Intelligent data analysis method for pension information service system

Abstract

The invention belongs to the technical field of intelligent old-age care and artificial intelligence crossing, and discloses an intelligent data analysis method for an old-age care information service system, which comprises the following steps: S1, collecting multi-modal time sequence data in real time from a heterogeneous sensing terminal deployed in the living environment of the old, and sending the multi-modal time sequence data to a server; the multi-modal time series data at least comprises physiological data, behavior data, environment data and social interaction data. According to the method, through multi-scale time sequence feature extraction and cross-modal correlation analysis, subtle modes and early symptoms before risks occur can be captured from massive multi-source data, the accuracy and timeliness of risk prediction are remarkably improved, and conversion from passive response to active intervention is achieved; through the real-time semantic matching of the dynamic risk portrait and the old-age service resource knowledge graph and the decision model based on multi-target reinforcement learning, the automatic and intelligent closed loop from risk identification to resource recommendation is realized, and the problem of broken link assessment and intervention is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wisdom and artificial intelligence, and particularly relates to a data intelligent analysis method for an old-age information service system. BACKGROUND

[0002] With the intensification of global population aging trend, it has become an urgent need for social development to build an intelligent old-age service system by using information technology. In the prior art, many old-age information service systems mainly focus on data collection, aggregation and static health assessment.

[0003] For example, the patent application No. CN202410245134.3 discloses a data intelligent analysis method for an old-age information service system, which is mainly based on multi-source data fusion for health state assessment, representing a typical technical path in this field. However, through in-depth analysis, the prior art generally has the following limitations: the risk perception is shallow and lagging, and the subtle, cross-modal precursor patterns before the occurrence of risks cannot be deeply mined; the evaluation and intervention are disconnected, and there is a lack of automatic and intelligent decision-making bridge between the evaluation results output by the system and the specific old-age service resource scheduling; the strategy is static and isolated, and the service recommendation strategy is usually based on predefined rules, lacking consideration of the dynamically changing environment, resource state and long-term intervention effect; the system is non-evolutionary, and once the traditional system model is deployed, its decision logic is relatively fixed, and cannot be continuously optimized with the accumulation of service data and the changes of the needs of the elderly. SUMMARY

[0004] In order to overcome the technical defects of risk early warning lag, passive service response and isolated and rigid decision-making in the prior art, the present application provides a data intelligent analysis method for an old-age information service system, which can independently perceive, intelligently predict, accurately match and dynamically optimize.

[0005] The technical solution adopted by the present application is as follows: a data intelligent analysis method for an old-age information service system, comprising the following steps:

[0006] S1: Real-time collection of multi-modal time series data from heterogeneous perception terminals deployed in the living environment of the elderly, prior to collection, the explicit authorization of the elderly or their guardians is required, and the principle of minimum necessity is followed. For data involving privacy (such as call content), localized processing or differential privacy technology is used to ensure that the data is desensitized before being transmitted to the cloud. The multi-modal time series data includes at least physiological data, behavior data, environmental data, and social interaction data, including call records, social application usage duration and frequency collected through intelligent terminals, and emotional analysis of call content through natural language processing technology, extracting emotional polarity as a social health indicator, and adaptive cleaning and spatio-temporal alignment fusion of the multi-modal time series data to form standardized multi-modal time series data stream;

[0007] S2: Construct a multi-scale time series feature extractor to extract micro, meso and macro time scale risk features from the standardized multi-modal time series data stream, and use graph attention network to mine the internal correlation between different modal features to generate cross-modal correlation features, which together constitute a high-dimensional dynamic feature vector;

[0008] S3: Input the high-dimensional dynamic feature vector into a pre-trained stacked integrated risk prediction model, which is composed of base learner layers including time series convolution network, bidirectional long short-term memory network and isolation forest, and a gradient boosting machine meta-learner layer for meta-feature learning, outputting calibrated risk probabilities for multiple categories of risks including falls, health deterioration, and loneliness;

[0009] S4: Based on the calibrated risk probability, combine the static profile and historical preferences of the elderly to generate a structured dynamic risk portrait, which represents the current risk state in vector form, and generates an interpretable report containing the main risk contributing factors through a visualization engine;

