed-loop update step, the system (especially the AHP weight library and the context classifier) ​​will be updated. All of them have evolved to possess the ability to identify and assess. The ability to predict and integrate scenarios. When a similar combination of factors leading to stable ozone pollution in the summer afternoon occurs again in the future, the online prediction and integration module 30 will be able to correctly match it to... and invoke the newly injected weights. This enables the adaptive evolution of the predictive ability of the health and wellness index.

[0104] See attached document Figure 1 The adaptive update and evolution module 40, after completing the human-machine collaborative knowledge base closed-loop step, can further perform adaptive evolution of the model as an optional or subsequent step in this stage. The purpose of this step is to incorporate new knowledge (i.e., new scenarios) confirmed by experts within the human-machine collaborative knowledge base closed loop. and its associated abnormal data This feedback is fed back to the M-SSM model trained in the M-SSM model training step, causing the M-SSM itself to evolve. The necessity of this evolution lies in the fact that the loss of control of the M-SSM indicates that its original parameters ( It is no longer possible to accurately describe the new scenarios. The system dynamics, including the AHP weight library and context classifier, are addressed. Updating only the AHP weight library and context classifier solves the evaluation-level problems, while evolutionary M-SSM solves the prediction-level problems.

[0105] The adaptive update and evolution module 40 first constructs a new, expanded training dataset. This dataset It was generated through data merging, and it includes: the offline modeling and initialization module 20 initially used to represent the steady state. dimensional historical observation vector sequence( ), and the abnormal time window dataset representing the new scenario captured in the step of extracting and defining new scenario features. Subsequently, the adaptive update and evolution module 40 uses this expanded training dataset. As input, the exact same training process as the offline modeling phase is invoked, and the parameter matrix of M-SSM is processed ( A re-estimation will be performed.

[0106] Specifically, the adaptive update and evolution module 40 maximizes the value based on Kalman filtering and maximum likelihood estimation. Dataset (containing) Data points, for The (logarithmic) likelihood function constructed from the size of the logarithm To optimize the objective, a numerical optimization algorithm is used iteratively until convergence. After convergence, a new, evolved set of optimal parameters for the M-SSM is obtained. It is used to replace and overwrite the old M-SSM model instance stored in the M-SSM model training step.

[0107] In a preferred embodiment, to ensure the effectiveness of the residual monitoring system, after the M-SSM model completes its evolution (i.e., retraining), the adaptive update and evolution module 40 should automatically trigger a re-initialization of the residual monitoring system. That is, using the new M-SSM model. exist New residual streams generated on the dataset Recalculate the covariance matrix of the multidimensional exponentially weighted moving average model. and new control limits By retraining the M-SSM model (prediction model), evolving the scenario classifier (matching model), expanding the AHP weight library (knowledge model), and reinitializing the residual monitoring system (monitoring benchmark), this invention achieves a complete, data-driven, and expert knowledge-coordinated adaptive and evolutionary closed loop, ensuring that the system's predictive ability and evaluation accuracy can be continuously improved when facing constantly changing environments.

[0108] See attached document Figure 1This invention provides a dynamic prediction and adaptive update system and method for a health and wellness index, which forms the core of this invention. For example... Figure 1 As shown, the system is clearly divided into three collaborative phases in its architecture: offline modeling and initialization module 20, online prediction and integration module 30, and adaptive update and evolution module 40.

[0109] The core technical problem solved by this invention is that existing health and wellness index models generally use fixed factor weights, which makes the models unable to adapt to the dynamic changes in the combination patterns (i.e., scenarios) of health and wellness factors such as weather and environment. The evaluation results under specific scenarios (such as hot and humid versus cold and dry) lack specificity and accuracy.

[0110] To address this issue, the present invention employs an M-SSM model in the offline modeling and initialization module 20. We jointly performed time-series modeling on several health and wellness factors to capture their dynamic interactions; simultaneously, we innovatively constructed a context-weighted knowledge base, which contains two key components: one generated by injecting expert knowledge using the Analytic Hierarchy Process (AHP). Baseline scenario weight vectors A context classifier trained using techniques such as Gaussian mixture models. The classifier learned Each scenario Dimensional factor data distribution characteristics.

[0111] In the online prediction and integration module 30, this invention demonstrates its key dynamic prediction mechanism. When prediction is required... When calculating the health and wellness index at a given time: the M-SSM model first outputs... Dimensional factor prediction vector The prediction vector Input Context Classifier . Instead of outputting a unique hard classification result, it outputs a result through Bayesian inference. dimensional posterior probability distribution vector .Should Value quantified Belongs to the The confidence level of each scenario enables soft matching of fuzzy boundaries between scenarios.

[0112] The system uses this probability vector as dynamic coefficients to apply them to the AHP weight library. Each benchmark weight Perform weighted linear combination Synthesize a unique, fully adapted Status 3D dynamic weight vector Finally, through dot product operation... The system integrates and outputs the health and wellness index at that moment.

[0113] Another core innovation of this invention lies in the adaptive update and evolution module 40. This invention recognizes that any offline model (M-SSM and...) All of these become ineffective due to long-term environmental evolution (such as the emergence of new pollution patterns or extreme weather combinations).

[0114] This invention uses a residual monitoring system to monitor the predicted residuals of M-SSM in real time. Statistical properties. When The statistical volume continues to exceed the control limit. The system determines that the M-SSM has encountered an out-of-context data pattern not included in its training set. This trigger signal activates new scenario extraction, automatically separating new scenarios from the anomalous data. Digital signature.

