Lightweight data fusion and privacy security intelligent recommendation system for home health

Abstract

The application discloses a lightweight data fusion and privacy security intelligent recommendation system for home health, and relates to the fields of health data processing, lightweight intelligent system and privacy computing technology. The system comprises a health data standardization processing module, a multi-source feature fusion and quantization module, a lightweight structured health knowledge base module, a three-layer constraint intelligent recommendation module, a multi-scale time series health trend analysis module, a lightweight federated privacy computing module and a local lightweight intelligent interaction terminal. The system improves data utilization efficiency through multi-source health data standardization and feature fusion, realizes safe and reliable personalized recommendation by adopting a three-layer constraint mechanism, completes health trend judgment through multi-scale time series analysis, and realizes privacy protection model optimization without data leaving the local area by relying on horizontal federated learning. The application has high lightweight degree, low computing power occupation and can be adapted to low computing power terminals such as small programs, the system is self-controllable and does not depend on third-party large models, the output content is lifestyle reference suggestions which are non-diagnostic and non-therapeutic, and the system can be widely applied to scenes such as home health management and community health services.

Description

Technical Field

[0001] This invention relates to the fields of health data processing, lightweight intelligent systems, privacy computing, and time-series data analysis. Specifically, it relates to a non-diagnostic, non-therapeutic health status assessment and lifestyle intelligent recommendation system suitable for low-computing-power mobile terminals, based on multi-source health data fusion and privacy protection. Background Technology

[0002] With the increasing demand for home-based health management, mobile health monitoring and intelligent recommendation products are gradually becoming more widespread. However, existing health management systems still have several technical shortcomings in practical applications. For example, the method and system for managing the lifestyle of diabetic patients that combines traditional Chinese medicine knowledge, proposed in patent CN121839037A, can provide health maintenance, recipes, and music recommendations based on TCM syndromes. However, this solution is only designed for a single disease (diabetes) and only analyzes and recommends specific health data. It lacks a universal fusion process for multi-source heterogeneous health data in daily home scenarios, resulting in limited data types, restricted applicability, and low overall data utilization efficiency. Furthermore, many systems use a centralized cloud computing model for health data analysis, requiring all user's original health data to be uploaded to the server, posing a significant risk of personal privacy data leakage and insufficient data security protection capabilities. In addition to the above, current health management systems suffer from several common technical problems: First, the analytical models used in these systems are complex and have a large number of parameters, requiring high computing power and storage resources from terminal devices. This makes them unsuitable for stable operation on low-computing-power devices such as mini-programs and portable smart terminals, limiting their applicability. Second, health recommendation schemes often use general, fixed rules for generation, resulting in highly homogenized recommendations. They fail to incorporate personalized filtering based on individual user health characteristics and lack security constraints, leading to insufficient rationality and practicality of the recommendations. Third, health data collected from multiple terminals is isolated, making it impossible to achieve collaborative model optimization while protecting local data privacy, resulting in low system iteration and upgrade efficiency. Fourth, most systems heavily rely on third-party large model platforms for inference calculations, leading to issues such as uncontrollable intellectual property rights, high data compliance risks, weak customization and expansion capabilities, and poor system autonomy and controllability. Summary of the Invention