[0010] S5: Real-time semantic matching of the dynamic risk portrait vector with a dynamically evolving knowledge graph of elderly care service resources, the dynamic evolution of the knowledge graph of elderly care service resources is realized through vector embedding of nodes and relationships through graph neural network, and direct and indirect related service resource nodes are discovered through multi-hop reasoning;

[0011] S6: Based on the matching results of S5, a multi-objective reinforcement learning-based decision model is used to generate personalized resource active recommendation strategy, the decision model takes risk state and resource state as input, resource combination recommendation as action, and comprehensive reward of risk reduction utility, resource cost and user satisfaction feedback as optimization target, under the constraints of resource availability, geographical accessibility and elderly taboo, outputs the long-term optimal strategy and automatically generates executable work order.

[0012] Preferably, the isomer perception terminal in step S1 comprises medical-grade biosensors for collecting heart rate, blood pressure, and blood oxygen; a non-contact behavior perception module based on millimeter wave radar for monitoring the elderly falling and abnormal behavior of staying; and indoor environmental sensor units for temperature, humidity, light intensity, and harmful gas concentration; the isomer perception terminal transmits the collected data to the cloud server through the built-in Internet of Things communication module.

[0013] Preferably, in step S2, the construction method of the multi-scale time sequence feature extractor is: for the micro-scale risk feature, the short-term variability of biological parameters and the burstiness index of behavior events are extracted for hour to day level data; for the meso-scale risk feature, the sliding window statistics of the trend change of behavior patterns are calculated for week level data; and for the macro-scale risk feature, the long-term drift trend line of chronic health indicators is evaluated for month level data.

[0014] Preferably, in step S3, the training process of the stacked integrated risk prediction model adopts a dynamic focal loss function, which dynamically adjusts the weight according to the imbalance ratio of the risk category to which the sample belongs and the difficulty of sample prediction.

[0015] Preferably, in step S5, the update mechanism of the dynamically evolved knowledge graph of pension service resources includes: based on real-time service records and user feedback, the weight of the adaptive relationship between nodes is dynamically adjusted through an online learning algorithm; when new service resources are introduced, the semantic correlation degree between the new service resources and the existing nodes is automatically calculated by the dynamically evolved knowledge graph representation learning model of the pension service resources, and the new service resources are integrated into the graph structure.

[0016] Preferably, in step S6, the training environment of the decision model based on multi-objective reinforcement learning is a pension service simulation system constructed based on historical data, and the model is trained through a near-end policy optimization algorithm. After actual deployment, the model parameters are fine-tuned according to the real effect feedback of the resource active recommendation strategy execution through an online learning mechanism.

[0017] Preferably, in step S1, the collection and transmission process of the multi-modal time sequence data further includes a privacy protection mechanism, specifically: for voice interaction data involving personal privacy, feature extraction is performed on local devices using a federated learning framework, and only anonymized feature vectors are uploaded; all data transmission uses an end-to-end encryption protocol, and homomorphic encryption processing is performed when stored in the cloud to ensure that the data is not decrypted during use.

[0018] Preferably, the visualization engine in step S4 also supports multi-role adaptive view generation, which automatically adjusts the level of detail and presentation form of the report content according to the viewer's identity (the elderly, family members, medical staff, and administrators).

[0019] Preferably, step S6 further includes a service resource scheduling optimization module, which monitors the service resource state in real time based on the recommendation strategy, dynamically adjusts the work order execution sequence, and supports priority rescheduling of emergency events.

[0020] Preferably, the system further includes a model continuous monitoring and evaluation module, which periodically detects the performance degradation of the prediction model of step S3 and the decision model of step S6, and triggers an automatic retraining process to ensure long-term effectiveness of the models.

