A Privacy-Preserving Joint Modeling System for Elderly Data Based on Federated Learning

Abstract

This invention discloses a privacy-preserving elderly care data joint modeling system based on federated learning, involving the interdisciplinary fields of artificial intelligence and privacy computing. The system includes: a data processing module that completes multi-source elderly care data collection and privacy desensitization, outputting a standardized dataset; a federated modeling module that uses horizontal federated learning to generate a global joint model; a multi-model fusion module that generates dynamic profiles of the elderly; a service recommendation module that uses heterogeneous association graphs to achieve service recall and adaptation, outputting service features and urgency; and a resource scheduling module that jointly optimizes and calculates decision values, generates scheduling schemes, and feeds back data. This invention securely breaks down data silos through privacy desensitization and federated learning, achieving collaborative modeling of multi-source data. By combining dynamic profiles, heterogeneous association graphs, and joint optimization scheduling, a closed-loop operating system is formed, improving service matching accuracy, response efficiency, and resource allocation rationality, comprehensively enhancing the operational quality and service capabilities of smart elderly care.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and privacy computing, specifically to a privacy-preserving joint modeling system for elderly data based on federated learning. Background Technology

[0002] As the smart elderly care industry continues to advance, wearable devices, home sensing devices, and community service terminals are constantly generating multi-dimensional elderly care data. This data covers key information such as physiological status, behavioral environment, and service operations, and is an important foundation for supporting precise elderly care services. Currently, elderly care data is scattered across different community units and various smart terminals, with each data node operating independently, forming data silos. Cross-node data fusion and value mining are difficult to carry out. At the same time, elderly care data contains a large amount of sensitive personal information, and data security and privacy protection are subject to strict regulatory constraints. Direct transmission and centralized use of raw data pose extremely high compliance risks. Federated learning technology provides a technical approach for cross-node data collaborative utilization. How to achieve joint modeling and efficient application of elderly care data while strictly adhering to privacy protection requirements has become a core pain point in the development of smart elderly care technology. The industry urgently needs technical solutions that combine data security and collaborative modeling capabilities to promote the upgrading of elderly care services towards intelligence and compliance.

[0003] Traditional elderly care data processing and service modeling technologies have many shortcomings and cannot meet the actual needs of modern smart elderly care. Traditional technologies mostly adopt a centralized data processing model, collecting raw data from various nodes before modeling, which is prone to leakage of sensitive information. Moreover, the desensitization methods are limited and cannot implement hierarchical protection for different types of sensitive data. The collection formats and feature dimensions of multi-source elderly care data vary greatly, making it difficult for traditional technologies to achieve standardization and time-series alignment. Poor data quality directly reduces the modeling effect. Traditional modeling cannot achieve collaborative training across community nodes, resulting in weak model generalization and applicability. Service recommendations rely solely on static data to generate fixed profiles, failing to match the real-time changes in the elderly's status. Elderly care resource scheduling relies heavily on manual experience and judgment, without coordinating service suitability and urgency for optimization. This leads to low resource allocation efficiency, slow service response speed, and difficulty in implementing personalized elderly care services. The core contradiction between data privacy protection and data value release remains unresolved. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a privacy-preserving joint modeling system for elderly care data based on federated learning. Through a horizontal federated learning architecture, it enables collaborative training of multiple elderly care community nodes. The original data does not leave the local machine; only the model parameters are uploaded, effectively protecting sensitive information such as the elderly's identity, health, and location. The multi-model fusion module runs a privacy fusion profiling algorithm to generate dynamic elderly profiles. The service recommendation module constructs a heterogeneous association graph to achieve accurate service recall and urgency assessment. The resource scheduling module jointly optimizes decisions to generate the optimal service personnel, time slots, and equipment configuration schemes.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a privacy-preserving joint modeling system for elderly care data based on federated learning, the system comprising: Data processing module: performs multi-source elderly care data collection and standardization processing, performs privacy desensitization processing on sensitive information, and outputs standardized desensitized dataset; Federated Modeling Module: Receives standardized and de-identified datasets, uses a horizontal federated learning architecture to build a multi-node collaborative framework, completes local model training and secure aggregation of global parameters, and generates a global joint model; Multi-modal fusion module: Receives standardized de-identified dataset and global joint model, runs federated privacy fusion profiling algorithm, and generates dynamic profile vectors of the elderly; Service recommendation module: Receives dynamic profile vectors of the elderly, constructs a heterogeneous association graph of elderly service scenarios to complete the recall and adaptation of elderly care services, and generates service adaptation features and demand urgency data; Resource scheduling module: Receives dynamic profile vectors of elderly people, service adaptation features and demand urgency data, runs a service scheduling joint optimization algorithm to calculate the recommended scheduling joint decision value, generates a resource scheduling scheme and executes the data feedback to the data processing module.

