Personnel health information recommendation method and device, equipment and medium
By leveraging the collaborative work of edge computing device clusters and cloud-based large models, and utilizing local large language models to generate contextual vectors combined with cloud-based analysis, the shortcomings of intelligent health systems in proactive perception and personalized recommendations are addressed, enabling rapid response and personalized health information recommendations.
Patent Information
- Application Number
- CN202511088733.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-11-18
AI Technical Summary
Existing smart health systems struggle to proactively perceive users' deeper needs and potential risks in home settings for the elderly. Furthermore, large cloud-based models are prone to delayed responses and privacy concerns, while purely edge models are limited by computing power and cognitive depth, resulting in insufficient personalized health information recommendations.
Multimodal data is collected by edge computing device clusters, contextual vectors are generated using local large language models, and in-depth analysis is performed in conjunction with cloud-based large models to collaboratively generate personalized health recommendations, including food and medicine recommendations and voice information recommendations.
While ensuring user privacy and rapid response, it has improved the depth of understanding of complex user scenarios and the accuracy of personalized health information recommendations, and optimized system response speed and resource utilization efficiency.
Smart Images
Figure CN120977478A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, in particular to a personnel health information recommendation method and device, equipment and medium. BACKGROUND
[0002] At present, although the intelligent health and pension technology has developed, the existing intelligent recommendation system is still insufficient in active care, especially in the scene of the elderly at home, which mainly shows that the user needs to actively initiate interaction (such as voice wake-up or click button wake-up), it is difficult to actively perceive and predict the user's deep needs and potential risks, and give recommended information to solve the problem (especially health problem). At the same time, simply relying on cloud big model is easy to cause delay reply problem, difficult to obtain comprehensive information problem and privacy concern, and pure edge model is limited by computing power and cognitive depth. Therefore, there is an urgent need for a personnel health information recommendation method that can actively perceive, deeply understand and provide personalized health information recommendation. SUMMARY
[0003] The purpose of the present application is to provide a personnel health information recommendation method, device, equipment and medium, which can actively perceive, deeply understand and provide personalized health information recommendation, aiming at solving the deficiency of the existing family health intelligent system in active perception, context understanding and personalized information recommendation.
[0004] To achieve the above purpose, the present application provides the following scheme:
[0005] In a first aspect, the present application provides a personnel health information recommendation method, comprising:
[0006] The edge computing device group collects multi-modal raw data of the user, and processes the multi-modal raw data by using a local large language model to generate a plurality of context vectors;
[0007] Using the local large language model, the scene complexity of the current scene is obtained by dynamically evaluating all context vectors, and the weight value of each context vector is obtained by assigning weights to each context vector;
[0008] Using the local large language model, a potential demand set is generated according to each context vector, the weight value of each context vector and the potential demand problem library; the potential demand set includes a plurality of potential demands corresponding to the current scene of the user;
[0009] For each potential demand in the potential demand set, the comprehensive weight of the potential demand is calculated by using the local large language model;
[0010] Based on the preset trigger logic, the comprehensive weight of the potential demand and the scene complexity of the current scene, it is judged whether to ask for interaction with the cloud large model, and when interaction is needed, a cloud query package is synthesized; the cloud query package is spliced by a core query task and a context vector;
[0011] The cloud large model is used for deep analysis and processing of the cloud query package to obtain cloud health suggestion feedback data;
[0012] The cloud health suggestion feedback data and the updated context vector are used as prompt words to ask the local large language model to generate active care behavior instructions.
[0013] Optionally, the local large language model is a Mistral model, and the cloud large model is a DeepSeekV3 model.
[0014] Optionally, the scene complexity of the current scene is obtained by dynamically evaluating all context vectors using the local large language model, specifically including:
[0015] The scene complexity of the current scene is obtained by weighted summation of all context vectors and the corresponding weight of each context vector using the local large language model.