[0021] The beneficial effects of the present application are:

[0022] 1. By multi-scale time series feature extraction and cross-modal correlation analysis, subtle patterns and early signs before risk occurrence can be captured from massive multi-source data, significantly improving the accuracy and timeliness of risk prediction, and realizing the transition from passive response to active intervention;

[0023] 2. Through real-time semantic matching of dynamic risk portrait and knowledge graph of pension service resources, and decision model based on multi-objective reinforcement learning, an automatic and intelligent closed loop from risk identification to resource recommendation is realized, effectively solving the problem of broken chain between evaluation and intervention;

[0024] 3. Through the setting of the decision model, not only the immediate effect is considered, but also the long-term comprehensive reward is taken as the optimization target, which can generate sustainable and personalized service strategies under complex and variable real constraints, avoiding the rigidity of static rules;

[0025] 4. The knowledge graph of pension service resources and the decision model both have online learning and dynamic updating mechanisms, which can continuously evolve with data accumulation, user feedback and introduction of new resources, making the system have long-term service capability and adaptability to meet the changing needs of the elderly;

[0026] 5. Through the visual engine, an interpretable report containing the main risk contribution factors is generated, so that the elderly and their families, caregivers can understand the basis of system decision-making, and enhance their trust and acceptance of the system. BRIEF DESCRIPTION OF DRAWINGS

[0027] One or more embodiments are illustrated by way of example in the accompanying drawings that are not intended to be limiting of the embodiments. Like reference numbers in the drawings indicate like elements unless otherwise specified. The drawings in which:

[0028] Figure 1 is a flowchart of a data intelligent analysis method for a pension information service system according to the present application. DETAILED DESCRIPTION

[0029] For the purposes of the present application, the technical solutions and advantages will be clearer, the embodiments of the present application will be described in detail below with reference to the drawings. However, those skilled in the art can understand that in the embodiments of the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, even without these technical details and based on various changes and modifications of the following embodiments, the technical solutions claimed in the claims of the present application can be implemented.

[0030] As Figure 1 shown, the present embodiment provides a data intelligent analysis method for an old-age information service system, comprising the following steps:

[0031] S1: Multimodal time series data acquisition and preprocessing, this step aims to collect raw data in real time from a plurality of heterogeneous sensing terminals deployed in the living environment of the elderly, including but not limited to:

[0032] Medical-grade biosensors: such as wearable devices (smart bracelet, chest patch) for collecting physiological parameters such as heart rate, blood pressure, blood oxygen saturation, etc., the sampling frequency can be set according to the needs (such as every 5 minutes) ;

[0033] Non-contact behavior sensing module: based on millimeter wave radar technology, without the need for the elderly to wear devices, it can monitor the activity trajectory of the elderly in the room, fall events, long-time retention (such as more than a set threshold in the bathroom) and other abnormal behaviors, millimeter wave radar has the advantages of good privacy protection, not affected by light, etc.

[0034] Environmental sensor unit: distributed in key positions in the living room (such as bedroom, living room, kitchen), used to monitor environmental parameters such as temperature and humidity, light intensity, carbon dioxide concentration, harmful gas (such as CO) concentration, etc.

[0035] All the above sensing terminals are built-in Internet of Things communication modules (such as NB-IoT, LoRa, Wi-Fi), and the collected data is transmitted in real time to the cloud server or edge computing node using the AES-256 encryption algorithm.

[0036] After data collection, data preprocessing is performed, including adaptive cleaning and spatio-temporal alignment fusion:

[0037] Adaptive cleaning: set cleaning rules for different data types. For physiological data, Z-score anomaly detection based on sliding window is used to remove spikes caused by signal interference; for behavior data, morphological filtering is used to remove short-time noise; for environmental data, range checking (such as temperature should be between 0-50 degrees Celsius) and continuous consistency checking are used.

[0038] Temporal-spatial alignment fusion: Due to the heterogeneity of data sources and different sampling frequencies, temporal-spatial alignment is needed. First, the timestamps of all data are unified to UTC time, and the low-frequency data is aligned to the time axis with the highest sampling frequency through interpolation (such as linear interpolation or spline interpolation); second, for spatial information, the sensor position is associated with the old person's activity area through the room layout map to form a unified spatial reference system; finally, the cleaned and aligned physiological, behavioral, environmental and social interaction data (such as call frequency and social application usage collected through smart devices) are fused into standardized multi-modal time series data streams, which are stored in the form of time series, and each time point contains synchronous observation values of all modalities.

[0039] The collection and transmission process of the multi-modal time series data also includes a privacy protection mechanism, specifically: for voice interaction data involving personal privacy, feature extraction is performed on the local device using a federated learning framework, and only anonymized feature vectors are uploaded; all data transmission uses an end-to-end encryption protocol, and homomorphic encryption processing is performed when stored in the cloud to ensure that the data is not decrypted during use.