[0003] (a) Technical problems to be solved 1. The heterogeneous health data from multiple sources has significant differences in format and dimension, lacks a complete standardized processing procedure, and cannot achieve unified feature fusion and efficient utilization, resulting in insufficient data processing accuracy. 2. Traditional health analysis models are highly complex, resource-intensive, and require high computing power, making them unsuitable for lightweight mobile terminals such as mini-programs and difficult to achieve real-time and efficient operation. 3. Existing systems mostly adopt centralized cloud computing, with inadequate mechanisms for protecting the privacy of users' sensitive health data. There are risks of leakage and misuse during the data upload and storage process, and a lack of secure distributed computing architecture. 4. The health recommendations lack personalization and fail to incorporate multi-level constraints and screening based on user characteristics and health and safety rules, resulting in insufficient targeting and safety of the recommendations. 5. Multi-terminal data cannot achieve collaborative modeling under privacy protection, resulting in low model optimization efficiency and high difficulty in overall system iteration and upgrade. 6. Over-reliance on third-party large model platforms leads to uncontrollable system intellectual property rights, weak customization and expansion capabilities, and an inability to guarantee data compliance and independent controllability. (II) Technical Solution This invention discloses a lightweight data fusion and privacy-preserving intelligent recommendation system for home health, including a health data standardization processing module, a multi-source feature fusion and quantification module, a lightweight structured health knowledge base module, a three-layer constraint-based intelligent recommendation module, a multi-scale time-series health trend analysis module, a lightweight federated privacy computing module, and a local lightweight intelligent interactive terminal. The overall system workflow is as follows: A local lightweight intelligent interactive terminal collects multi-dimensional raw health data, including basic user information, physiological indicators, daily routines, exercise and diet records, and behavioral feedback. This data is then transmitted to a health data standardization processing module. This module performs missing value interpolation, outlier filtering, and dimension normalization on the raw data to form standardized health data with a unified format. The processed data is then fed into a multi-source feature fusion and quantization module, where discrete and continuous data are encoded and converted into vectors respectively. Redundant information is removed through feature filtering to construct a comprehensive health feature vector. The system leverages a lightweight, structured health knowledge base to store health behavior entries, adaptation rules, contraindications, and safety thresholds. It then performs inference operations based on comprehensive health feature vectors. A three-layer constrained intelligent recommendation module sequentially completes feature matching, safety filtering, and personalized ranking to generate lifestyle recommendations tailored to the user's individual condition. Simultaneously, a multi-scale time-series health trend analysis module retrieves the user's historical health data, using multi-time-window analysis to fit health status trends and determine fluctuations, outputting a probability result for the health trend. The system builds a horizontal federated learning architecture based on a lightweight federated privacy computing module. Each local terminal independently completes model inference and parameter updates, encrypting only the model update parameters before uploading them, while the original health data remains on the local terminal throughout the process. Privacy protection is further enhanced through gradient pruning and noise perturbation, while model pruning and quantization compression techniques are used to reduce the model's computational requirements, ensuring efficient system operation on low-computing-power terminals. This invention only provides health trend references and lifestyle guidance, and does not involve the diagnosis, treatment, or prevention of any diseases, nor does it constitute a disease diagnosis and treatment method. (III) Technical Effects This invention solves the problem of heterogeneous data not being used uniformly by standardizing and fusing multi-source health data, effectively improving the accuracy of health data processing and the reliability of subsequent analysis. It also adopts lightweight model optimization technology to significantly reduce system computing power and resource consumption, achieve native adaptation to mini-programs and low-computing-power mobile terminals, broaden the system's applicable scenarios, and improve ease of use. Through a lightweight federated privacy computing architecture, the system technically ensures that users' original health data does not leave their local devices, eliminating the risk of privacy leaks caused by data uploads, while balancing model collaborative optimization and data security. The three-layer constraint-based intelligent recommendation module combines user characteristics and security rules to achieve personalized and secure recommendation output, solving the problems of homogeneity and poor usability of traditional recommendations, and improving the rationality of recommended content. The system does not rely on third-party large model platforms throughout the entire process, and its overall architecture is independently controllable, avoiding intellectual property and data compliance risks, and possessing good customization and expansion capabilities. Multi-scale time series analysis enables quantitative trend judgment of health status, coupled with an interpretable output mechanism to improve the credibility of system results and user acceptance. At the same time, it clearly limits non-diagnostic and non-treatment application scenarios. Attached Figure Description