[0016] Optionally, the calculation formula of the comprehensive weight of the potential demand is as follows:
[0017] W(q(t,i))=alpha*Imp(q(t,i))+beta*Rel(q(t,i),Context(t))+gamma*Urg(q(t,i))+0.7*Context(t);
[0018] Wherein, W(q(t,i)) is the comprehensive weight of the i-th potential demand q(t,i) at t time point; Imp(q(t,i)) represents the importance of the i-th potential demand q(t,i) at t time point; Rel(q(t,i),Context(t)) represents the matching degree of the potential demand q(t,i) and the context vector Context(t) at t time point; Urg(q(t,i)) represents the urgency of the potential demand q(t,i); alpha, beta and gamma are weighting coefficients, and the sum is 0.3.
[0019] Optionally, the preset trigger logic includes whether the scene complexity of the current scene exceeds the complexity threshold, whether the comprehensive weight of any potential demand exceeds the importance threshold, or whether the user has issued an explicit instruction requiring complex understanding.
[0020] Optionally, the personnel health information recommendation method further includes:
[0021] The edge computing device group drives a corresponding execution unit to complete the active information recommendation service according to the generated active care behavior instruction, and the active information recommendation service includes food information recommendation, medicine information recommendation or voice information recommendation.
[0022] Optionally, the personnel health information recommendation method further includes:
[0023] The frequency of data collection and data processing of the edge computing device group is dynamically adjusted according to the scene complexity of the current scene.
[0024] In a second aspect, the present application provides a personnel health information recommendation device, comprising an edge computing device group and a server:
[0025] The edge computing device group is configured to collect multi-modal raw data of a user.
[0026] The server is configured to process the multi-modal raw data using a local large language model to generate a plurality of context vectors.
[0027] The local large language model is used to dynamically evaluate the scene complexity of the current scene according to all the context vectors, and to assign weights to each context vector to obtain a weight value of each context vector.
[0028] The local large language model is used to generate a potential demand set according to each context vector, the weight value of each context vector and a potential demand problem library; the potential demand set includes potential demands corresponding to the current scene of the user.
[0029] For each potential demand in the potential demand set, the local large language model is used to calculate a comprehensive weight of the potential demand.
[0030] Based on a preset trigger logic, the comprehensive weight of the potential demand and the scene complexity of the current scene, it is determined whether to interact with a cloud large model, and when interaction is needed, a cloud query package is synthesized; the cloud query package is spliced from a core query task and a context vector.
[0031] The cloud large model is used to perform deep analysis and processing on the cloud query package to obtain cloud health suggestion feedback data.
[0032] The cloud health suggestion feedback data and the updated context vector are used as prompt words to ask the local large language model, and an active care behavior instruction is generated.
[0033] In a third aspect, the present application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor executes the computer program to realize the personnel health information recommendation method described above.
[0034] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above-mentioned personnel health information recommendation method.
[0035] According to the specific embodiments provided in the present application, the following technical effects are disclosed:
[0036] The present application provides a personnel health information recommendation method, device, equipment and medium, through the edge computing device group to collect the multi-modal original data of the user, and the local large language model is used to process the multi-modal original data to generate a plurality of context vectors; using the local large language model, the scene complexity of the current scene is obtained by dynamically evaluating all context vectors, and the weight value of each context vector is obtained by assigning weights to each context vector; using the local large language model, the potential demand set is generated according to each context vector, the weight value of each context vector and the potential demand problem library; for each potential demand in the potential demand set, the comprehensive weight of the potential demand is calculated by using the local large language model; based on the preset trigger logic, the comprehensive weight of the potential demand and the scene complexity of the current scene, it is judged whether it is necessary to ask for interaction with the cloud large model, and when interaction is needed, a cloud query package is synthesized; the cloud large model is used to analyze and process the cloud query package in depth to obtain cloud health suggestion feedback data; the cloud health suggestion feedback data and the updated context vector are used as prompt words to ask the local large language model, and active care behavior instructions are generated; the present application is based on the collaborative working method of edge computing (local large language model) and cloud large model, which can realize the active care service of intelligent agent through real-time perception of edge, prediction of user's potential demand, and provide timely and personalized health information for users. Through the intelligent scheduling of the coordination of edge computing and cloud large model, combined with the rapid response of edge and the deep cognition of cloud, the depth and accuracy of understanding complex user context are improved, the system response speed and the utilization efficiency of data transmission and computing resources are optimized through dynamic decision of cloud interaction, and the ability to efficiently utilize edge lightweight model and cloud heavyweight model is realized on the premise of guaranteeing fast response and user privacy. The limitations of the prior art in the aspects of double-model coordination efficiency, resource optimization and active service depth are overcome. BRIEF DESCRIPTION OF DRAWINGS
[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any inventive labor.