[0040] S2: Multi-scale time series feature extraction and cross-modal correlation analysis, this step extracts risk features of different time granularities from the standardized data stream by constructing a multi-scale time series feature extractor, which includes three scale feature extraction units of micro, meso and macro, and a graph attention network.

[0041] Micro-scale risk feature extraction unit: for high-frequency data from hours to days. The extracted features include: short-term variability of biological parameters: for example, calculate the time domain indicators (SDNN, RMSSD) and frequency domain indicators (LF, HF power) of heart rate variability (HRV), and the short-term fluctuation standard deviation of blood pressure; suddenness indicators of behavioral events: for example, the acceleration peak value and duration of a fall event, the switching frequency of activity states (such as the number of sitting-walking transitions);

[0042] This micro-scale risk feature extraction unit usually uses short-time Fourier transform or wavelet transform to capture the transient features of the signal.

[0043] Meso-scale risk feature extraction unit: for weekly data, focusing on the trend changes of behavior patterns, using a sliding window (window size can be 7 days, step size 1 day) to calculate statistics such as: moving average and moving standard deviation of daily activity volume (estimated by radar monitoring of moving trajectory length); weekly trend slope of sleep quality indicators (such as total sleep time, sleep onset time variability); autocorrelation coefficient of social interaction frequency (such as daily call frequency).

[0044] Macro-scale risk feature extraction unit: For monthly data, assess long-term health trends. Fit a long-term drift trend line for chronic health indicators (e.g., weight, blood glucose trends) using linear regression or exponential smoothing, and calculate the slope, intercept, and variance of the residuals to reflect the trend of deterioration or improvement of long-term health status.

[0045] The extracted features of each scale are then input into a graph attention network for cross-modal correlation analysis. In this process, the features of different modalities are treated as nodes in a graph, and the edges between nodes represent the potential correlation between modalities. The graph attention network uses a two-layer graph attention network structure, with 8 heads and a hidden dimension of 64. The GraphSAGE sampler uses mean aggregation with 10 neighbor samples. The graph attention network learns the importance weights between nodes through self-attention mechanisms, aggregates neighbor node information, and generates cross-modal correlation features rich in contextual information. For example, it may find that there is a strong correlation between "nighttime heart rate variability reduction" and "daytime activity reduction", which itself may be an important risk precursor.

[0046] Finally, all scale features and cross-modal correlation features are concatenated and reduced in dimension (e.g., using principal component analysis PCA or autoencoder) in the feature fusion layer to form a unified high-dimensional dynamic feature vector, which serves as input to the downstream prediction model.

[0047] S3: Stacked ensemble risk prediction. The high-dimensional dynamic feature vector is input into a pre-trained stacked ensemble risk prediction model, which outputs calibrated multi-class risk probabilities. The model uses a two-layer structure: base learner layer and meta-learner layer.

[0048] Base learner layer: composed of three heterogeneous base learners, designed to capture data patterns from different perspectives: temporal convolutional network base learner: uses dilated convolution to capture long-term temporal dependencies, particularly suitable for extracting periodic physiological patterns; bidirectional long short-term memory network base learner: learns temporal context from both forward and backward directions, good at capturing forward and backward associations in behavior sequences; Isolation Forest base learner: an unsupervised anomaly detection algorithm for identifying rare events or abnormal patterns that deviate significantly from normal patterns, serving as a complement to other supervised learners. The size of the convolution kernel of the temporal convolutional network ranges from , with dilated coefficients set to 1, 2, 4, 8; the hidden layer dimension of the bidirectional long short-term memory network is 128; the number of trees in the Isolation Forest is set to 100, and the subsampling number is 256.

[0049] Each base learner independently trains the input features and outputs the initial prediction probabilities for each risk category (fall, health deterioration, loneliness).

[0050] Meta-learner layer: The prediction results of base learners (i.e., meta-features) are concatenated and input into the meta-feature learning layer for further feature transformation, and then the final decision is made by the gradient boosting machine meta-learner. GBM can effectively learn the complex combination relationship between the predictions of base learners by integrating multiple weak decision trees, and improve the overall generalization ability. The maximum depth of the tree of the meta-learner gradient boosting machine is 6, and the learning rate is 0.1.