[0006] Furthermore, the data processing module includes multi-source elderly care data collection and standardization processing. Specifically, it collects physiological time-series data transmitted from wearable devices, behavioral and environmental data transmitted from home sensing devices, and business record data transmitted from community service terminals. The collected data is sorted and organized frame by frame according to the collection time sequence, redundant data collected repeatedly is removed, abnormal data is filtered, and the time sequence of all types of data is aligned at a uniform time granularity of 1 minute. Data with different collection formats and different feature dimensions are converted into numerical features with a unified structure. Normalization transformation is performed on continuous physiological and environmental data, and encoding transformation is performed on discrete business and behavioral data to form standardized data with unified dimensions and coherent time sequence, providing a well-organized data foundation for subsequent privacy desensitization and model training.

[0007] Furthermore, in the data processing module, privacy desensitization is implemented in a tiered manner for four types of sensitive data in elderly care data: identity information, health information, location information, and contact information. Irreversible hash encryption is used for core identity information such as ID card numbers and mobile phone numbers; data generalization is used for health information such as disease diagnoses, chronic disease history, and medication plans, converting precise values ​​into interval features; regional fuzzification is used for location information such as home addresses and precise location information, masking specific building and room numbers; and field masking is used for contact information. This process completely severs the direct association between data and personal identity, does not retain original sensitive plaintext data, and only retains non-sensitive feature data required for modeling. The desensitized data is then integrated with standardized data to form a standardized desensitized dataset.

[0008] Furthermore, in the federated modeling module, the multi-node collaborative framework consists of multiple local elderly care nodes and a global aggregation node. Each local elderly care node corresponds to an independent elderly care community unit, independently stores and uses standardized and desensitized datasets, and only conducts model iterative training within the local node without transmitting the original data to the outside. Each local computing node uploads the locally trained model parameters to the global model aggregation service node according to a fixed iteration cycle of 20 rounds. The global model aggregation service node uniformly receives the parameters from each local node and completes secure aggregation. After generating the global joint model, it is synchronously distributed to all local computing nodes, forming a closed-loop collaborative process of local training, parameter uploading, global aggregation, and model distribution.

[0009] Furthermore, in the multi-modal fusion module, the mathematical expression of the federated privacy fusion profiling algorithm is: ; in, Let be the dynamic portrait vector of the i-th elderly person after the t-th iteration, where i is the elderly person's index; The fusion coefficient is specific to the elderly care scenario, with a value range of 0-1. It is used to balance the weights of local multimodal data features and global model features. It is adjusted according to the health level of the elderly, with a value range of 0.7-0.9 for high-risk elderly and 0.4-0.6 for low-risk elderly. K is the total number of types of multimodal data for elderly care, with a value of 4, corresponding to four types of data: physiological time series data, behavioral environment data, service record data, and static archive data. The privacy-preserving adaptive weight for the k-th modality data in the t-th iteration is automatically generated by the local node based on the frequency of abnormal data fluctuations. The weight value is positively correlated with the abnormal frequency of the corresponding modality data. k is the index of the elderly multimodal data type. This provides the standardized and desensitized data for the k-th modality of the i-th elderly person in the t-th iteration. These are the global joint model parameters generated in the (t-1)th iteration.

[0010] Furthermore, in the service recommendation module, the heterogeneous association graph of elderly service scenarios consists of a node set and an edge set. The node set includes three types of nodes: elderly nodes with the elderly dynamic profile vector as feature attributes, service nodes with the type, content, and attributes of elderly care services as feature attributes, and scenario nodes with the application environment and conditions of elderly care services as feature attributes. The edge set includes five types of association edges: historical interaction edges between the elderly and services, feature adaptation edges between the elderly and scenarios, condition matching edges between services and scenarios, profile similarity edges between elderly people, and business association edges between services. Each edge is generated based on historical interaction records and feature similarity data, forming a complete heterogeneous association graph structure.

[0011] Furthermore, in the service recommendation module, the elderly care service recall and adaptation process is specifically based on the constructed heterogeneous association graph of elderly service scenarios. Starting from the target elderly node, the association edges are traversed, and service nodes with an association strength of not less than 0.6 are selected to form an initial service set. The service nodes in the initial service set are matched with the elderly dynamic profile. During the matching, the built-in elderly care domain knowledge graph rules are referenced. The rules include three categories: disease services, physical activities, and seasonal care. Based on the feature information in the elderly dynamic profile, the services in the initial service set are filtered, services that do not meet the rules are removed, the order of the remaining services is adjusted, and service adaptation feature data is formed. Then, combined with the real-time status in the elderly dynamic profile, demand urgency data is generated.

[0012] Furthermore, the service adaptation features include the basic attributes of the service node, the association strength of heterogeneous association edges, and the matching results of knowledge graph rules. The basic attributes include the service type, service duration, and service implementation conditions. The association strength is the weight value of the association edge between the service node and the elderly node. The matching results are the rule matching tags between the service and the elderly profile features. The service adaptation features are stored in the form of numerical vectors, with each dimension corresponding to the adaptation information of one service. The demand urgency data is generated based on the real-time status in the elderly dynamic profile and includes two items: status level and response time limit. The status level corresponds to the elderly's current health and risk status, and the response time limit is the time requirement for the corresponding service to be started. The demand urgency data is stored in the form of structured identifiers, with each data item corresponding to one service to be recommended for one elderly person.