[0038] Figure 1A flowchart of a personnel health information recommendation method provided for Embodiment 1 of the present application.
[0039] Figure 2 A detailed flowchart of the personnel health information recommendation method provided for Embodiment 1 of the present application.
[0040] Figure 3 A schematic diagram of a double-large model collaborative architecture provided for Embodiment 1 of the present application.
[0041] Figure 4 A structural schematic diagram of a computer device provided for Embodiment 3 of the present application. DETAILED DESCRIPTION
[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.
[0043] The above purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.
[0044] Embodiment 1.
[0045] In an exemplary embodiment, as shown in Figure 1 a personnel health information recommendation method is provided, which comprises the following steps 201 to 207.
[0046] Step 201, the edge computing device group collects multi-modal raw data of a user, and processes the multi-modal raw data using a local large language model to generate a plurality of context vectors.
[0047] Step 202, using the local large language model, the scene complexity of the current scene is obtained by dynamic evaluation according to all context vectors, and a weight value of each context vector is obtained by assigning a weight to each context vector.
[0048] Step 203, using the local large language model, a potential demand set is generated according to each context vector, the weight value of each context vector, and a potential demand problem library; the potential demand set includes potential demands corresponding to the current scene of the user.
[0049] Step 204, for each potential demand in the potential demand set, the comprehensive weight of the potential demand is calculated using the local large language model.
[0050] Step 205, based on the preset trigger logic, the comprehensive weight of the potential demand and the scene complexity of the current scene, judge whether it is necessary to ask the cloud large model for interaction, when interaction is needed, synthesize the cloud query package; the cloud query package is spliced by the core query task and the context vector.
[0051] Step 206, using the cloud large model to perform deep analysis and processing on the cloud query package, obtaining cloud health suggestion feedback data.
[0052] Step 207, taking the cloud health suggestion feedback data and the updated context vector as the prompt word, asking the local large language model, generating active care behavior instructions.
[0053] Implementing the above steps 201 to 207, the control method of edge computing (local large language model) and cloud large model collaborative work can realize the active care service of intelligent agent through edge real-time perception and prediction of user's potential demand, and provide timely and personalized health information for users. Through intelligent scheduling of edge computing and cloud large model collaboration, combining the fast response of edge and the deep cognition of cloud, the depth and accuracy of complex user context understanding are improved. Through dynamic decision-making of cloud interaction, the system response speed and the utilization efficiency of data transmission and computing resources are optimized. It has the ability to efficiently utilize edge lightweight model and cloud heavyweight model under the premise of guaranteeing fast response and user privacy, and overcomes the limitations of existing technology in double model collaboration efficiency, resource optimization and active service depth. Through the intelligent scheduling of edge model and cloud model collaborative control algorithm, the application integrates multi-party information to generate personalized behavior instructions, enhances the personalized service ability and user emotional connection; at the same time, focusing on edge processing and optional local emergency plan design, it helps to protect user privacy and service continuity. The application realizes the efficient collaboration of end-to-cloud intelligence, effectively improves the comprehensive performance of active care intelligent agent in service initiative, context understanding, system efficiency and user experience.
