Student nutrition balance monitoring and individualized intervention system based on multi-source data fusion

Abstract

The present application relates to the field of artificial intelligence and intelligent health data processing technology, especially to a student nutrition balance monitoring and personalized intervention system based on multi-source data fusion, comprising: a multi-source heterogeneous perception acquisition module, which acquires the dietary images, continuous sensing signals and regular physical examination archives of a target object within a preset time period, and generates an original physiological and behavioral data set; a feature level fusion extraction module, which extracts the nutrition intake, energy consumption and body state index of the target object based on the original physiological and behavioral data set, and calculates the historical index change rate of the body state index. In the present application, by constructing a state prediction model, the traditional responsive monitoring is upgraded to a predictive closed-loop control, the time series model is used to predict the future development trajectory, the adaptive baseline is calculated combined with individual differences, and the early warning parameters are dynamically generated through derivative comparison, which effectively removes the index growth caused by natural growth and reduces the evaluation error caused by a single threshold.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and smart health data processing technology, and in particular to a student nutrition balance monitoring and personalized intervention system based on multi-source data fusion. Background Technology

[0002] With the increasing demand for physical health management among school-aged children, schools and families are gradually adopting information technology to continuously monitor students' body mass index, dietary structure, and exercise consumption data in order to achieve a scientific assessment of students' nutritional status and physiological development. Existing systems typically collect multi-source information such as physical examination data, dietary records, and exercise data to quantitatively analyze students' current physical condition and determine whether there are risks such as overweight, obesity, or nutritional imbalance based on preset health standards. This provides schools and administrators with a basis for health intervention and promotes the standardized and refined management of students' physical health.

[0003] However, existing technologies mostly rely on current monitoring data and fixed threshold standards to determine the status, which is a reactive analysis method. It lacks the ability to predict students' future physiological development trends and makes it difficult to identify potential risk changes in a timely manner. At the same time, since different students are at different stages of growth and development, their physical indicators have natural growth differences. Using a uniform fixed threshold standard can easily affect the accuracy of the assessment results, resulting in insufficient individual adaptability in health status evaluation. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a student nutrition balance monitoring and personalized intervention system based on multi-source data fusion, aiming to improve the problem that existing technologies mostly rely on current monitoring data and fixed threshold standards for status determination, which is a post-event response analysis method.

[0005] This invention provides the following technical solution: a student nutrition balance monitoring and personalized intervention system based on multi-source data fusion, comprising: The multi-source heterogeneous sensing and acquisition module acquires dietary images, continuous sensor signals and regular physical examination records of the target object within a preset time period, and generates raw physiological and behavioral datasets. The feature-level fusion extraction module extracts the nutritional intake, energy consumption, and body shape indicators of the target object based on the original physiological and behavioral dataset, calculates the historical rate of change of the body shape indicators, and aligns the nutritional intake, energy consumption, body shape indicators, and historical rate of change of the indicators according to their respective timestamps to generate a multidimensional metabolic feature matrix and a historical state sequence. The time-series state prediction module, based on the multidimensional metabolic feature matrix and the historical state sequence, inputs the multidimensional metabolic feature matrix and the historical state sequence into a pre-trained time-series deep learning model and outputs a physiological development prediction trajectory. The dynamic threshold adaptive evaluation module calculates an adaptive baseline based on the predicted physiological development trajectory and the historical index change rate, combined with the body posture index. It then compares the derivatives of the predicted physiological development trajectory and the adaptive baseline to generate state warning parameters. The decision-level fusion and closed-loop intervention module, based on the multidimensional metabolic feature matrix and the state warning parameters, compares the multidimensional metabolic feature matrix and the state warning parameters through expert system rules, outputs an intervention plan, and inputs the subsequent continuous sensing signals of the target object into the multi-source heterogeneous sensing and acquisition module.

[0006] Preferably, in the multi-source heterogeneous sensing acquisition module, the generation of the original physiological and behavioral dataset specifically includes the following steps: Extract the dietary images, the continuous sensor signals, and the timestamps from the periodic physical examination records, respectively. Based on multiple timestamps, the dietary images, the continuous sensor signals, and the regular physical examination records are correlated in a time series to generate time series correlation data. The time-series correlated data are combined to generate the original physiological and behavioral dataset.

[0007] Preferably, in the feature-level fusion extraction module, the extraction of the target object's nutrient intake, energy consumption, and body shape indicators, and the calculation of the historical rate of change of the body shape indicators specifically include the following steps: Analyze the periodic physical examination records in the original physiological and behavioral dataset to extract body posture indicators; Perform time-series difference calculations on the aforementioned body posture indicators to generate historical indicator change rates; A visual semantic segmentation algorithm is used to extract pixel features and calculate volume mapping on the dietary images in the original physiological and behavioral dataset to generate nutrient intake. A time-series signal analysis algorithm is used to perform frequency domain transformation and feature matching on the continuous sensor signals in the original physiological and behavioral dataset to generate energy consumption data.

[0008] Preferably, in the feature-level fusion extraction module, the fusion to generate a multidimensional metabolic feature matrix and a historical state sequence specifically includes the following steps: Obtain the timestamps corresponding to the nutrient intake, energy expenditure, body shape indicators, and the rate of change of the historical indicators; Based on the timestamp, the nutrient intake and energy consumption at the same time point are combined to generate a metabolic feature vector; Multiple metabolic feature vectors are stacked in multiple dimensions to generate a multidimensional metabolic feature matrix. Based on the timestamp, the change rates of the body posture indicators and the historical indicators are sorted and correlated in a time sequence to generate a historical state sequence.