[0013] Furthermore, in the resource scheduling module, the mathematical expression of the joint optimization algorithm for service scheduling is: ; in, Let i be the recommended scheduling joint decision value for the s-th elderly care service corresponding to the i-th elderly person. Each decision value corresponds to one service to be executed for an elderly person. i is the elderly person index, s is the service index, and t is the iteration round. This is the service scheduling priority coefficient, a scalar value ranging from 0 to 1. It is used to adjust the weight ratio of the adaptability dimension and the urgency dimension in the joint decision-making process and is set by the resource scheduling module according to the type of elderly care service and the operational scenario. The dynamic profile vector of the i-th elderly person after the t-th iteration is output by the multi-modal fusion module and is a multi-dimensional vector containing the elderly person's physiological state, behavior pattern, risk level, and service preferences. The feature vector for the s-th elderly care service is output by the service recommendation module. It includes the service type, duration, resource consumption, implementation condition attributes, and the matching strength information between the service and the elderly profile. The urgentness vector of the service demand for the i-th elderly person corresponding to the s-th service is output by the service recommendation module and includes the elderly person’s health risk level, service response time limit, and task priority information corresponding to the service. This is the global optimal scheduling factor vector for elderly care resources, which includes the number of service personnel available for allocation in the current time period, the status of service equipment resources, and the distribution information of service locations.

[0014] Furthermore, in the aforementioned resource scheduling module, the resource scheduling scheme generated by the resource scheduling module includes service execution sequence, service personnel allocation information, service time period arrangement information, and service equipment configuration information; the service execution sequence is arranged from high to low according to the recommended scheduling joint decision value, and each service corresponds to a unique service personnel, service time period, and service equipment; the service personnel allocation information includes the service personnel number, the work group to which they belong, and the type of service that can be executed; the service time period arrangement information includes the service start time, the service duration, and the service end time; the service equipment configuration information includes the equipment number, the equipment type, and the equipment usage status identifier; the scheduling scheme is stored in the form of structured data, forming a structured scheduling list containing elderly identifiers, service identifiers, personnel identifiers, time period identifiers, and equipment identifiers.

[0015] Compared with existing technologies, this privacy-preserving elderly data joint modeling system based on federated learning has the following advantages: I. This invention standardizes and hierarchically desensitizes multi-source elderly care data, severing the direct link between sensitive data and personal identity throughout the process, and not retaining original sensitive plaintext data. It builds a multi-node collaborative training framework based on a horizontal federated learning architecture, where each local node trains the model locally, transmitting only model parameters rather than the original data. This effectively solves the industry problem of fragmented and independent multi-source elderly care data and high privacy leakage risks, enabling the secure and compliant collaborative use of elderly care data, breaking down data silos between communities, and efficiently constructing a global joint model while protecting data privacy rights. This provides secure and reliable technical support for the in-depth mining and application of elderly care data, allowing the value of multi-node elderly care data to be fully released, while improving the stability and versatility of model training and adapting to elderly care data application scenarios in different communities.

[0016] Second, this invention integrates multimodal data features and global model features through a federated privacy fusion profiling algorithm to generate dynamic profiles that fit the actual state of the elderly. It relies on heterogeneous association graphs of elderly service scenarios to achieve accurate recall and adaptive screening of elderly care services. It optimizes service matching logic by combining knowledge graph rules in the elderly care field, and then uses a service scheduling joint optimization algorithm to coordinate elderly needs, service suitability, and global resource status to generate a scientific and reasonable resource scheduling scheme. This forms a closed-loop operating system of data processing, modeling, profiling, recommendation, and scheduling, improving the matching accuracy between elderly care services and elderly needs, shortening service response time, optimizing the allocation efficiency of elderly care service resources, avoiding resource idleness or waste, and making elderly care services more personalized and real-time. This comprehensively improves the operational efficiency and service quality of elderly care services, providing an efficient and feasible technical solution for the implementation of smart elderly care services.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 A flowchart of a privacy-preserving joint modeling system for elderly data based on federated learning; Figure 2 This is a schematic diagram of data transmission in a privacy-preserving joint modeling system for elderly care data based on federated learning. Figure 3 This is a schematic diagram of data transmission in the federated modeling module of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: This embodiment focuses on smart elderly care services for seniors living at home within a single community. It leverages a privacy-preserving collaborative modeling system for elderly care data based on federated learning to fully implement the entire process of multi-source elderly care data collection and processing, federated collaborative modeling, dynamic profile generation for seniors, precise service recommendations, and intelligent resource scheduling. This comprehensively strengthens the privacy and security of elderly care data, transforming scattered home-based elderly care data into actionable, precise service capabilities. It enables community-based home care services to shift from passive response to proactive prediction, and from generalized services to personalized adaptation, comprehensively improving the refinement of community-based home care services and the service experience for seniors. Figure 1 As shown.