[0054] The purpose of the present application is to provide a personnel health information recommendation method based on double large model cooperation, which is used to drive active caring intelligent agent, provide health information recommendation service for users, and solve the deficiencies of existing family health intelligent system in active perception, context understanding and personalized information recommendation. Through the intelligent recommendation and scheduling mechanism proposed in the present application, an overall intelligent agent system capable of active perception, deep understanding and personalized health information recommendation is constructed to overcome the limitations of existing technology in double model cooperation efficiency, resource optimization and active service depth. Through efficient and intelligent scheduling of the cooperative work of edge lightweight model and cloud heavyweight model, an overall intelligent agent capable of low delay response, deep understanding of user demand and active provision of health recommendation information is constructed. The present application provides an active intelligent recommendation algorithm based on double large model cooperation, including food recommendation, drug recommendation and other information recommendation.
[0055] As shown in Figure 2 and Figure 3 The edge computing device group collects video data, voice data and multi-modal raw data synthesized by health monitoring devices of the user environment, processes, filters and fuses the initial context representation using the edge lightweight model (i.e. local large language model) to generate the initial context representation, i.e. the scene vector. Figure 3 The third party device in the formula can be a health status monitoring bracelet that monitors the user's blood pressure and other data. The host computer controls the edge computing device and the cloud large model through the control layer through sync and multi technology.
[0056] The edge computing device group includes cameras, microphones, health monitoring devices, etc. The user's data stream is obtained through the edge computing device group, including camera picture stream, microphone audio stream and health monitoring device data stream, etc. According to the data stream, a multi-modal large model (i.e. local large language model) comprehensive processing method is used for real-time context perception and complexity evaluation to obtain the scene vector and corresponding weight value in multiple directions including emotion, health status, sick state, etc. The above-mentioned local large language model can be a Mistral model.
[0057] The edge computing device group continuously collects multi-modal raw data stream of the user. The local Mistral small model converts the multi-modal raw data into structured context vector Context(t) containing timestamp, user activity state, location, preliminary emotional tendency, key physiological indicators and environmental parameters, etc. The key physiological indicators monitored by the health monitoring device include heart rate, blood pressure, blood oxygen saturation, sleep state, etc. The user activity type includes standing, sitting, lying, etc. The environmental parameters can include the distance of the user from the special place, the object the user is interacting with identified through the camera picture stream, other important objects in the environment, etc. The special place can be the kitchen, bedroom, etc.
[0058] Subsequently, the local large language model dynamically evaluates the scene complexity C(Context(t)) of the current scene according to the context vector Context(t) of the current time point. The evaluation can be based on the weighted combination of each feature in the context vector or a small machine learning model (i.e., the Mistral model). For example, in step 202 described above, the scene complexity of the current scene is dynamically evaluated according to all context vectors using the local large language model, which specifically includes: using the local large language model to perform weighted summation according to all context vectors and the corresponding weights of each context vector to obtain the scene complexity of the current scene.
[0059] The scene complexity C(Context(t)) based on the context vector Context(t) can be represented as:
[0060] C(Context(t)) = f complex(w_A*Score(A(t))+w_L*ChangeRate(L(t))+…);
[0061] where f complex() represents a scene complexity function; Score(A(t)) is the complexity score of the user activity type A(t) at time t; ChangeRate(L(t)) is the change rate of the user location L(t) at time t; w_A and w_L are the corresponding weights, which are set artificially according to user needs. The context vector includes user activity state, location, preliminary sentiment tendency, key physiological indicators, and environmental parameters, etc.
[0062] When the context vector of the current scene awareness information is analyzed by the local large language model (Mistral model), the local large language model will generate a context vector containing multiple directions such as emotion, health status, and sick state, and give a value of 0-1 as the weight value of its context vector, preparing to predict the user's needs.
[0063] According to the multiple context vectors and the corresponding weight values obtained in the above steps, potential demand prediction, question raising, and priority sorting are performed to obtain a plurality of possible questions and the corresponding weights.
[0064] Based on the current context vector Context(t) and the potential demand question library, the local large language model predicts and generates a set of potential demands or questions that the user may currently have (i.e., a potential demand set), and generates a question table Q(t) containing multiple potential demands or questions q(t,i). Q(t) represents the potential demand set predicted at time t, and q(t,i) represents the i-th potential demand at time t.