[0009] Preferably, in the time-series state prediction module, the step of inputting the multidimensional metabolic feature matrix and the historical state sequence into a pre-trained time-series deep learning model and outputting a physiological development prediction trajectory specifically includes the following steps: Temporal features are extracted from the multidimensional metabolic feature matrix to generate metabolic temporal features; Trend features are extracted from the historical state sequence to generate growth and development trend features; The metabolic time-series features are concatenated with the growth and development trend features to generate a joint input feature matrix; The joint input feature matrix is ​​input into a pre-trained time series deep learning model based on a long short-term memory network to perform temporal state evolution calculation, outputting a multi-time step predicted state vector sequence, and combining the predicted state vector sequences to generate a physiological development prediction trajectory.

[0010] Preferably, the training steps of the pre-trained time-series deep learning model based on a long short-term memory network include: Obtain the joint input feature matrix of multiple sets of samples and the corresponding labels of real physiological developmental trajectories; The multiple sets of samples are jointly input into the feature matrix and then input into the initial long short-term memory network model to output the sample predicted trajectory. The deviation between the predicted trajectory of the sample and the actual physiological development trajectory label is calculated to generate a temporal loss value; Based on the time-series loss value, the weight parameters of the initial long short-term memory network model are updated using the backpropagation algorithm to generate a pre-trained time-series deep learning model.

[0011] Preferably, in the dynamic threshold adaptive evaluation module, the step of comparing the physiological development prediction trajectory with the adaptive baseline by its derivative to generate state warning parameters specifically includes the following steps: The time-dimensional derivative of the predicted physiological development trajectory is calculated to generate a sequence of predicted trajectory change rates; The time-dimensional derivative of the adaptive baseline is calculated to generate a baseline change rate sequence; The difference between the predicted trajectory rate of change sequence and the baseline rate of change sequence is calculated to generate a time-series deviation matrix. The peak features in the time-series deviation matrix are extracted as state warning parameters.

[0012] Preferably, in the decision-level fusion and closed-loop intervention module, the step of comparing the multidimensional metabolic feature matrix with the state warning parameters through expert system rules and outputting the intervention plan specifically includes the following steps: Extract the state features of the current time node from the multidimensional metabolic feature matrix to generate the current metabolic state feature vector; Risk feature mapping is performed on the aforementioned status warning parameters to generate a warning status feature vector; The current metabolic state feature vector is combined with the warning state feature vector to generate a comprehensive risk assessment vector; Based on the comprehensive risk assessment vector query, a pre-set intervention strategy rule base is used for matching and calculation, and an intervention plan is output.

[0013] The present invention has the following beneficial effects: 1. In this invention, by constructing a state prediction model, the traditional response monitoring is upgraded to predictive closed-loop control. The time series model is used to predict the future development trajectory, and an adaptive baseline is calculated in combination with individual differences. Early warning parameters are dynamically generated by comparing derivatives, which effectively removes the index growth brought about by natural growth, reduces the evaluation error caused by a single threshold, and provides an objective basis for precise intervention.

[0014] 2. In this invention, a hierarchical fusion strategy is adopted to process multi-source heterogeneous data. Nutritional and physical features are extracted through feature-level fusion, and metabolic status and early warning risks are matched by decision-level fusion. This effectively reduces the computational complexity of high-dimensional data and accurately establishes the intrinsic relationship between the underlying physical perception data and the upper physiological development trend.

[0015] 3. In this invention, a business closed loop covering data perception, decision intervention and execution feedback is constructed. By matching rules, abstract risks are transformed into specific diet and exercise management plans. Real-time sensor data from feedback is used to trigger continuous iteration, thereby realizing dynamic adjustment of intervention strategies and improving the scientific nature and consistency of health management. Attached Figure Description

[0016] Figure 1 This is an architecture diagram of the student nutrition balance monitoring and personalized intervention system based on multi-source data fusion proposed in this invention. 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] This invention provides a student nutrition balance monitoring and personalized intervention system based on multi-source data fusion, such as... Figure 1 As shown, it includes: The multi-source heterogeneous sensing and acquisition module acquires dietary images, continuous sensor signals and regular physical examination records of the target object within a preset time period, and generates raw physiological and behavioral datasets. Furthermore, in the multi-source heterogeneous sensing acquisition module, generating the original physiological and behavioral dataset specifically includes the following steps: Extract the timestamps from dietary images, continuous sensor signals, and regular physical examination records, respectively. Time-series correlation is performed on dietary images, continuous sensor signals, and regular physical examination records based on multiple timestamps to generate time-series correlation data. By combining time-series correlated data, a raw physiological and behavioral dataset is generated.

[0019] Specifically, in the multi-source heterogeneous sensing and acquisition module, the system first collects multimodal data of the target object within a preset time period through hardware devices deployed on students' smart terminals and in the campus environment. Specifically, the system acquires dietary images taken by students during breakfast, lunch and dinner times, and at the same time, it collects continuous sensing signals including triaxial acceleration and photoplethysmography pulse waves in real time through smart wearable devices worn by students at a fixed sampling frequency, and obtains regular physical examination records including height, weight and skinfold thickness by connecting to the campus medical management system.