[0022] The data processing module, specifically designed for home-based elderly care within this community, comprehensively collects physiological time-series data transmitted from wearable devices, behavioral and environmental data transmitted from home sensors, and elderly care service record data transmitted from community service terminals. This data covers core information such as daily physiological indicators, home activity status, and historical service usage, serving as the core basis for subsequent modeling and service implementation. The module sorts and organizes all data frame-by-frame according to the collection sequence, eliminating redundant data from repeated collections, filtering abnormal data caused by equipment malfunctions or environmental interference, and aligning all data types at a unified time granularity. It transforms scattered data with different collection formats and feature dimensions into numerical features with a unified structure, performs normalization transformation on continuous physiological and environmental data, and performs encoding transformation on discrete business and behavioral data, forming standardized data with unified dimensions and coherent time sequences. This gives the originally chaotic multi-source data a unified processing standard, laying a solid data foundation for subsequent privacy desensitization and model training. To avoid modeling biases caused by inconsistent data formats and disordered time sequences, and to reduce the drag on system efficiency caused by invalid data, the module further implements hierarchical desensitization for four types of sensitive data in elderly care data: identity information, health information, location information, and contact information. Irreversible hash encryption is used for core identity information, data generalization is used for health information, regional fuzzification is used for location information, and field masking is used for contact information. The direct association between data and personal identity is severed throughout the process, and no original sensitive plaintext data is retained. Only non-sensitive feature data required for modeling is retained. The desensitized data is integrated with standardized data to form a standardized desensitized dataset. While fully preserving the data modeling value, the risk of leakage of elderly people's sensitive personal information is eliminated from the source. This meets the core requirements of privacy protection and compliant use of elderly care data, ensuring data security in the data processing stage and providing a safe and usable data carrier for subsequent federated modeling, ensuring that the data plays its maximum value under the premise of security and compliance.

[0023] The federated modeling module receives a standardized, anonymized dataset from the data processing module. It employs a horizontal federated learning architecture to build a multi-node collaborative framework consisting of local support nodes and a global aggregation node. Each community has an independent local support node, which independently stores and uses the standardized, anonymized dataset. All model iterations and training are completed within the local node, never transmitting any raw data externally, thus completely mitigating the risk of data leakage during transmission. Local computing nodes upload their trained model parameters to the global model aggregation service node at fixed iteration cycles. The global model aggregation service node uniformly receives the parameters from all local nodes and performs secure aggregation, generating a global federated dataset. After model aggregation, the data is synchronously distributed to all local computing nodes, forming a closed-loop collaborative process of local training, parameter uploading, global aggregation, and model distribution. This eliminates the need to collect raw data from each node to build a high-quality global model, ensuring local data remains within the community. Simultaneously, the generalization ability of the global model enhances the service adaptation accuracy of local models, overcoming modeling bottlenecks caused by limited data within the community. This allows the model to adapt to the complex needs of home-based elderly care scenarios, continuously optimizing model training effects and service prediction capabilities. This provides accurate model support for subsequent elderly profile generation and service recommendations, removing privacy compliance restrictions on cross-node data collaboration. Figure 3 As shown.

[0024] The multi-modal fusion module receives the standardized, de-identified dataset output by the data processing module and the global joint model output by the federated modeling module, and runs the federated privacy fusion profiling algorithm. The mathematical expression of the federated privacy fusion profiling algorithm is: ; in, Let be the dynamic portrait vector of the i-th elderly person after the t-th iteration; This is a fusion coefficient specific to the elderly care scenario; K represents the total number of data types in the elderly care multimodal dataset. The privacy-preserving adaptive weights for the k-th modality data in the t-th iteration; This provides the standardized and desensitized data for the k-th modality of the i-th elderly person in the t-th iteration. The algorithm generates global joint model parameters for the (t-1)th iteration. It balances the weights of local multimodal data features and global model features, adjusts the fusion coefficients based on the elderly person's health level, and automatically generates adaptive weights for corresponding modal data based on the frequency of abnormal data fluctuations. It deeply integrates four types of multimodal data—physiological time-series data, behavioral environment data, service record data, and static archive data—with the global model parameters to generate a dynamic profile vector for each elderly person living at home. The dynamic profile vector fully carries the core information of the elderly person's real-time physiological state, behavioral patterns, health risks, and service preferences. It can update in real time according to changes in the elderly person's state, overcoming the lag defects of traditional static profiles, accurately depicting the elderly person's personalized elderly care needs, and providing the most suitable basis for subsequent service recommendations. This allows service recommendations to no longer rely on fixed labels but to achieve accurate matching based on real-time dynamic features. At the same time, relying on federated privacy protection features, it always adheres to the bottom line of data privacy during the profile generation process, without disclosing any sensitive personal information of the elderly person, so that personalized profile generation and privacy protection are achieved simultaneously.