[0051] The training process of the stacked ensemble risk prediction model uses a dynamic focal loss function to solve the class imbalance problem commonly found in elderly care data (e.g., fall events are much less common than normal activities); this loss function is improved based on the standard focal loss, and its weight factor α and modulation factor γ are no longer fixed values, but are dynamically adjusted according to the real-time distribution ratio of each class in each training batch and the difficulty level (confidence) of each sample predicted by the current model; the formula can be expressed as:

[0052]

[0053] wherein, is the prediction probability of the model for the true class is the weight calculated based on the frequency of class in the current batch, is the focus parameter dynamically adjusted based on the average confidence of the sample, , is the frequency of class in the current batch; , the average confidence is the average prediction probability of the samples belonging to class in the current batch. This allows the model to continuously focus on difficult samples and minority class samples during training, significantly improving the detection ability of rare risks.

[0054] Finally, the model outputs the risk probability calibrated by Platt Scaling or Isotonic Regression, ensuring that the probability value is consistent with the true risk frequency and improving the reliability of the decision.

[0055] S4: Dynamic risk profile generation and explainability report, based on the calibrated risk probability, combined with the old person's static profile (such as age, chronic disease history, allergic drugs) and historical preferences (such as preference for certain service types, refusal history), a structured dynamic risk profile is generated.

[0056] ​Risk profile vectorization: In the risk profile vectorization module, the probabilities of each category of risk, the key indicators of the static profile (encoded as one-hot or embedding vectors), and the historical preference vector are concatenated and compressed and fused through a fully connected neural network, outputting a fixed-length dynamic risk profile vector; this vector comprehensively represents the current comprehensive risk status of the elderly;

[0057] Interpretable factor analysis: The interpretable factor analysis module uses post-hoc interpretability techniques such as SHAP or LIME to analyze which input features (such as "average nighttime heart rate in the past week" and "number of falls yesterday") contribute most to the final risk probability in the prediction model of S3; these contributing factors are extracted and stored in a structured manner;

[0058] Visual report generation: The visual report generator uses web front-end technologies (such as ECharts and D3.js) to convert the dynamic risk profile vector and the main risk contributing factors into intuitive visual reports. The report may include: risk level dashboard: showing the current level (low, medium, high) of each risk; trend comparison chart: showing the comparison of key physiological / behavioral indicators with historical averages; contribution factor bar chart: clearly listing the top N key factors affecting risk decision-making and their contribution.

[0059] The visualization engine also supports multi-role adaptive view generation, automatically adjusting the level of detail and presentation form of the report content according to the viewer's identity (elderly, family members, medical staff, administrators).

[0060] The report can be viewed through a PC, tablet, or family member's mobile phone APP, greatly enhancing the transparency of the system and the trust of users.

[0061] S5: Semantic matching of service resources based on dynamic knowledge graph, this step aims to convert the abstract risk profile into specific service resource recommendations. The core is a dynamically evolving knowledge graph of elderly care service resources.

[0062] Knowledge graph construction: Knowledge graph nodes represent various service resources, such as "home visits by medical staff", "psychological counseling", "meal delivery service", "rehabilitation training", and "emergency call"; relationship edges represent semantic associations between nodes, such as "synergistic effect", "substitution relationship", and "precondition"; the initial graph can be constructed by domain experts or extracted from structured service databases.

[0063] Vector embedding and updating: Use graph neural networks (such as GraphSAGE) to map nodes and relationships in the knowledge graph to a low-dimensional vector space; update mechanisms include:

[0064] Based on real-time service records and user feedback, the weight of the relationship edge is dynamically adjusted through online learning algorithms (such as stochastic gradient descent); for example, if multiple records show that the user's satisfaction is high after the joint recommendation of "psychological counseling" and "social activity organization", the weight of the "collaboration" relationship between the two nodes will automatically increase;

[0065] When a new service resource (such as "drone delivery of medicine") is introduced, the system uses its attribute description to automatically calculate its semantic similarity with existing nodes through a graph representation learning model, and integrates it into the graph as a new node, while establishing the corresponding relationship edge.