[0020] After acquiring the aforementioned multi-source data, the system extracts the timestamps from dietary images, continuous sensor signals, and regular physical examination records. For dietary images, the system reads the interchangeable image file format attribute of the image file and extracts the absolute physical time at the time of capture as the image timestamp. For continuous sensor signals, the system parses the frame header information of the underlying data packets and extracts the microsecond-level high-frequency sampling timestamp. For regular physical examination records, the system reads the entry time field of the medical system database as the examination timestamp.

[0021] After extracting the timestamps, the system performs time-series correlation on the three types of heterogeneous data based on multiple timestamps. Since dietary images are discrete event data, continuous sensor signals are high-frequency continuous data, and regular health check records are extremely low-frequency discrete data, the system constructs a time-series correlation algorithm based on anchored time windows to align these three types of heterogeneous data. It sets daily as the basic time-series benchmark and divides the time window into sliding time windows with hourly steps. Let the set of time windows be... ,in Indicates the first The start-to-end time interval of a time window, for any time window The system calculates the time-series correlation data set within this period using the following formula. : ; In the formula, This represents the input set of dietary images. This indicates that the extracted timestamp falls within the current time window. Discrete mapping function for internal diet images; This represents a continuous sequence of input sensing signals. This represents the truncation function that performs discrete sampling and mean filtering on the triaxial acceleration and pulse wave within the time window, generating a smoothed sensing time series data segment; This indicates the input of regular medical examination records. This represents the nearest matching function, used to find the most recent medical examination record data with the smallest absolute time difference from the current time window as the baseline state for that time window; This refers to the structured combination operation of multimodal data, which involves aligning input data from different dimensions to the same time window node to form a composite data dictionary.

[0022] Through the aforementioned temporal correlation algorithm, the system maps scattered multi-source heterogeneous data onto a unified time axis to generate temporal correlation data. Subsequently, the system combines this temporal correlation data in multiple dimensions according to the time window sequence and encapsulates it into a standard data structure containing image matrices, sensor sequences, and physical examination scalar features, thereby generating the original physiological and behavioral dataset.

[0023] This achieves spatiotemporal alignment and structured integration of student health and behavior data collected at different frequencies, eliminates asynchronous bias in heterogeneous data, and provides a standardized underlying data source for the subsequent feature-level fusion extraction module to accurately extract nutrient intake and energy consumption.

[0024] The feature-level fusion extraction module extracts the target object's nutrient intake, energy consumption, and body shape indicators based on the original physiological and behavioral dataset. It also calculates the historical rate of change of the body shape indicators and aligns the nutrient intake, energy consumption, body shape indicators, and historical rate of change according to their respective timestamps, fusing them to generate a multidimensional metabolic feature matrix and a historical state sequence. Furthermore, the feature-level fusion extraction module extracts the target object's nutrient intake, energy expenditure, and body shape indicators, and calculates the historical rate of change of body shape indicators, specifically including the following steps: Analyze the periodic physical examination records in the original physiological and behavioral dataset to extract body posture indicators; Perform time-series difference calculations on body posture indicators to generate historical rate of change of indicators; A visual semantic segmentation algorithm is used to extract pixel features and calculate volume mapping from dietary images in the original physiological and behavioral dataset to generate nutrient intake. A time-series signal analysis algorithm is used to perform frequency domain transformation and feature matching on continuous sensor signals in the original physiological and behavioral dataset to generate energy consumption data.

[0025] Furthermore, in the feature-level fusion extraction module, the specific steps for fusing and generating a multidimensional metabolic feature matrix and historical state sequence include: Obtain the timestamps corresponding to nutrient intake, energy expenditure, body shape indicators, and historical rate of change of these indicators. Based on the timestamp, the nutritional intake and energy consumption at the same time point are combined to generate a metabolic feature vector; Multiple metabolic feature vectors are stacked in different dimensions to generate a multidimensional metabolic feature matrix. Based on timestamps, the change rates of body posture indicators and historical indicators are sorted and correlated in a time sequence to generate a historical state sequence.

[0026] Specifically, after obtaining the raw physiological and behavioral dataset generated by the aforementioned modules, the feature-level fusion extraction module first parses the physical examination scalar features in the composite data dictionary corresponding to each time window, extracting the target object's height, weight, and skinfold thickness as body posture indicators. The system performs temporal difference calculations on the body posture indicators of adjacent physical examination nodes to generate historical indicator change rates. Let the formula for calculating this historical indicator change rate be... In the formula Indicates the first Historical rate of change of indicators for individual testing nodes. Indicates the first A comprehensive score of body posture indicators extracted from individual inspection nodes. Indicates the first A physical examination node is the comprehensive score of physical condition indicators from the previous physical examination. and The denominator represents the absolute physical timestamps corresponding to the current physical examination and the previous physical examination, respectively. This represents the actual physical time span between two physical examination timestamps, and the difference reflects the student's absolute growth and development rate within a specific period.