[0065] Potential demand example: For example, it is monitored that the user is processing shrimp, and it is given whether a shrimp-related recipe is needed; the user's blood pressure is higher than the set value, and the potential demand is whether to take medicine, etc.
[0066] For each potential demand q(t, i) in the potential demand set, the algorithm calculates its comprehensive weight W(q(t, i)), which comprehensively considers the scenario vector Context(t) just calculated, the importance Imp(q(t, i)) of the potential demand itself, the matching degree Rel(q(t, i), Context(t)) of the potential demand and the current context vector, and the potential urgency Urg(q(t, i)). The calculation formula of the comprehensive weight of the potential demand is as follows:
[0067] W(q(t, i)) = alpha * Imp(q(t, i)) + beta * Rel(q(t, i), Context(t)) + gamma * Urg(q(t, i)) + 0.7 * Context(t);
[0068] Wherein, W(q(t, i)) is the comprehensive weight of the i-th potential demand q(t, i) at t time point; Imp(q(t, i)) represents the importance of the i-th potential demand q(t, i) at t time point; Rel(q(t, i), Context(t)) represents the matching degree of the potential demand q(t, i) and the context vector Context(t) at t time point; Urg(q(t, i)) represents the urgency of the potential demand q(t, i); alpha, beta and gamma are adjustable weighting coefficients, the sum is 0.3, which is set according to user experience.
[0069] It should be noted that the weighting coefficient beta of the matching degree of the potential demand and the current context vector will be determined according to the user preference, which means that the user can set the system (through the mobile APP or the network link that can be connected to the device), such as the user wants the system to care a little more or a little less, or can set to care more about health problems or more about people's emotions. These settings will affect the weighting coefficient beta of the matching degree of the potential demand and the current context vector.
[0070] According to the comprehensive weights of all potential demands in descending order, it is judged whether to ask the cloud large model for interaction from the potential demand with the largest comprehensive weight value.
[0071] The comprehensive weight of the potential demand obtained above is the weight given to each question (potential demand) after considering multiple factors, which determines whether the potential demand is to be asked. After obtaining the question table (i.e., the set of potential demands) and the corresponding comprehensive weight, the system will monitor whether there is a question with a comprehensive weight exceeding a certain threshold, and if so, proceed to the next step of interacting with the cloud large model. The cloud large model is the DeepSeekV3 model.
[0072] Meanwhile, the personnel health information recommendation method further comprises: dynamically adjusting the frequency of data collection and data processing of the edge computing device group according to the scene complexity of the current scene. The frequency F_edge of edge data collection and processing is dynamically adjusted according to the scene complexity C(Context(t)) in order to save computing power in simple scenes or improve reaction speed and judgment accuracy in complex scenes. The adjustment formula is:
[0073] F_edge=F_min+(F_max-F_min)*sigma(k*(C(Context(t))-c_0));
[0074] Where F_min and F_max are the minimum and maximum frequencies respectively, sigma is an activation function (such as the Sigmoid function), and k and c_0 are adjustment parameters.
[0075] When the comprehensive weight of the potential demand q(t,i) obtained by the above steps exceeds the set threshold, the cloud is interacted with through API calling to obtain the cloud health suggestion feedback data. Based on the preset trigger logic, it is determined whether to ask the cloud large model for interaction, and a cloud query package containing the core question and auxiliary scene information is synthesized.
