A smart home scene design intelligent agent construction method and device and related equipment
By decomposing smart home scenarios into multiple descriptive dimensions and constructing a dialogue dataset, training a large language model, and utilizing multi-agent system collaboration, the problem of insufficient understanding of complex business logic by general models is solved, and efficient and accurate generation of customized design solutions is achieved.
Patent Information
- Application Number
- CN202611436978.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, general-purpose large language models lack sufficient understanding of the complex business logic of smart home design, and the generated solutions are prone to business logic errors, making it difficult to meet the needs of actual applications.
The smart home scenario is decomposed into multiple independent descriptive dimensions, a dialogue dataset is constructed and a large language model is trained to generate design agents, and design solutions are generated through collaboration of a multi-agent system.
It improves the accuracy and stability of the generated schemes, reduces the output illusion caused by multi-task coupling of a single intelligent agent, and realizes the efficient generation of customized design schemes.
Smart Images

Figure CN123070169A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of smart home large-scale models, specifically to a method, device, and related equipment for constructing intelligent agents in smart home scene design. Background Technology
[0002] In recent years, with the improvement of people's living standards and requirements for the quality of living, smart home products have gradually evolved from single-function independent devices to systematic and scenario-based whole-house smart home solutions. Currently, smart homes cover multiple application areas such as home security, environmental regulation, energy management, convenient control, and health monitoring.
[0003] How to efficiently and rationally configure numerous smart devices to form scene combinations that meet user preferences has become a key aspect of improving user experience. The whole-house smart home scene design model is a new type of intelligent auxiliary tool proposed to meet this need. Its core function is to automatically or semi-automatically generate home scene configuration solutions based on multi-dimensional information such as residential space layout, user living habits, environmental changes, and device capabilities.
[0004] Currently, existing technologies mainly use a general large language model combined with prompt word engineering to generate solutions. That is, the model output is constrained by writing large sections of rule prompt words. However, the general large language model has insufficient understanding of the complex business logic in smart home design, and the generated solutions are prone to business logic errors. The accuracy and stability of the solution output are difficult to meet the needs of practical applications. Summary of the Invention
[0005] The purpose of this application is to provide a method, apparatus, and related equipment for constructing intelligent agents in smart home scene design, which can solve the problems that existing models are difficult to understand the complex business logic of smart home design, and the generated solutions are prone to business logic errors and fail to meet the needs of actual applications.
[0006] To achieve the above objectives, this application provides a method for constructing intelligent agents in smart home scene design, comprising:
[0007] Construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values. Each combination of parameter values for multiple descriptive dimensions corresponds to a smart home scene.
[0008] Construct a dialogue dataset, which includes multiple dialogue data, each corresponding to a smart home scenario. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scenario and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scenario.
[0009] The large language model is trained based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on the user's interactive text.
[0010] A multi-agent system is constructed, comprising an interactive agent and a design agent; the interactive agent interacts with the user to obtain key design information corresponding to parameter values of multiple descriptive dimensions, and transmits the key design information to the design agent, which generates a design scheme based on the key design information.
[0011] Optionally, the plurality of independent description dimensions include at least two of the following:
[0012] The floor plan dimension reflects the room and living room structure of the house, and its multiple parameter values represent different room and living room structure types;
[0013] The dialogue hierarchy dimension reflects the completeness of information provided during user interaction, and its multiple parameters take different levels of completeness.
[0014] The intelligent function scope dimension reflects the scope of a user's intelligent function needs, and its multiple parameter values take different scope types.
[0015] The family member dimension reflects the composition of a household, and its multiple parameters can take different member composition types.
[0016] The boundary condition dimension is used to reflect the special constraints in user requirements, and its multiple parameters can take different special constraint types.
[0017] The smart function density dimension reflects the richness of smart functions in a smart home scenario, and its parameter values take different levels of richness.
[0018] Optionally, constructing the dialogue dataset includes:
[0019] Calculate all parameter values for multiple description dimensions to obtain the theoretical full set of scenarios, which includes all smart home scenarios;
[0020] Select a preset number of basic scenarios from the theoretical full set of scenarios;
[0021] For each basic scenario, generate one original dialogue data, and generate multiple variant dialogue data based on the original dialogue data; both the original dialogue data and the variant dialogue data include at least one input field and at least one output field;
[0022] The dialogue dataset is constructed based on the original dialogue data and the variant dialogue data.
[0023] Optionally, a preset number of basic scenarios are selected from the theoretical full set of scenarios, including:
[0024] Select a first preset number of first basic scenarios from the theoretical full set of scenarios, wherein the first basic scenarios are smart home scenarios with the highest business priority;
[0025] Select a second preset number of second basic scenarios from the theoretical full set of scenarios. The second basic scenarios are smart home scenarios with medium business priority.
[0026] Select a third preset number of third basic scenarios from the theoretical full set of scenarios. The third basic scenarios are smart home scenarios with the lowest business priority.
[0027] Among them, the first preset quantity is greater than the second preset quantity, and the second preset quantity is greater than the third preset quantity.