[0027] The system then employs a visual semantic segmentation algorithm to process the dietary image matrix in the composite data dictionary. The system extracts food category pixel features from the images using a pre-trained convolutional neural network model, calculates the pixel area of ​​each food category, and performs volume mapping calculations using depth mapping coefficients. Finally, it generates the nutritional intake for that time window based on the campus food nutrition facts table. Let the formula for calculating nutritional intake be... In the formula This represents the total calorie and macronutrient intake of the target individual within the current time window. This represents the set of all food categories contained in the current dietary image, identified by a visual semantic segmentation algorithm. Represents a set A specific food category, Indicate category The total pixel area occupied by food in a two-dimensional dietary image. This represents the depth mapping constant coefficient used to map two-dimensional pixel areas to three-dimensional physical volumes. This indicates the category obtained by looking up the nutritional information of campus food. Metabolic heat and macronutrient density scalars corresponding to a unit physical volume of food.

[0028] Simultaneously, the system employs a time-series signal analysis algorithm to process the sensor time-series data segments in the composite data dictionary. The system performs a Fast Fourier Transform on the triaxial acceleration sequence, converting it from the time domain to the frequency domain, extracting the power spectral density of the dominant motion frequency band, and performing feature matching with a pre-set database of metabolic equivalents of physical activity in school-aged children. This, combined with heart rate features extracted from the pulse wave, generates energy consumption. Let the energy consumption calculation formula be... In the formula This indicates the energy consumption of the target object within the current time window. This represents the basal metabolic rate of the target subject calculated based on the aforementioned extracted body shape indicators such as height and weight. This represents the specific discrete frequency points within the main frequency band of motion extracted after the Fast Fourier Transform. Indicates frequency The power spectral density value obtained from the analysis, Indicates frequency The activity metabolic equivalent coefficients obtained after matching with a preset metabolic equivalent database are mapped. This represents the average heart rate value for the current time window extracted from the photoplethysmography (PPG) pulse wave sensing time-series data segment. This represents the empirical conversion weight constant used to convert the physiological index of heart rate into the corresponding calorie consumption value.

[0029] After completing the feature extraction, the system obtains the timestamps corresponding to nutrient intake, energy expenditure, body shape indicators, and historical rate of change of these indicators. For high-frequency daily metabolic behaviors, the system uses the timestamps of time windows to group data from the same time points. Nutritional intake With energy consumption Feature combination is performed to generate metabolic feature vectors for a single time point. In the formula Indicates the first Metabolic feature vectors for each time window, with superscript This represents the matrix transpose operation. The system stacks the metabolic feature vectors of multiple consecutive time windows according to temporal order to generate a two-dimensional multidimensional metabolic feature matrix. For low-frequency physiological development features, the system sorts and associates the change rates of body shape indicators and historical indicators according to absolute timestamps, packages features corresponding to the same physical examination time into state nodes, and concatenates them in chronological order to generate a historical state sequence.

[0030] By transforming multi-source images and underlying sensor signals into structured metabolic and developmental trend features with clear biological significance, the dimensionality reduction and decoupling of high-frequency metabolic fluctuations and low-frequency body shape changes were achieved, providing feature-aligned and dimensionally unified data input support for subsequent time-series deep learning prediction models.

[0031] The time-series state prediction module, based on the multidimensional metabolic feature matrix and historical state sequence, inputs the multidimensional metabolic feature matrix and historical state sequence into a pre-trained time-series deep learning model and outputs a physiological development prediction trajectory. Furthermore, in the time-series state prediction module, the multidimensional metabolic feature matrix and historical state sequence are input into a pre-trained time-series deep learning model, and the output physiological development prediction trajectory specifically includes the following steps: Temporal features are extracted from the multidimensional metabolic feature matrix to generate metabolic temporal features; Trend features are extracted from historical state sequences to generate growth and development trend features; Metabolic time-series features and growth and development trend features are concatenated to generate a joint input feature matrix; The joint input feature matrix is ​​input into a pre-trained time series deep learning model based on a long short-term memory network to perform temporal state evolution calculations, outputting a multi-time step predicted state vector sequence, and combining the predicted state vector sequences to generate a physiological development prediction trajectory.

[0032] Furthermore, the training steps for a pre-trained time-series deep learning model based on a long short-term memory network include: Obtain the joint input feature matrix of multiple sets of samples and the corresponding labels of real physiological developmental trajectories; Multiple sets of samples are jointly input into the feature matrix and then input into the initial long short-term memory network model to output the predicted trajectory of the samples. The deviation between the predicted trajectory of the sample and the actual physiological development trajectory label is calculated to generate a time-series loss value; Based on the temporal loss value, the weight parameters of the initial long short-term memory network model are updated using the backpropagation algorithm to generate a pre-trained time series deep learning model.

[0033] Specifically, after obtaining the multidimensional metabolic feature matrix and historical state sequence generated by the aforementioned feature-level fusion extraction module, the temporal state prediction module first extracts temporal features from the multidimensional metabolic feature matrix. Since the multidimensional metabolic feature matrix contains high-frequency nutrient intake and energy consumption within a time window dimension, the system uses a one-dimensional convolutional neural network layer to extract its local metabolic temporal dependencies. Let the temporal feature extraction formula be: ; In the formula, Indicates at time step Extract the generated one-dimensional metabolic time-series feature vector; Represents a non-linear activation function; This represents the kernel weight matrix of a one-dimensional convolutional network layer. This represents the standard convolution operation; This represents the multidimensional metabolic feature matrix at time step. The corresponding extracted local sliding window feature sub-matrix specifically contains the continuous fluctuation values ​​of the target student's daily calorie intake and energy consumption within the sliding window; This represents the bias vector of a one-dimensional convolutional network layer.