[0066] Semantic matching and multi-hop reasoning: input the dynamic risk profile vector generated by S4 into the semantic matching engine to calculate the cosine similarity with all service resource node vectors in the knowledge graph; at the same time, start the multi-hop reasoning module to explore indirectly related resources; for example, for "high risk of loneliness", the direct match may be "community activity recommendation", while through two-hop reasoning (loneliness -> need social interaction -> can participate in interest groups) may find "calligraphy interest class" as an indirect, but may be a more suitable resource for user historical preferences.

[0067] S6: personalized strategy generation and work order execution based on multi-objective reinforcement learning, based on the matched service resource candidate set, a decision model based on multi-objective reinforcement learning is used to generate a long-term recommendation strategy that considers effectiveness, cost, and user satisfaction.

[0068] Decision model construction:

[0069] State space: includes the current dynamic risk profile vector, the state of available service resources (such as busy degree, geographical location), environmental context (such as weather, time);

[0070] Action space: defined as selecting one or more combinations of service resources from the resource candidate set (i.e. recommendation strategy);

[0071] Reward function: a multi-objective function, including: risk reduction utility: the risk dimension of the next period risk profile vector decreases after the execution of the recommendation; resource cost: economic cost, time cost; user satisfaction feedback: quantified through post-questionnaire or implicit feedback (such as service completion rate, complaint rate);

[0072] Constraints: including resource availability (whether the nurse is on duty), geographical accessibility (whether the service distance is within the old person's acceptance range), and old person taboo (such as allergy to certain food or activities).

[0073] Model training and deployment:

[0074] Offline training: In the simulation system of elderly care services built based on historical data, the proximal policy optimization algorithm is used for training; the PPO algorithm updates the step length by clipping the policy, ensuring the stability of the training, and can effectively handle high-dimensional state and action space;

[0075] Online learning and fine-tuning: After actual deployment, the system establishes an online learning cycle, and online learning triggers fine-tuning every 24 hours, with an experience replay pool capacity of 10,000, using the Adam optimizer and a learning rate of 1e-5; whenever a recommended strategy is executed and produces real feedback (such as risk changes, user ratings), the interaction experience (state, action, reward, new state) will be stored in the experience replay pool and used regularly to fine-tune the PPO model parameters, so that the decision-making model can continuously adapt to changes in the needs of the elderly and the evolution of the service ecosystem.

[0076] It also includes a service resource scheduling optimization module that monitors service resource status in real time based on recommended strategies, dynamically adjusts the execution order of work orders, and supports priority rescheduling for emergencies.

[0077] The system also includes a model continuous monitoring and evaluation module that regularly performs performance degradation detection on the prediction model of step S3 and the decision-making model of step S6, and triggers an automated retraining process to ensure long-term effectiveness of the models.

[0078] Work order generation: The optimal action (i.e., resource combination) output by the decision-making model is passed to the work order generation and execution module. This module automatically generates structured executable work orders containing service content, execution time, responsible personnel (such as caregivers, doctors), required materials, and other information, and automatically distributes them to the corresponding service provider system through a workflow engine, completing the last mile from decision-making to execution.

[0079] In actual deployment, the system and method described in the present application can achieve 7x24 hours of uninterrupted operation; the full process automation from data collection to work order generation greatly reduces the burden of manual care and achieves precise and personalized elderly care services; through continuous learning and evolution, the system can continuously optimize its prediction and decision-making capabilities, providing increasingly personalized care and services to the elderly.

[0080] Meanwhile, the contents not described in detail in the present specification are all existing technologies known to those skilled in the art.