[0025] The service recommendation module receives the dynamic profile vector of the elderly from the multi-modal fusion module and constructs a heterogeneous association graph of elderly service scenarios. This graph consists of three types of nodes: elderly nodes, service nodes, and scenario nodes, as well as five types of association edges: historical interaction edges, feature adaptation edges, condition matching edges, profile similarity edges, and business association edges. Elderly nodes use the dynamic profile vector as their feature attribute, service nodes use the type, content, and attributes of elderly care services as their feature attributes, and scenario nodes use the application environment and conditions of elderly care services as their feature attributes. All types of association edges are generated based on historical interaction records and feature similarity data, constructing a complete elderly care service association network. Starting from the target elderly node, the module traverses the association edges, selecting service nodes with high association strength to form an initial service set, and then... Services in the initial service set are screened according to the rules of knowledge graphs in the field of elderly care, eliminating services that do not meet the requirements of the elderly's condition and scenario. The ranking of the remaining services is adjusted, and service adaptation features and demand urgency data are generated. The service adaptation features fully cover the basic attributes of the services, the strength of the association, and the rule matching results, and clearly present the service adaptation information in the form of numerical vectors. The demand urgency data clearly defines the elderly's health risk level and the service response time limit, so that service recommendation can get rid of the subjectivity of human experience judgment. Relying on data and association rules, accurate adaptation is achieved, quickly identifying the most suitable elderly care services for the elderly. At the same time, the response priority of services is clarified, providing a clear service basis for subsequent resource scheduling, and improving the accuracy and efficiency of service matching.

[0026] The resource scheduling module receives the elderly dynamic profile vector output by the multi-modal fusion module and the service adaptation features and demand urgency data output by the service recommendation module. It then runs a joint service scheduling optimization algorithm, the mathematical expression of which is: ; in, Let t be the joint decision value for recommending and scheduling the s-th elderly care service corresponding to the i-th elderly person, where i is the elderly person index, s is the service index, and t is the iteration round. This is a service scheduling priority coefficient. Let be the dynamic portrait vector of the i-th elderly person after the t-th iteration; Let s be the adaptive feature vector for the s-th elderly care service; Let be the urgency vector of the service demand for the i-th elderly person corresponding to the s-th service; To determine the globally optimal scheduling factor vector for elderly care resources, this algorithm adjusts the weighting of the adaptability and urgency dimensions in joint decision-making. It sets weight coefficients based on the type of elderly care service and the operational scenario, integrating globally optimal scheduling factors such as the number of available service personnel, the status of service equipment resources, and the distribution of service locations in the current time period. This results in a recommended joint decision value for scheduling. Service execution sequences are generated by arranging these decision values ​​from highest to lowest, clearly defining the service personnel, service time slots, and service equipment corresponding to each service. This forms a resource scheduling scheme that includes service execution sequences, service personnel allocation information, service time slot arrangements, and service equipment configuration information, thus creating a structured... The data is stored as a complete scheduling list, and the scheduling data is fed back to the data processing module, forming a closed-loop operation system of data processing, modeling, profiling, recommendation, and scheduling. This allows elderly care service resources to be accurately allocated according to the urgency of the elderly's needs and the suitability of the services, avoiding resource idleness and waste, shortening service response time, improving the operational efficiency of community elderly care services, and enabling elderly people living at home to obtain the most suitable elderly care services in the shortest possible time. At the same time, the data fed back in the closed loop continuously optimizes the data processing and model training effects, allowing the entire system to continuously iterate and upgrade, continuously improve the quality of home-based elderly care services, and provide stable support for the long-term operation of community-based home care services.

[0027] This embodiment, within a single community-based home care scenario, achieves secure organization and privacy anonymization of multi-source home care data through end-to-end system operation. It utilizes federated learning to complete local data security modeling, employs a federated privacy fusion profiling algorithm to generate dynamic profiles of the elderly, achieves precise service recommendations through heterogeneous association graphs and knowledge graphs, and finally completes intelligent resource allocation through a service scheduling joint optimization algorithm, forming a closed-loop operating system. This embodiment fully protects the privacy and security of elderly data, breaks through the limitations of data within the community, enables elderly care services to accurately match the real-time needs of the elderly, optimizes the efficiency of community resource allocation, and improves the refinement and intelligence of home care services, providing a complete and feasible technical solution for the implementation of home care services in a single community.

[0028] Example 2: This embodiment addresses a scenario where multiple adjacent communities collaborate to provide elderly care services. It leverages a privacy-preserving joint modeling system for elderly care data based on federated learning to achieve secure cross-community collaboration on elderly care data, joint construction of a global model, unified generation of elderly profiles, precise service adaptation, and coordinated resource allocation. This breaks down data silos between communities, enabling efficient cross-regional allocation of elderly care resources. It allows multi-community elderly care services to form a collaborative whole, improving the overall supply capacity and service balance of regional elderly care services. This provides a complete technical implementation solution for cross-community smart elderly care operations. Figure 2 As shown.