[0034] Simultaneously, the system extracts trend features from the historical state sequence. Since the historical state sequence contains postural indicators and their rates of change at low-frequency absolute physical time points, the system employs a fully connected mapping layer to transform its feature space to align with the dimensions of high-frequency features. Let the formula for this trend feature extraction be: ; In the formula, Indicates at time step The generated growth and development trend feature vector; This represents the weight matrix of the fully connected mapping layer; This indicates that the copy is aligned to the time step after the time step. The historical state sequence feature vector, which specifically contains the scalar values ​​of the target student's height, weight, and skinfold thickness at that time point; This represents the bias vector of the fully connected mapping layer. After the above extraction, the system concatenates the metabolic time-series features and growth and development trend features at the same time step according to their feature dimensions to generate a joint input feature matrix. Let the concatenation formula be: ; In the formula, Indicates time step The joint input feature matrix; This indicates a concatenation operation that links the two feature vectors together at the beginning and end along the feature channel dimension.

[0035] After concatenation, the system inputs the joint input feature matrix into a pre-trained time-series deep learning model based on a long short-term memory network for temporal state evolution calculation. This deep learning model includes an input layer, a long short-term memory hidden layer containing a forget gate, an input gate, and an output gate, and a linear output layer. Let the core hidden layer evolution calculation formula be: ; In the formula, This indicates that the model is at the current time step. The output hidden state feature vector; This represents the nonlinear gated aggregation mapping function inside the hidden layer of a Long Short-Term Memory (LSTM) network. For the current time step The joint input feature matrix; This indicates that the propagation from the previous time step to the current time step is... The hidden state feature vector; This indicates that the propagation from the previous time step to the current time step is... The hidden cellular state vector is used to store long-term developmental trend memories. The model then maps the hidden state feature vector to student physiological development index values ​​at the corresponding time steps through a linear output layer. Let the output formula be: ; In the formula, Indicates the current time step The output predicted state vector, whose specific data corresponds to the predicted height, weight or body mass index value of the student at a specific future time point; This represents the weight matrix of the linear output layer; This represents the bias vector of the linear output layer. The system sequentially combines multiple predicted state vectors from continuous outputs at multiple time steps to generate a continuous physiological development prediction trajectory.

[0036] The training of the pre-trained time-series deep learning model based on a long short-term memory network is completed offline before model deployment. The system first obtains a joint input feature matrix and corresponding real physiological development trajectory labels from multiple groups of school-aged children in the school's historical health database. These real physiological development trajectory labels are derived from time-series data of real student physical examination records with known long-term follow-up results. The system inputs the joint input feature matrix of multiple samples into the initial, unoptimized long short-term memory network model, and outputs the corresponding predicted trajectory of the samples according to the aforementioned forward propagation calculation of the network hierarchy.

[0037] After obtaining the predicted trajectory of the sample, the system uses the mean squared error function to calculate the deviation between the predicted trajectory and the actual physiological development trajectory label, generating a time-series loss value. Let the formula for calculating the time-series loss value be: ; In the formula, Represents the global time series loss value as a scalar; This indicates the total number of batches of samples in a single training iteration. This represents the total length of the temporal nodes contained in a single physiological developmental trajectory; This indicates the index number of the current sample in the summation operation; This indicates the index number of the current time step in the summation operation; Indicates the first A sample of students at time step The predicted state vector output by the model's forward propagation; Indicates the first A sample of students at time step The corresponding real physiological development characteristics are specifically the real height and weight values ​​recorded in the historical database. Based on the calculated global temporal loss value, the system uses an adaptive moment estimation optimizer to calculate the network gradient through backpropagation, and uses this to update all weight matrices and bias vector parameters in the initial long short-term memory network model. After multiple iterative calculations until the global temporal loss value converges to below a preset minimum threshold, the system saves the converged network model parameters and formally generates the pre-trained time-series deep learning model.

[0038] By using a time-series deep learning model, the complex temporal nonlinear mapping relationship between students' daily high-frequency dietary consumption in the multidimensional metabolic feature matrix and long-term low-frequency physical development in the historical state sequence was captured. This enabled a forward-looking quantitative prediction of the future healthy development trend of individual students, providing a precise data foundation with clear physical meaning and time dimension alignment for the subsequent dynamic threshold adaptive evaluation module to compare the predicted trajectory with the baseline to generate state warning parameters.

[0039] The dynamic threshold adaptive assessment module calculates an adaptive baseline based on the predicted trajectory of physiological development and the rate of change of historical indicators, combined with body posture indicators. It then compares the derivatives of the predicted trajectory of physiological development and the adaptive baseline to generate status warning parameters. Furthermore, in the dynamic threshold adaptive evaluation module, the physiological development prediction trajectory is compared with the adaptive baseline by its derivative to generate state warning parameters. This process specifically includes the following steps: The time-dimensional derivative of the predicted physiological development trajectory is calculated to generate a sequence of the rate of change of the predicted trajectory. Calculate the time-dimensional derivative of the adaptive baseline to generate a baseline change rate sequence; The difference between the predicted trajectory rate of change sequence and the baseline rate of change sequence is calculated to generate a time series deviation matrix. Peak features in the time-series deviation matrix are extracted as state warning parameters.