[0076] According to the preset trigger logic, it is judged whether the interaction with the cloud large model needs to be started. The preset trigger logic can comprehensively consider whether the scene complexity C(Context(t)) of the current scene exceeds the complexity threshold Thresh complex, whether the comprehensive weight W(q(t, i)) of any potential demand exceeds the corresponding importance threshold Thresh weight, or whether the user has issued an explicit instruction requiring complex understanding. If it is judged that cloud interaction is needed, the cloud query information Json package (i.e. cloud query package) Q_cloud formed by assembling the core query task Task_core and the current context vector Context(t) is uploaded to the cloud large model for deep analysis, processing, and the result is returned to the edge computing device. The API of the corresponding private cloud service platform is used to access the cloud large model, upload the file and the prompt word, and obtain the cloud health suggestion feedback data fed back by the cloud large model. The priority of the three conditions in the preset trigger logic from high to low is whether the user has issued an explicit instruction requiring complex understanding, whether the comprehensive weight W(q(t, i)) of any potential demand exceeds the corresponding importance threshold Thresh weight, and whether the scene complexity C(Context(t)) of the current scene exceeds the complexity threshold Thresh complex.
[0077] The cloud health suggestion feedback data of the cloud large model obtained according to the above steps is used to obtain accurate information that can provide services for the user. The cloud large model feeds back the Json file containing the required cloud health suggestion feedback data to the edge computing device for health information recommendation.
[0078] The edge computing device integrates the cloud health suggestion feedback data, the updated context vector and the user preference, generates and executes the active care behavior instruction, such as controlling the Internet of Things device, active voice interaction or information push.
[0079] The edge end uploads the cloud query information package Q_cloud to the cloud large model for reaction calculation. The cloud large model refers to the DeepSeekV3 model. The upper host can receive the cloud feedback result (cloud health suggestion feedback data) R_cloud, and use it and the current latest local context (i.e. updated context vector) Context(t_prime) (t_prime represents the updated time point) as a prompt word to ask the local large language model, and comprehensively consider to generate one or a series of active care behavior instructions Action(t) to be executed finally. Action(t) represents the active care behavior instruction at time point t.
[0080] Subsequently, the personnel health information recommendation method provided by the embodiment further includes: the edge computing device drives a corresponding execution unit to complete the active information recommendation service according to the generated multiple active care behavior instructions Action(t), and the active information recommendation service includes food information recommendation, medicine information recommendation or voice information recommendation.
[0081] The system can selectively collect user satisfaction feedback data Feedback(t) (indicating user feedback collected at time point t) on this information recommendation, which refers to whether the user performs activities or has positive operation feedback according to the content of the information recommendation, rather than ignoring the feedback result of the system, and is used for long-term offline updating of a weight addition value k of a context vector Context(t) in the system. If the user is satisfied, the weight addition value will be increased by a little, otherwise the addition value will be reduced by a little. The k value is only multiplied by the context vector once, so as to realize continuous improvement of algorithm performance and user experience. The specific formula is as follows:
[0082] Content'(t) = Content(t) * k;
[0083] Wherein, Content'(t) represents the context vector after the addition in the long-term offline updating system, and k is the addition value of the context weight.
[0084] In another exemplary embodiment of the application, in the process of interaction with the cloud, when the cloud communication is abnormal, the embodiment can also perform basic active care based on the local plan. When the cloud is abnormal, the basic problem is given in the local plan library.
[0085] Through the combination of the above processing flow, the whole double-large model solution realizes accurate perception of user state, intelligent prediction of potential demand, optimization of computing resources, and personalized generation and continuous optimization of health information. Through intelligent scheduling of edge computing resources and cloud large model resources, active, personalized and efficient data recommendation care services are realized.
[0086] The application dynamically decides the method of collaborative interaction and information transmission between edge computing and cloud large model according to real-time scene complexity and user potential demand weight, especially in the interaction of double-large model; predicts user potential demand based on current context, and calculates its comprehensive priority weight to drive active health information recommendation. The application generates personalized behavior instructions by integrating multiple information, enhances personalized service capability and user emotional connection; at the same time, focuses on edge processing and optional local emergency plan design, which helps to protect user privacy and service continuity. The application realizes efficient collaboration of end-to-cloud intelligence, effectively improves the comprehensive performance of active care intelligent agent in service initiative, context understanding, system efficiency and user experience.