[0029] The data processing module collects and processes elderly care data for each linked community. It independently collects physiological time-series data from wearable devices, behavioral and environmental data from home sensors, and elderly care business records from community service terminals for each community. Each community's data is processed independently and stored locally. The module sorts and organizes the data frame by frame according to the collection sequence, removing redundant data from duplicate collections, filtering abnormal data caused by equipment failure or environmental interference, and aligning all data types according to a unified time granularity. It converts scattered data with different collection formats and feature dimensions into numerical features with a unified structure, performs normalization transformation on continuous physiological and environmental data, and performs encoding transformation on discrete business and behavioral data, forming standardized data with unified dimensions and coherent time sequences for each community. This provides a unified processing standard for elderly care data from different communities, offering a well-structured data foundation for cross-community federated modeling and avoiding issues arising from... The module addresses the obstacles to collaborative modeling caused by differences in data standards between communities, while simultaneously improving the system's computational efficiency in processing multi-community data. Furthermore, it implements tiered desensitization for four types of sensitive data in each community's elderly care data: identity information, health information, location information, and contact information. Irreversible hash encryption is used for core identity information, data generalization is used for health information, regional obfuscation is used for location information, and field masking is used for contact information. This process completely severs the direct link between data and individual identity, leaving no original sensitive plaintext data, retaining only the non-sensitive feature data required for modeling. The desensitized data is then integrated with standardized data to form standardized desensitized datasets for each community. This fully preserves the modeling value of each community's data while ensuring privacy and security in cross-community data processing, ensuring that each community's data remains locally. This eliminates the risk of privacy leaks from cross-community data transmission at the source, meeting the core requirements for compliant collaborative use of regional elderly care data.

[0030] The federated modeling module receives standardized, anonymized datasets from the data processing modules of each community. It employs a horizontal federated learning architecture to build a multi-community, multi-node collaborative framework. Each linked community corresponds to an independent local node, while the global aggregation node coordinates all local nodes. Each local node independently stores and uses its own standardized, anonymized dataset. All model iterations and training are completed within the community's local node, never transmitting any raw data externally, thus ensuring complete local data security for each community. Each local computing node uploads its trained model parameters to the global model aggregation service node at a fixed iteration cycle. The global model aggregation service node uniformly receives all local node data from each community. The parameters are collected and securely aggregated to generate a global joint model adapted to multi-community collaborative scenarios. This model is then simultaneously distributed to the local computing nodes of all communities, forming a closed-loop collaborative process across communities: local training, parameter uploading, global aggregation, and model distribution. High-quality global model construction can be completed without collecting raw data from any community, breaking down data silos between communities and enabling joint modeling of data from each community without leaving their local area. The generalization ability of the global model compensates for the shortcomings of insufficient data volume and single scenario in a single community, improving the adaptability and prediction accuracy of the model in multi-community collaborative scenarios. This provides unified and accurate model support for cross-community elderly profile generation and service recommendation.

[0031] The multimodal fusion module receives standardized, anonymized datasets from the data processing modules of each community and a cross-community global joint model from the federated modeling module. It then runs a federated privacy fusion profiling algorithm. This algorithm balances the weights of local multimodal data features from each community with the global model features, adjusts the fusion coefficients based on the health levels of the elderly in each community, and automatically generates adaptive weights for corresponding modal data based on the frequency of abnormal data fluctuations in each community. It deeply integrates four types of multimodal data—physiological time-series data, behavioral environment data, service record data, and static archive data—with the global model parameters to generate a unified, standardized dynamic profile vector for each elderly person in each community. This dynamic profile vector fully carries core information such as the elderly person's real-time physiological state, behavioral patterns, health risks, and service preferences. It can update in real time following changes in the elderly person's status, balancing the differences in elderly care scenarios across communities with the uniformity of regional services. It accurately depicts the personalized elderly care needs of elderly people across communities, providing a unified and accurate basis for cross-community service recommendations. This allows elderly people in different communities to receive service recommendations tailored to their individual circumstances. Furthermore, relying on federated privacy protection features, it adheres to the data privacy bottom line of each community throughout the profile generation process, ensuring no leakage of any sensitive personal information of the elderly, thus achieving the dual goals of cross-community personalized profile generation and data privacy protection.