[0040] Specifically, after obtaining the physiological development prediction trajectory output by the aforementioned time-series state prediction module, the dynamic threshold adaptive evaluation module first calculates an adaptive baseline based on the time span corresponding to the physiological development prediction trajectory, combined with the historical index change rate and body shape index output by the feature-level fusion extraction module. The system extracts the actual body mass index scalar calculated from the height and weight of the target object at the current physical examination node as the initial body shape index benchmark value, and accumulates it with the product of the historical index change rate multiplied by the future prediction time step to generate an adaptive baseline that fits the natural growth and development rhythm of the student. Let the adaptive baseline calculation formula be: ; In the formula, Indicates future discrete time steps The adaptive baseline expected body mass index scalar value calculated at the location; This represents the current true physical fitness index benchmark scalar obtained by parsing regular physical examination records based on the feature-level fusion extraction module; This represents the historical body mass index change rate scalar calculated based on the comprehensive score of body shape indicators at adjacent physical examination nodes; It represents the discrete time step index number for projecting into the future, specifically indicating which future calendar day is being predicted.

[0041] After calculating the adaptive baseline, the system compares the derivatives of the predicted physiological development trajectory with the adaptive baseline. The system uses a first-order forward differencing algorithm to calculate the time-dimensional derivative of the predicted physiological development trajectory, generating a sequence of predicted trajectory change rates. Let the formula for generating the predicted trajectory change rate sequence be: ; In the formula, This indicates the predicted trajectory of physiological development at time step The scalar of the rate of change of the predicted trajectory at a given point in time is used to characterize the student’s expected rate of development at that specific point in time. This represents the physiological developmental prediction trajectory output by the time series deep learning model at time step [missing information]. The expected body mass index scalar value at the location; This indicates the predicted trajectory of physiological development at time step The expected body mass index scalar value at the location; This represents the fixed physical time interval constant between two adjacent prediction time steps, which is taken as a single natural day here.

[0042] Similarly, the system calculates the time-dimensional derivative of the adaptive baseline to generate a baseline rate of change sequence. Let the formula for generating the baseline rate of change sequence be: ; In the formula, Indicates the adaptive baseline at time step The baseline rate of change scalar is used to characterize the baseline developmental rate of the student under natural growth conditions; Indicates the adaptive baseline at time step The expected body mass index scalar value at the location; Indicates the adaptive baseline at time step The expected body mass index scalar value at the location; Similarly, it represents a fixed physical time interval constant between adjacent prediction time steps.

[0043] The system then calculates the difference between the predicted trajectory rate of change sequence and the baseline rate of change sequence by time alignment, generating a time deviation matrix. The formula for calculating the elements of the time deviation matrix is ​​as follows: ; In the formula, The timing deviation matrix represents the time step. The corresponding developmental deviation scalar element specifically refers to the pure fluctuation difference between the predicted developmental rate and the natural baseline developmental rate. This represents the scalar value representing the rate of change of the predicted trajectory calculated above; The baseline change rate scalar represents the aforementioned calculation. The developmental deviation scalar elements of multiple predicted time steps are arranged in chronological order to form a complete temporal deviation matrix. The system uses a sliding window extreme value search algorithm to traverse the temporal deviation matrix in a temporal sequence, extracts local peak features in the temporal deviation matrix whose absolute values ​​are greater than the preset medical safety threshold, and encapsulates the abnormal developmental deviation scalar values ​​corresponding to the local peak features and their future warning time nodes in a structured data manner to formally generate state warning parameters.

[0044] By using first-derivative difference comparison, the natural increase in absolute values ​​of indicators brought about by normal physiological development in school-aged children was stripped away, and the deviation trend of poor posture caused by recent abnormal metabolism was accurately quantified, providing an objective risk assessment basis for subsequent modules to generate de-trend-based precise intervention plans.

[0045] The decision-level fusion and closed-loop intervention module, based on the multidimensional metabolic feature matrix and state warning parameters, compares the multidimensional metabolic feature matrix and state warning parameters through expert system rules, outputs the intervention plan, and inputs the subsequent continuous sensing signals of the target object into the multi-source heterogeneous sensing acquisition module.

[0046] Furthermore, in the decision-level fusion and closed-loop intervention module, the intervention plan is output by comparing the multidimensional metabolic feature matrix with the state warning parameters through expert system rules, specifically including the following steps: Extract the state features of the current time node from the multidimensional metabolic feature matrix to generate the current metabolic state feature vector; Risk feature mapping is performed on the status warning parameters to generate a warning status feature vector; The current metabolic state feature vector is combined with the warning state feature vector to generate a comprehensive risk assessment vector; The intervention plan is output by matching and calculating the pre-set intervention strategy rule base based on the comprehensive risk assessment vector query.