[0081] Those skilled in the art can understand that the above embodiments are specific examples for implementing the present application, and in actual applications, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. A data intelligent analysis method for a pension information service system, characterized by: The method comprises the following steps: S1: Real-time collection of multi-modal time series data from heterogeneous perception terminals deployed in the living environment of the elderly, the multi-modal time series data at least including physiological data, behavior data, environmental data and social interaction data, and adaptive cleaning and spatio-temporal alignment fusion of the multi-modal time series data to form a standardized multi-modal time series data stream; S2: Construction of a multi-scale time series feature extractor to extract micro, meso and macro three time scale risk features from the standardized multi-modal time series data stream, and use a graph attention network to mine the internal correlation between different modal features to generate cross-modal correlation features, which together constitute a high-dimensional dynamic feature vector; S3: Inputting the high-dimensional dynamic feature vector into a pre-trained stacked integrated risk prediction model to output calibrated risk probabilities of multiple categories of risks including falls, health deterioration and loneliness; S4: Based on the calibrated risk probability, combining the static profile and historical preferences of the elderly, a structured dynamic risk portrait is generated, which represents the current risk state in vector form, and an interpretable report containing the main risk contributing factors is generated through a visualization engine; S5: Real-time semantic matching of the dynamic risk portrait vector with a dynamically evolving knowledge graph of elderly service resources, and through multi-hop reasoning, directly and indirectly related service resource nodes are found; S6: Based on the matching results of S5, a personalized resource active recommendation strategy is generated by using a multi-objective reinforcement learning based decision model, and an executable work order is automatically generated.

2. The data intelligent analysis method for the aged information service system according to claim 1, characterized in that: The heterogeneous perception terminal in step S1 includes medical-grade biological sensors for collecting heart rate, blood pressure and blood oxygen; a non-contact behavior perception module based on millimeter wave radar for monitoring the elderly's fall and abnormal behavior of staying still; And indoor distributed temperature and humidity, light intensity, harmful gas concentration environmental sensor units; the heterogeneous perception terminal transmits the collected data to the cloud server through the built-in Internet of Things communication module. 3.The data intelligent analysis method for the aged information service system according to claim 1, wherein: In step S2, the construction method of the multi-scale time series feature extractor is: The micro-scale risk features are short-term variability of biological parameters and suddenness indicators of behavior events for hourly to daily data; the meso-scale risk features are sliding window statistics of behavior pattern trend changes for weekly data; The macro-scale risk features are long-term drift trend lines of chronic health indicators for monthly data.

4. The data intelligent analysis method for the aged information service system according to claim 1, characterized in that: In step S3, the training process of the stacked integrated risk prediction model uses a dynamic focal loss function, which dynamically adjusts the weights according to the imbalance ratio of the risk categories to which the samples belong and the difficulty of sample prediction.

5. The data intelligent analysis method for the aged information service system according to claim 1, characterized in that: In step S5, the update mechanism of the dynamically evolving knowledge graph of elderly service resources includes: based on real-time service records and user feedback, the weights of the adaptive relationship between nodes are dynamically adjusted through an online learning algorithm; when new service resources are introduced, the semantic correlation degree between the new service resources and the existing nodes is automatically calculated through a dynamically evolving knowledge graph representation learning model of the knowledge graph of elderly service resources and integrated into the graph structure.

6. The data intelligent analysis method for the aged information service system according to claim 1, characterized in that: In step S6, the training environment of the multi-objective reinforcement learning-based decision model is a pension service simulation system constructed based on historical data. The model is trained by a proximal policy optimization algorithm, and after actual deployment, the model parameters are fine-tuned according to the real effect feedback of the resource active recommendation strategy execution through an online learning mechanism.

7. The data intelligent analysis method for the aged information service system according to claim 1, characterized in that: In step S1, the collection and transmission process of the multi-modal time series data also includes a privacy protection mechanism. Specifically, for voice interaction data involving personal privacy, feature extraction is performed on the local device using a federated learning framework, and only anonymized feature vectors are uploaded. End-to-end encryption protocols are used for all data transmission, and homomorphic encryption is performed when storing data in the cloud to ensure that the data is not decrypted during use. 8.The data intelligent analysis method for the aged information service system according to claim 1, wherein: The visualization engine in step S4 also supports multi-role adaptive view generation, automatically adjusting the level of detail and presentation form of the report content according to the viewer's identity (elderly, family members, medical staff, administrators). 9.The data intelligent analysis method for the aged information service system according to claim 1, wherein: Step S6 also includes a service resource scheduling optimization module that monitors the service resource state in real time based on the recommendation strategy, dynamically adjusts the execution order of work orders, and supports priority rescheduling of emergency events. 10.The data intelligent analysis method for the aged information service system according to claim 1, wherein: The system also includes a model continuous monitoring and evaluation module that periodically detects the performance degradation of the prediction model in step S3 and the decision model in step S6, and triggers an automated retraining process to ensure long-term effectiveness of the models.

Images