[0032] The service recommendation module receives dynamic profile vectors of elderly people from the multi-modal fusion modules of each community. It then constructs a heterogeneous association graph of elderly service scenarios for each linked community. Each association graph consists of three types of nodes: elderly nodes, service nodes, and scenario nodes, as well as five types of association edges: historical interaction edges, feature adaptation edges, condition matching edges, profile similarity edges, and business association edges. Elderly nodes use dynamic profile vectors as feature attributes, service nodes use elderly care service type, content, and attributes as feature attributes, and scenario nodes use elderly care service application environment and conditions as feature attributes. All types of association edges are generated based on historical interaction records and feature similarity data of each community, constructing an elderly care service association network adapted to the scenarios of each community. Starting from the target elderly node in each community, the module traverses the association edges to select service nodes with high association strength. An initial service set is formed, and then the services in the initial service set are screened according to the knowledge graph rules in the field of elderly care. Services that do not meet the requirements of the elderly's condition and community scenario are removed, and the ranking of the remaining services is adjusted. Service adaptation features and demand urgency data are generated for the elderly in each community. The service adaptation features fully cover the basic attributes of the service, the strength of the association, and the rule matching results, and clearly present the service adaptation information in the form of numerical vectors. The demand urgency data clearly defines the elderly's health risk level and the service response time limit, so that cross-community service recommendation can get rid of the subjectivity of human experience judgment. Based on data and association rules, the accurate adaptation of services in each community can be achieved, and the most suitable elderly care services for the elderly in each community can be quickly identified. At the same time, the response priority of services is clarified, providing a clear and unified service basis for the overall scheduling of cross-community resources.

[0033] The resource scheduling module receives dynamic profile vectors of the elderly from the multi-modal fusion modules of each community, and service suitability features and urgency data from the service recommendation module. It then runs a joint optimization algorithm for service scheduling. This algorithm adjusts the weighting of suitability and urgency dimensions in the joint decision-making process, sets weighting coefficients based on the type and operational scenario of multi-community collaborative elderly care services, and integrates globally optimal scheduling factors such as the number of available service personnel, service equipment resource status, and service location distribution across all communities in the region. It calculates the recommended joint decision value for services corresponding to the elderly in each community, and generates a cross-community coordinated service execution sequence based on these decision values ​​from highest to lowest. This sequence clearly defines the service personnel, service time slots, and service equipment for each service in each community, forming a comprehensive service execution sequence containing service personnel allocation information, service time slot arrangement information, and service details. The cross-community resource scheduling scheme for equipment configuration information stores a complete regional scheduling list in structured data format. Simultaneously, the scheduling data is fed back to the data processing modules of each community, forming a closed-loop operation system for cross-community data processing, modeling, profiling, recommendation, and scheduling. This allows for the coordinated allocation of elderly care service resources within the region based on the urgency of the elderly's needs and service suitability, balancing resource supply across communities, avoiding resource idleness and waste, shortening cross-community service response time, and improving the overall operational efficiency of multi-community collaborative elderly care services. This ensures that elderly people in different communities can fairly access high-quality and suitable elderly care services. Furthermore, the data fed back through the closed loop continuously optimizes the data processing and model training effects in each community, enabling the entire system to continuously iterate and upgrade, continuously improving the balance and quality of cross-community elderly care services, and providing solid technical support for the integrated operation of regional elderly care services.

[0034] This embodiment, in a multi-community collaborative elderly care scenario, achieves local secure processing of elderly care data and cross-node federated modeling through cross-community collaborative operation. It generates standardized dynamic profiles of the elderly based on a federated privacy fusion profiling algorithm, providing precise service adaptation for each community. A service scheduling joint optimization algorithm then completes the overall scheduling of regional resources. This embodiment breaks down data silos between communities, enabling cross-community collaborative modeling with data remaining within the community, balancing the supply of regional elderly care resources, ensuring data privacy and service fairness for the elderly across the region, and improving the operational efficiency and service quality of multi-community collaborative elderly care. It provides a mature implementation model for regional integrated smart elderly care services.

[0035] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A privacy-preserving joint modeling system for elderly care data based on federated learning, characterized in that, The system includes: Data processing module: performs multi-source elderly care data collection and standardization processing, performs privacy desensitization processing on sensitive information, and outputs standardized desensitized dataset; Federated Modeling Module: Receives standardized and de-identified datasets, uses a horizontal federated learning architecture to build a multi-node collaborative framework, completes local model training and secure aggregation of global parameters, and generates a global joint model; Multi-modal fusion module: Receives standardized de-identified dataset and global joint model, runs federated privacy fusion profiling algorithm, and generates dynamic profile vectors of the elderly; Service recommendation module: Receives dynamic profile vectors of the elderly, constructs a heterogeneous association graph of elderly service scenarios to complete the recall and adaptation of elderly care services, and generates service adaptation features and demand urgency data; Resource scheduling module: Receives dynamic profile vectors of elderly people, service adaptation features and demand urgency data, runs a service scheduling joint optimization algorithm to calculate the recommended scheduling joint decision value, generates a resource scheduling scheme and executes the data feedback to the data processing module.