[0115] Most importantly, the system guides domain experts to new scenarios through a closed-loop human-machine collaboration mechanism. Perform semantic verification and inject new AHP weight knowledge into it. This new knowledge It was used to extend the AHP weight library and classifier The M-SSM model itself was also triggered to evolve, enabling it to recognize and predict the new scenario.

[0116] In summary, the system disclosed in this invention has significant advantages over traditional technologies: Context Adaptability: By combining M-SSM prediction with GMM soft matching, dynamic configuration of health and wellness factor weights is achieved, greatly improving the accuracy of index assessment under different environmental scenarios. Knowledge Evolution: Through a closed-loop adaptive mechanism based on residual monitoring, the system achieves the ability to learn from the unknown. It integrates data-driven pattern discovery... The inclusion of DBSCAN (Database-Based Search and Data Detection) and expert knowledge injection (AHP, Human Verification) enables both the model base and knowledge base to evolve collaboratively with environmental changes, ensuring the long-term robustness and effectiveness of the system.

Claims

1. A method for predicting a health and wellness index based on a time series model, characterized in that, include: Data is obtained from the data source, and a multidimensional historical observation vector is constructed. Training a multidimensional state-space model; Based on historical observation vectors and their corresponding expert-annotated scenarios, a scenario classifier is trained, which is used to output the probability distribution of multiple initial health and wellness meteorological scenarios. Store multiple sets of multidimensional weight vectors corresponding to the initial health and wellness meteorological scenarios; Drive the multidimensional state-space model to output the multidimensional factor prediction vector at the prediction time; The multidimensional factor prediction vector is input into the scenario classifier to obtain the probability distribution of the multiple initial health and wellness meteorological scenarios; Based on the probability distribution and the multiple sets of multidimensional weight vectors, calculate the dynamic weight vector at the prediction time; The multidimensional factor prediction vector and the dynamic weight vector are multiplied by a dot product to generate and output the health and wellness index prediction value.

2. The method for predicting a health and wellness index based on a time series model according to claim 1, characterized in that, Also includes: Obtain the current observation value; Calculate the true prediction residual of the multidimensional state-space model; Drive the multidimensional statistical process control model and update the statistical vector of the multidimensional exponentially weighted moving average model; Calculate the Hotelling T-squared statistic of the statistical vector of the multidimensional exponentially weighted moving average model, and continuously compare the Hotelling T-squared statistic with an anomaly control upper limit; When the multidimensional state space model is determined to encounter an out-of-context anomaly, a new scenario discovery workflow is triggered.

3. The method for predicting a health and wellness index based on a time series model according to claim 2, characterized in that, The new scenario discovery workflow includes: Capture observation data within an abnormal time window; A density-based clustering algorithm is applied to the observation data to identify one or more new scene clusters; Calculate a multidimensional digital signature for the new scenario cluster and define the new scenario cluster as a new scenario.

4. The method for predicting a health and wellness index based on a time series model according to claim 3, characterized in that, Also includes: The new scenario and its multidimensional digital signature are stored in an AHP weight library and its status is marked as pending review. Trigger the human-machine collaboration interface to provide domain experts with the opportunity to review the new scenario; Receive the corresponding new AHP weights injected by the domain expert for the new scenario.

5. The method for predicting a health and wellness index based on a time series model according to claim 4, characterized in that, Also includes: After the new AHP weights are confirmed and activated, the retraining of the scenario classifier is triggered. This enables the scenario classifier to identify all scenarios, including the new scenario.

6. The method for predicting a health and wellness index based on a time series model according to claim 1, characterized in that, The step of storing multiple sets of multidimensional weight vectors includes: Receive multiple initial health and wellness weather scenarios defined by domain experts; For each of the multiple scenarios, the corresponding analytic hierarchy process (AHP) judgment matrix is ​​received; Perform a consistency check on the analytic hierarchy process (AHP) judgment matrix; For the judgment matrix that passes the consistency test, calculate the normalized eigenvector corresponding to its largest eigenvalue, and use it as the multidimensional weight vector.

7. The method for predicting a health and wellness index based on a time series model according to claim 1, characterized in that, The steps for training the scenario classifier include: Construct a labeled training set, which includes historical observation vectors and their corresponding scene labels; A Gaussian mixture model is used as the scenario classifier, and the Gaussian mixture model contains multiple Gaussian components that correspond to the initial health and wellness meteorological scenario. Using the expectation-maximization algorithm, parameters of multiple scenario Gaussian components are estimated based on the training set. These parameters include prior probabilities, multidimensional mean vectors, and multidimensional covariance matrices.

8. The method for predicting a health and wellness index based on a time series model according to claim 1, characterized in that, The method further includes data preparation and management steps, which include: Obtain historical time-series data of multiple heterogeneous health and wellness factors from one or more external data sources; Preprocessing operations are performed on the historical time series data, including data cleaning, missing value imputation, and data standardization. Set a uniform time frequency and perform resampling on multiple factor sequences; By aligning all the standardized factor values ​​generated through the data standardization operation according to the timestamp, a multidimensional historical observation vector is constructed.

9. The method for predicting a health and wellness index based on a time series model according to claim 1, characterized in that, The multidimensional state-space model is constructed using the form of a structural time series model; The state vector of the multidimensional state-space model is decomposed into a stack of multiple components, the components including: Components with local linear trends, cyclical seasonality, and vector autoregression.

10. An application for predicting a health and wellness index based on a time series model, characterized in that, It includes one or more processors, said one or more processors being configured to perform the method of any one of claims 1 to 9.

Images