[0047] Specifically, after acquiring the feature data generated by the aforementioned dynamic threshold adaptive evaluation module, the decision-level fusion and closed-loop intervention module first extracts the state features of the multidimensional metabolic feature matrix at the current time node. The system sets a short-term sliding window containing the most recent seven natural days, extracts the nutrient intake sequence and energy consumption sequence within this window from the multidimensional metabolic feature matrix, and generates the current metabolic state feature vector by calculating the mean. Let the formula for calculating the current metabolic state feature vector be: ; In the formula, This represents the current metabolic state feature vector, used to characterize the average lifestyle and daily metabolic habits of the target student in the past week; This represents the time step constant of the short-term sliding window, specifically a value of seven calendar days. This indicates the index number of the current time step in the summation operation; This represents the first element in the aforementioned multidimensional metabolic feature matrix. The scalar value of total daily nutrient intake calories corresponding to each time step; This represents the first element in the aforementioned multidimensional metabolic feature matrix. The scalar value of total daily energy consumption for each time step; superscript This indicates that the generated row vector is transposed to generate a standard column vector.

[0048] Simultaneously, the system performs risk feature mapping on the status warning parameters output by the aforementioned modules, generating a warning status feature vector. The system analyzes the local peak features contained in the status warning parameters, extracts the abnormal development deviation scalar and its corresponding future warning time node, and maps it into a vector feature containing the severity of the deviation and the urgency of the warning. Let the formula for calculating the warning status feature vector be: ; In the formula, The feature vector representing the early warning status is used to quantify the severity and urgency of the student's future health risks of obesity, overweight, or malnutrition. This represents the local peak feature of the temporal deviation matrix extracted from the status warning parameters, i.e., the abnormal development deviation scalar. This indicates that the absolute value of the scalar deviation from abnormal development is taken to characterize the magnitude of the deviation. This represents the difference in physical natural days between the current actual time point and the future warning time point where this peak characteristic occurs. The smaller this value, the higher the degree of imminent health risk. This represents the first weighted mapping coefficient used to normalize the magnitude of abnormal development deviations to the standard risk dimension. This represents the second weighted mapping coefficient used to normalize the numerical value of time urgency to the standard risk dimension.

[0049] After generating the above feature vectors, the system combines the current metabolic state feature vector and the warning state feature vector by dimension to generate a comprehensive risk assessment vector. Let the formula for generating the comprehensive risk assessment vector be: ; In the formula, This represents a comprehensive risk assessment vector, which fully defines the coordinates of the overall health status of the school-aged child in a multidimensional feature mathematical space, encompassing the child's current metabolic behavior and future developmental risks. This represents the two-dimensional current metabolic state feature column vector obtained from the aforementioned calculation; This represents the two-dimensional warning status feature column vector obtained from the aforementioned calculation; This represents a vector concatenation operation, which combines the two two-dimensional column vectors in the row dimension into a four-dimensional comprehensive risk assessment column vector.

[0050] The system then performs matching calculations based on the comprehensive risk assessment vector query within a pre-defined intervention strategy rule base, outputting an intervention plan. The intervention strategy rule base stores multiple sets of standard risk template vectors developed by clinical nutrition experts and their corresponding standard intervention plans. The system uses the Euclidean distance algorithm to calculate the spatial similarity distance between the currently generated comprehensive risk assessment vector and each standard risk template vector in the rule base. Let the matching calculation formula be: ; In the formula, This represents the standard risk template index number that best matches the student's current overall health status, calculated after global distance comparison. This indicates the search for the independent variable that minimizes the subsequent distance function. The mathematical extremum search operation; This represents the traversal index number of the standard risk template vector stored in the intervention strategy rule base; This represents the total dimension of the feature vectors involved in the distance calculation; here, it is set to four dimensions. This indicates the index number of the current vector dimension in the summation operation; This represents the first element in the currently generated four-dimensional comprehensive risk assessment vector. Numerical values ​​of feature components in each dimension; This indicates the first rule in the intervention strategy rule base. The first standard risk template vector Numerical values ​​of each dimension of feature components.

[0051] The system retrieves the corresponding bound intervention plan based on the calculated optimal matching index number. This intervention plan specifically includes the recommended daily calorie intake limit for the target student's three meals in the next cycle and the recommended daily duration of moderate-to-vigorous physical activity. Finally, the system distributes the intervention plan to the target student's wearable smart terminal via the campus network to provide health reminders. The system also feeds back the continuous sensor signals and dietary images generated after the target student follows the guidance plan to the multi-source heterogeneous sensing acquisition module in real time, triggering a new round of closed-loop data acquisition and evolutionary calculation.

[0052] This step uses expert system rules to accurately map abstract, high-dimensional health risk characteristics into specific calorie intake and exercise expenditure management instructions that students can intuitively execute. It establishes a seamless business loop between underlying physical perception and upper-level medical intervention, ensuring that the campus health management plan can be continuously and dynamically adjusted and optimized based on students' actual implementation feedback and fluctuations in vital signs.