[0087] The application further provides an application scenario of the personnel health information recommendation method. Specifically, the personnel health information recommendation method provided in the embodiment can be applied in an intelligent health information recommendation scenario. The intelligent health information recommendation scenario includes a content production link, a content processing link and a content distribution link. The multi-modal original data of the user enters the content processing link from the content production link, obtains a corresponding active care behavior instruction, and enters the downstream content distribution link. The personnel health information recommendation method provided in the embodiment belongs to the content processing link. Specifically, in the content processing link process for the multi-modal original data of the user, the local large language model can be used to process the multi-modal original data to generate a plurality of context vectors. The local large language model is used to dynamically evaluate the scene complexity of the current scene according to all the context vectors, and a weight value of each context vector is obtained by assigning a weight to each context vector. The local large language model is used to generate a potential demand set according to the context vectors, the weight value of each context vector and the potential demand problem library. For each potential demand in the potential demand set, the local large language model is used to calculate the comprehensive weight of the potential demand. Based on the preset trigger logic, the comprehensive weight of the potential demand and the scene complexity of the current scene, it is determined whether to ask for interaction with the cloud large model. When interaction is needed, a cloud query package is synthesized. The cloud large model is used to perform deep analysis and processing on the cloud query package to obtain cloud health suggestion feedback data. The cloud health suggestion feedback data and the updated context vector are used as prompt words to ask the local large language model, and an active care behavior instruction is generated.
[0088] Embodiment 2.
[0089] Based on the same inventive concept, the embodiment of the application further provides a personnel health information recommendation device for implementing the personnel health information recommendation method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more personnel health information recommendation device embodiments provided below can refer to the limitations of the personnel health information recommendation method described above, which will not be repeated here.
[0090] In one exemplary embodiment, a personnel health information recommendation device is provided, including an edge computing device group and a server.
[0091] The edge computing device group is configured to collect multi-modal original data of a user.
[0092] The server is configured to process the multi-modal original data using a local large language model to generate a plurality of context vectors.
[0093] The scene complexity of the current scene is obtained by using the local large language model and dynamically evaluating all the context vectors, and a weight value of each context vector is obtained by assigning a weight to each context vector;
[0094] A potential demand set is generated by using the local large language model, according to the context vectors, the weight value of each context vector, and the potential demand problem library; the potential demand set includes potential demands corresponding to the current scene of the user;
[0095] For each potential demand in the potential demand set, the comprehensive weight of the potential demand is calculated by using the local large language model;
[0096] Based on the preset trigger logic, the comprehensive weight of the potential demand, and the scene complexity of the current scene, it is determined whether to ask for interaction with the cloud large model, and when interaction is needed, a cloud query package is synthesized; the cloud query package is composed of a core query task and a context vector;
[0097] The cloud large model is used to perform deep analysis and processing on the cloud query package to obtain cloud health suggestion feedback data;
[0098] The cloud health suggestion feedback data and the updated context vector are used as prompt words to ask the local large language model, and an active care behavior instruction is generated.
[0099] Embodiment 3.
[0100] In an exemplary embodiment, a computer device, which can be a server or a terminal, has an internal structure as shown in Figure 4 The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is used to store health information recommendation data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a personnel health information recommendation method.
[0101] Those skilled in the art can understand, Figure 4The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0102] In an exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above method embodiments when executing the computer program.
[0103] Embodiment 4.
[0104] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.
[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.
[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0107] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.
[0108] The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.
[0109] The principles and implementation manners of the present application are described herein by using specific examples, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will have changes. In conclusion, the content of the specification should not be understood as a limitation of the present application.