[0004] Figure 1 This is a system overall architecture block diagram according to an embodiment of the present invention; 1- Local lightweight intelligent interactive terminal; 2- Health data standardization processing module; 3- Multi-source feature fusion and quantification module; 4- Three-layer constraint-based intelligent recommendation module; 5- Lightweight structured health knowledge base module; 6- Multi-scale time-series health trend analysis module; 7- Lightweight federated privacy computing module; 8- Output non-diagnostic health trends and lifestyle recommendations. Figure 2 This is a flowchart of the three-layer constraint-based intelligent recommendation process according to an embodiment of the present invention; 1-User health feature vector; 2-Feature matching layer; 3-Rule security layer; 4-Lightweight reinforcement learning layer; 5-Output personalized lifestyle suggestions. Figure 3 This is a schematic diagram of a lightweight federated learning architecture according to an embodiment of the present invention; 1-First local terminal; 2-Second local terminal; 3-Third local terminal; 4-Multi-terminal collaborative computing node. Figure 4 This is a flowchart of the multi-scale time-series health trend analysis according to an embodiment of the present invention; 1-Historical health time series data; 2-7-day short-term window analysis; 3-30-day medium-to-long-term window analysis; 4-Moving average and linear regression calculations; 5-Output trend probability results. Detailed Implementation

[0005] (I) Data Collection and Standardization Combined with appendix Figures 1-4 The various modules of this invention work together in a coordinated manner, and the complete workflow is as follows: The local lightweight intelligent interactive terminal 1 collects multimodal data such as user heart rate, sleep duration, steps, dietary structure, and health behavior feedback, and initiates a data preprocessing process: missing data is filled using linear interpolation to ensure data continuity; values ​​exceeding the normal physiological range are identified as outliers and directly filtered out to avoid interfering with subsequent analysis. For different types of health data, the Min-Max normalization formula is used to uniformly map the data to the [0,1] numerical range, completing the unification of units and standardization of format, forming standardized input data, and providing a foundation for subsequent feature fusion. (ii) Multi-source feature fusion After standardization, the multi-type data enters the multi-source feature fusion and quantization module 3 for feature encoding transformation. Discrete features are processed using one-hot encoding, while continuous features are directly vectorized to achieve a unified representation of different types of data. Effective features highly correlated with health status are selected using the correlation coefficient method, and redundant and irrelevant features are eliminated. The retained effective features are then subjected to dimensional fusion processing to construct a comprehensive health feature vector with compact dimensions and strong stability, providing high-quality input for subsequent recommendation inference and trend analysis. (III) Lightweight Structured Health Knowledge Base The system has a built-in lightweight structured health knowledge base module 5, which uses a storage structure that combines rule tables and key-value pairs to store content such as suitable health behaviors, contraindications, behavioral conflict rules, safety thresholds, and compatibility relationships. It does not require a large amount of storage resources and has a fast query response speed. During the recommendation reasoning process, feature matching and rule verification are completed in real time, and recommended content that is not suitable for the user's current status, poses health risks, or violates safety rules is quickly filtered out, ensuring that the output results are safe and compliant. (iv) Three-layer constraint-based intelligent recommendation As attached Figure 2 As shown, the workflow of the three-layer constraint-based intelligent recommendation module 4 is as follows: The first layer is the feature matching layer 2: based on the comprehensive health feature vector, cosine similarity is used to calculate the degree of matching between user features and recommended item features, and the content with the highest matching degree is selected to generate a preliminary recommendation candidate set; The second layer is the rule security layer 3: it calls the lightweight structured health knowledge base module 5 to verify the recommended candidate set one by one, and removes content that does not comply with the security rules, poses health risks, or conflicts with the user's status, thus forming an optimized candidate set; The third layer is a lightweight reinforcement learning layer 4: It adopts a lightweight Q-Learning algorithm, using user suggestion compliance and historical feedback satisfaction as reward functions, dynamically sorts the optimization candidate set, and outputs the personalized lifestyle recommendation results with the highest priority and best suited to the user 5. (V) Multi-scale time-series health trend analysis As attached Figure 4 As shown, the multi-scale time-series health trend analysis module 6 retrieves the user's historical health time-series data 1, sets a 7-day short-term analysis window 2 and a 30-day medium-to-long-term analysis window 3, and conducts short-term fluctuation monitoring and medium-to-long-term trend fitting, respectively; through the moving average method and linear regression analysis calculation unit 4, it determines whether health indicators are rising, falling, or stable, calculates the corresponding trend probability, and outputs the results 5. This module only performs trend analysis at the data level and does not generate any disease diagnosis conclusions, nor does it provide any treatment plans or medical advice. (vi) Lightweight Federated Privacy Computing As attached Figure 3 As shown, the system achieves privacy protection through the lightweight federated privacy computing module 7. It adopts a horizontal federated learning architecture, where each local terminal 1, 2, and 3 independently completes model inference, training, and parameter updates. Only the model update parameters are encrypted and uploaded to the collaborative computing node 4. The original health data is stored on the local terminal throughout the process, achieving data usability without visibility. The gradient pruning threshold is set to 0.1~1.0 to limit the parameter update range, and Gaussian noise perturbation is added to further enhance the data privacy protection strength. Through model pruning and 8-bit quantization technology, the model size is compressed and the number of parameters is reduced, achieving lightweight system operation and adapting to low-computing-power mobile terminals and mini-program environments. (vii) Interpretable output While outputting health trend analysis results and lifestyle recommendations, the system simultaneously generates corresponding explanatory evidence, including the core characteristics of the matched users, the triggered health and safety rules, the data sources for trend judgment, and the recommendation logic. This presents the system's operational logic in a clear and easy-to-understand manner, ensuring that the results are traceable and explainable, thereby increasing users' acceptance and compliance with the system's output.