[0053] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A student nutrition balance monitoring and personalized intervention system based on multi-source data fusion, characterized in that, include: The multi-source heterogeneous sensing and acquisition module acquires dietary images, continuous sensor signals and regular physical examination records of the target object within a preset time period, and generates raw physiological and behavioral datasets. The feature-level fusion extraction module extracts the nutritional intake, energy consumption, and body shape indicators of the target object based on the original physiological and behavioral dataset, calculates the historical rate of change of the body shape indicators, and aligns the nutritional intake, energy consumption, body shape indicators, and historical rate of change of the indicators according to their respective timestamps to generate a multidimensional metabolic feature matrix and a historical state sequence. The time-series state prediction module, based on the multidimensional metabolic feature matrix and the historical state sequence, inputs the multidimensional metabolic feature matrix and the historical state sequence into a pre-trained time-series deep learning model and outputs a physiological development prediction trajectory. The dynamic threshold adaptive evaluation module calculates an adaptive baseline based on the predicted physiological development trajectory and the historical index change rate, combined with the body posture index. It then compares the derivatives of the predicted physiological development trajectory and the adaptive baseline to generate state warning parameters. The decision-level fusion and closed-loop intervention module, based on the multidimensional metabolic feature matrix and the state warning parameters, compares the multidimensional metabolic feature matrix and the state warning parameters through expert system rules, outputs an intervention plan, and inputs the subsequent continuous sensing signals of the target object into the multi-source heterogeneous sensing and acquisition module.

2. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the multi-source heterogeneous sensing acquisition module, the generation of the original physiological and behavioral dataset specifically includes the following steps: Extract the dietary images, the continuous sensor signals, and the timestamps from the periodic physical examination records, respectively. Based on multiple timestamps, the dietary images, the continuous sensor signals, and the regular physical examination records are correlated in a time series to generate time series correlation data. The time-series correlated data are combined to generate the original physiological and behavioral dataset.

3. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the feature-level fusion extraction module, the extraction of the target object's nutrient intake, energy expenditure, and body shape indicators, and the calculation of the historical rate of change of the body shape indicators, specifically includes the following steps: Analyze the periodic physical examination records in the original physiological and behavioral dataset to extract body posture indicators; Perform time-series difference calculations on the aforementioned body posture indicators to generate historical indicator change rates; A visual semantic segmentation algorithm is used to extract pixel features and calculate volume mapping on the dietary images in the original physiological and behavioral dataset to generate nutrient intake. A time-series signal analysis algorithm is used to perform frequency domain transformation and feature matching on the continuous sensor signals in the original physiological and behavioral dataset to generate energy consumption data.

4. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the feature-level fusion extraction module, the fusion to generate a multidimensional metabolic feature matrix and a historical state sequence specifically includes the following steps: Obtain the timestamps corresponding to the nutrient intake, energy expenditure, body shape indicators, and the rate of change of the historical indicators; Based on the timestamp, the nutrient intake and energy consumption at the same time point are combined to generate a metabolic feature vector; Multiple metabolic feature vectors are stacked in multiple dimensions to generate a multidimensional metabolic feature matrix. Based on the timestamp, the change rates of the body posture indicators and the historical indicators are sorted and correlated in a time sequence to generate a historical state sequence.

5. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the time-series state prediction module, the step of inputting the multidimensional metabolic feature matrix and the historical state sequence into a pre-trained time-series deep learning model and outputting a physiological development prediction trajectory specifically includes the following steps: Temporal features are extracted from the multidimensional metabolic feature matrix to generate metabolic temporal features; Trend features are extracted from the historical state sequence to generate growth and development trend features; The metabolic time-series features are concatenated with the growth and development trend features to generate a joint input feature matrix; The joint input feature matrix is ​​input into a pre-trained time series deep learning model based on a long short-term memory network to perform temporal state evolution calculation, outputting a multi-time step predicted state vector sequence, and combining the predicted state vector sequences to generate a physiological development prediction trajectory.

6. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 5, characterized in that, The training steps for the pre-trained time-series deep learning model based on a long short-term memory network include: Obtain the joint input feature matrix of multiple sets of samples and the corresponding labels of real physiological developmental trajectories; The multiple sets of samples are jointly input into the feature matrix and then input into the initial long short-term memory network model to output the sample predicted trajectory. The deviation between the predicted trajectory of the sample and the actual physiological development trajectory label is calculated to generate a temporal loss value; Based on the time-series loss value, the weight parameters of the initial long short-term memory network model are updated using the backpropagation algorithm to generate a pre-trained time-series deep learning model.

7. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the dynamic threshold adaptive evaluation module, the step of comparing the physiological development prediction trajectory with the adaptive baseline to generate state warning parameters specifically includes the following steps: The time-dimensional derivative of the predicted physiological development trajectory is calculated to generate a sequence of predicted trajectory change rates; The time-dimensional derivative of the adaptive baseline is calculated to generate a baseline change rate sequence; The difference between the predicted trajectory rate of change sequence and the baseline rate of change sequence is calculated to generate a time-series deviation matrix. The peak features in the time-series deviation matrix are extracted as state warning parameters.

8. The student nutrition balance monitoring and personalized intervention system based on multi-source data fusion according to claim 1, characterized in that, In the decision-level fusion and closed-loop intervention module, the step of comparing the multidimensional metabolic feature matrix with the state warning parameters using expert system rules and outputting the intervention plan specifically includes the following steps: Extract the state features of the current time node from the multidimensional metabolic feature matrix to generate the current metabolic state feature vector; Risk feature mapping is performed on the aforementioned status warning parameters to generate a warning status feature vector; The current metabolic state feature vector is combined with the warning state feature vector to generate a comprehensive risk assessment vector; Based on the comprehensive risk assessment vector query, a pre-set intervention strategy rule base is used for matching and calculation, and an intervention plan is output.

Images