2. The privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the data processing module, multi-source elderly care data collection and standardization processing specifically involves collecting physiological time-series data transmitted from wearable devices, behavioral and environmental data transmitted from home sensing devices, and business record data transmitted from community service terminals. The collected data is sorted and organized frame by frame according to the collection time sequence, redundant data collected repeatedly is removed, abnormal data is filtered, and the time sequence of all types of data is aligned at a unified time granularity. Data with different collection formats and different feature dimensions are converted into numerical features with a unified structure. Normalization conversion is performed on continuous physiological and environmental data, and encoding conversion is performed on discrete business and behavioral data to form standardized data with unified dimensions and coherent time sequence.

3. The privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the data processing module, privacy desensitization is performed on four types of sensitive data in elderly care data: identity information, health information, location information, and contact information. Irreversible hash encryption is used for identity information; data generalization is used for health information; regional fuzzification is used for location information, and specific building and room number information is hidden; and field masking is used for contact information. The direct association between data and personal identity is completely severed, and no original sensitive plaintext data is retained. Only non-sensitive feature data required for modeling is retained. The desensitized data is integrated with standardized data to form a standardized desensitized dataset.

4. The privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the federated modeling module, the multi-node collaborative framework consists of multiple local elderly care nodes and a global aggregation node. Each local elderly care node corresponds to an independent elderly care community unit, independently stores and uses standardized and de-identified datasets, and only conducts model iterative training within the local node without transmitting the original data to the outside. Each local computing node uploads the locally trained model parameters to the global model aggregation service node according to a fixed iteration cycle of 20 rounds. The global model aggregation service node uniformly receives the parameters of each local node and completes secure aggregation, generates a global joint model, and then synchronously distributes it to all local computing nodes.

5. A privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the multi-modal fusion module, the mathematical expression of the federated privacy fusion profiling algorithm is: ； in, Let be the dynamic portrait vector of the i-th elderly person after the t-th iteration; This is a fusion coefficient specific to the elderly care scenario; K represents the total number of data types in the elderly care multimodal dataset. The privacy-preserving adaptive weights for the k-th modality data in the t-th iteration; This provides the standardized and desensitized data for the k-th modality of the i-th elderly person in the t-th iteration. These are the global joint model parameters generated in the (t-1)th iteration.

6. A privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the service recommendation module, the heterogeneous association graph of elderly service scenarios consists of a node set and an edge set. The node set contains three types of nodes: elderly nodes with the elderly dynamic profile vector as feature attributes, service nodes with the type, content, and attributes of elderly care services as feature attributes, and scenario nodes with the application environment and conditions of elderly care services as feature attributes. The edge set contains five types of association edges: historical interaction edges between the elderly and services, feature adaptation edges between the elderly and scenarios, condition matching edges between services and scenarios, profile similarity edges between elderly people, and business association edges between services. Each edge is generated based on historical interaction records and feature similarity data, forming a complete heterogeneous association graph structure.

7. A privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the service recommendation module, the elderly care service recall and adaptation process is specifically based on the constructed heterogeneous association graph of elderly service scenarios. Starting from the target elderly node, the association edges are traversed, and service nodes with an association strength of not less than 0.6 are selected to form an initial service set. The service nodes in the initial service set are matched with the elderly dynamic profile. During the matching, the built-in elderly care domain knowledge graph rules are referenced. The rules include three categories: disease services, physical activities, and seasonal care. Based on the feature information in the elderly dynamic profile, the services in the initial service set are filtered, services that do not meet the rules are removed, and the order of the remaining services is adjusted to form service adaptation feature data. Then, combined with the real-time status in the elderly dynamic profile, the urgency data of the demand is generated.

8. A privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the resource scheduling module, the mathematical expression of the service scheduling joint optimization algorithm is: ； in, Let t be the joint decision value for recommending and scheduling the s-th elderly care service corresponding to the i-th elderly person, where i is the elderly person index, s is the service index, and t is the iteration round. This is a service scheduling priority coefficient; Let be the dynamic portrait vector of the i-th elderly person after the t-th iteration; Let s be the adaptive feature vector for the s-th elderly care service; Let be the urgency vector of the service demand for the i-th elderly person corresponding to the s-th service; This is the globally optimal scheduling factor vector for elderly care resources.

9. A privacy-preserving elderly care data joint modeling system based on federated learning according to claim 1, characterized in that, In the resource scheduling module, the resource scheduling scheme generated by the module includes service execution sequence, service personnel allocation information, service time period arrangement information, and service equipment configuration information. The service execution sequence is arranged from high to low according to the recommended joint scheduling decision value, and each service corresponds to a unique service personnel, service time period, and service equipment. The service personnel allocation information includes the service personnel number, the work group to which they belong, and the type of service that can be executed. The service time period arrangement information includes the service start time, the service duration, and the service end time. The service equipment configuration information includes the equipment number, the equipment type, and the equipment usage status identifier. The scheduling scheme is stored in the form of structured data, forming a structured scheduling list containing elderly identifiers, service identifiers, personnel identifiers, time period identifiers, and equipment identifiers.

Images