Claims

1. A lightweight data fusion and privacy-secure intelligent recommendation system for home health, characterized in that, The system includes a health data standardization and processing module, a multi-source feature fusion and quantification module, a lightweight structured health knowledge base module, a three-layer constrained intelligent recommendation module, a multi-scale time-series health trend analysis module, a lightweight federated privacy computing module, and a local lightweight intelligent interactive terminal. The system standardizes and fuses user multi-source health data, combines it with a lightweight structured health knowledge base to complete secure inference and personalized recommendations, outputs health trend probabilities through multi-scale time-series analysis, and achieves privacy-preserving model collaborative optimization under a federated learning architecture. The system outputs non-diagnostic and non-treatment lifestyle reference suggestions.

2. The system according to claim 1, characterized in that, The three-layer constrained intelligent recommendation module consists of: a feature matching layer, a rule-based safety layer, and a lightweight reinforcement learning layer.

3. The system according to claim 1, characterized in that, The multi-scale time-series health trend analysis module sets up a 7-day short-term analysis window and a 30-day medium-to-long-term analysis window. It uses moving averages and linear regression to determine the trend of health indicators and output probabilities, but does not provide disease diagnosis or treatment conclusions.

4. The system according to claim 1, characterized in that, The lightweight federated privacy computing module adopts a horizontal federated architecture. The user's original health data is stored locally and not uploaded to the cloud. Only the model update parameters are uploaded, and gradient pruning and Gaussian noise perturbation are used to enhance privacy protection.

5. The system according to claim 1, characterized in that, The system achieves model lightweighting through model pruning and 8-bit quantization, enabling it to run efficiently on low-computing-power mobile terminals such as mini-programs.

6. The system according to claim 1, characterized in that, The lightweight, structured health knowledge base uses rule tables and key-value pairs to store health behavior entries, adaptation relationships, safety constraints, and conflict rules, which are used for filtering recommended content and verifying the security of results.

7. The system according to claim 1, characterized in that, The overall system architecture is independent and controllable, does not rely on third-party large model platforms, and all inference and calculation processes are completed on local terminals or within the proprietary architecture.

8. The system according to claim 1, characterized in that, The system provides explanations and justifications when outputting recommendation results and trend analysis, making the reasoning process explainable and traceable.

Images