Claims
1. A method for recommending personnel health information, characterized in that, The method for recommending personnel health information includes: Edge computing device clusters collect users' multimodal raw data and use local large language models to process the multimodal raw data to generate multiple contextual vectors; Using a local large language model, the scene complexity of the current scene is dynamically evaluated based on all context vectors, and weights are assigned to each context vector to obtain the weight value of each context vector. Using a local large language model, a potential demand set is generated based on each context vector, the weight value of each context vector, and a potential demand question library; the potential demand set includes several potential demands corresponding to the user's current scenario. For each potential demand in the set of potential demands, the comprehensive weight of the potential demand is calculated using the local large language model; Based on the preset triggering logic, the comprehensive weight of the potential needs, and the scenario complexity of the current scenario, it is determined whether it is necessary to ask questions and interact with the cloud-based large model. When interaction is required, a cloud query package is synthesized. The cloud query package is composed of core query tasks and context vectors. The cloud query package is analyzed and processed in depth using the cloud-based big model to obtain cloud-based health advice feedback data. Using the cloud-based health advice feedback data and the updated contextual vector as prompts, questions are posed to the local large language model to generate proactive care behavior instructions.
2. The method for recommending personnel health information according to claim 1, characterized in that, The local large language model is the Mistral model, and the cloud-based large model is the DeepSeekV3 model.
3. The method for recommending personnel health information according to claim 1, characterized in that, By utilizing a local large language model, the scene complexity of the current scene is dynamically evaluated based on all context vectors, specifically including: By utilizing a local large language model, the scene complexity of the current scene is obtained by weighted summation based on all context vectors and the corresponding weights of each context vector.
4. The method for recommending personnel health information according to claim 1, characterized in that, The formula for calculating the comprehensive weight of the potential demand is as follows: W(q(t,i))=alpha*Imp(q(t,i))+beta*Rel(q(t,i),Context(t))+ gamma*Urg(q(t,i))+0.7*Context(t); Where W(q(t,i)) is the comprehensive weight of the i-th potential demand q(t,i) at time t; Imp(q(t,i)) represents the importance of the i-th potential demand q(t,i) at time t; Rel(q(t,i),Context(t)) represents the matching degree between the potential demand q(t,i) and the context vector Context(t) at time t; Urg(q(t,i)) represents the urgency of the potential demand q(t,i); alpha, beta, and gamma are weighting coefficients, with a total sum of 0.
3.
5. The method for recommending personnel health information according to claim 1, characterized in that, The preset triggering logic includes whether the complexity of the current scenario exceeds a complexity threshold, whether the overall weight of any potential demand exceeds an importance threshold, or whether the user has issued an explicit instruction that requires complex understanding.
6. The method for recommending personnel health information according to claim 1, characterized in that, The method for recommending personnel health information also includes: The edge computing device cluster drives the corresponding execution units to complete proactive information recommendation services based on the generated proactive care behavior instructions. The proactive information recommendation services include food information recommendations, drug information recommendations, or voice information recommendations.
7. The method for recommending personnel health information according to claim 1, characterized in that, The method for recommending personnel health information also includes: The frequency of data acquisition and processing by the edge computing device cluster is dynamically adjusted based on the complexity of the current scenario.
8. A personnel health information recommendation device, characterized in that, The personnel health information recommendation device includes an edge computing device cluster and a server: The edge computing device group is used to collect users' multimodal raw data; The server is used to: process multimodal raw data using a local large language model to generate multiple context vectors; Using a local large language model, the scene complexity of the current scene is dynamically evaluated based on all context vectors, and weights are assigned to each context vector to obtain the weight value of each context vector. Using a local large language model, a potential demand set is generated based on each context vector, the weight value of each context vector, and a potential demand question library; the potential demand set includes several potential demands corresponding to the user's current scenario. For each potential demand in the set of potential demands, the comprehensive weight of the potential demand is calculated using the local large language model; Based on the preset triggering logic, the comprehensive weight of the potential needs, and the scenario complexity of the current scenario, it is determined whether it is necessary to ask questions and interact with the cloud-based large model. When interaction is required, a cloud query package is synthesized. The cloud query package is composed of core query tasks and context vectors. The cloud query package is analyzed and processed in depth using the cloud-based big model to obtain cloud-based health advice feedback data. Using the cloud-based health advice feedback data and the updated contextual vector as prompts, questions are posed to the local large language model to generate proactive care behavior instructions.
9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the personnel health information recommendation method according to any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the personnel health information recommendation method as described in any one of claims 1-7.