[0028] Optionally, the step of generating one original dialogue data for each basic scenario and generating multiple variant dialogue data based on the original dialogue data includes:
[0029] The input fields for constructing the original dialogue data include multi-turn dialogue text, which at least includes apartment type confirmation information, family member confirmation information, and special needs expression information. The apartment type confirmation information is an apartment type description text corresponding to the apartment type dimension of the basic scenario; the family member confirmation information is a family member description text corresponding to the family member dimension of the basic scenario; and the special needs expression information is a special needs description text corresponding to the boundary condition dimension of the basic scenario. The interaction rounds of the multi-turn dialogue text and the methods of providing the apartment type confirmation information, family member confirmation information, and special needs expression information are determined by the parameter values of the dialogue hierarchy dimension of the basic scenario.
[0030] The output fields of the original dialogue data are constructed, and the output fields include multiple room configuration schemes. Each room configuration scheme includes at least a room name, member tags, and a smart function configuration list. The room name is determined by the parameter value of the apartment type dimension of the basic scene. The member tags are determined by the parameter value of the family member dimension of the basic scene. The smart function configuration list includes multiple smart function items, and the number of smart function items is determined by the parameter value of the smart function density dimension of the basic scene. Multiple room configuration schemes constitute the design scheme.
[0031] The original dialogue data is adjusted based on the preset variant strategy to obtain multiple variant dialogue data; the variant strategy is used to change the expression style of the input fields in the original dialogue data, and / or to change the room configuration scheme of the output fields in the original dialogue data.
[0032] Optionally, when the parameter value of the boundary condition dimension of the basic scenario causes the first room configuration scheme corresponding to the first member label in the output field to be excluded, the first member label is transferred to a room configuration scheme of the same type as the first room configuration scheme.
[0033] Optionally, when there are multiple room configuration schemes of the same type as the first room configuration scheme, the member tag is transferred to the room configuration scheme with the highest priority according to a preset priority order.
[0034] Optionally, the multi-agent system further includes: a product matching agent;
[0035] The product matching agent receives the design scheme output by the design agent, matches specific products from the product database according to the design scheme, and outputs the final scheme containing specific product recommendations.
[0036] This application also provides a smart home scene design intelligent agent construction device, including:
[0037] The dimension construction module is used to construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values, and each combination of parameter values of the multiple descriptive dimensions corresponds to a smart home scene.
[0038] The dataset construction module is used to construct a dialogue dataset, which includes multiple dialogue data, each corresponding to a smart home scene. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scene and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scene.
[0039] The training module is used to train the large language model based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on the user's interactive text.
[0040] A multi-agent construction module is used to construct a multi-agent system, which includes an interactive agent and a design agent. The interactive agent interacts with the user to obtain key design information corresponding to parameter values of multiple description dimensions, and transmits the key design information to the design agent. The design agent generates a design scheme based on the key design information.
[0041] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the above-described method for constructing intelligent agents for smart home scene design.
[0042] This application also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the steps in the above-described method for constructing intelligent agents in smart home scene design.
[0043] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the steps in the above-described method for constructing intelligent agents in smart home scene design.
[0044] Compared with existing technologies, the intelligent agent construction method, device and related equipment for intelligent home scene design provided in this application decompose the intelligent home scene into multiple independent descriptive dimensions and define the intelligent home scene by combining parameter values. It does not rely on human experience, so that the scene coverage of the training data is systematic and comprehensive. The trained design agent is dedicated to generating solutions in the multi-agent system and cooperates with the interactive agent. The task boundaries are clear, reducing the output illusion caused by the multi-task coupling of a single agent, and enabling the system to generate customized design solutions with high accuracy. Attached Figure Description
[0045] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0046] Figure 1 This is one of the flowcharts illustrating a method for constructing an intelligent agent in a smart home scene design according to an embodiment of this application.
[0047] Figure 2 This is a second flowchart illustrating a method for constructing an intelligent agent in a smart home scene design, as provided in one embodiment of this application.
[0048] Figure 3The third flowchart illustrates a method for constructing an intelligent agent in a smart home scene design, as provided in one embodiment of this application.
[0049] Figure 4 This is a schematic diagram of the structure of an intelligent agent construction device for designing a smart home scene, provided in one embodiment of this application. Detailed Implementation
[0050] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.
[0051] Reference Figure 1 , Figure 1 This is a flowchart of a smart home scene design intelligent agent construction method provided in this embodiment, such as... Figure 1 As shown, the method includes:
[0052] Step 101: Construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values. Each combination of parameter values for multiple descriptive dimensions corresponds to a smart home scene.
[0053] In one embodiment of the present invention, the plurality of mutually independent description dimensions include at least two of the following:
[0054] The apartment layout dimension reflects the room and living room structure of the house, and its multiple parameter values represent different room and living room structure types. The parameter values of the apartment layout dimension include different room and living room structure types such as 2R (two bedrooms and one living room), 3R (three bedrooms and one living room), and 4R (four bedrooms and two living rooms). Different parameter values correspond to different default functional room lists. For example, 2R, that is, a two-bedroom, one-living room apartment, corresponds to the preset functional room list corresponding to the 2R parameter value, such as the living room, master bedroom, secondary bedroom, kitchen, bathroom, and balcony. 3R, that is, a three-bedroom, one-living room apartment, corresponds to the preset functional room list corresponding to the 3R parameter value, such as the living room, master bedroom, secondary bedroom one, secondary bedroom two, kitchen, bathroom, and balcony.
[0055] The dialogue hierarchy dimension reflects the completeness of information provided during user interaction, with multiple parameters taking different completeness levels. In one embodiment of the invention, the parameters of the dialogue hierarchy dimension include multiple completeness levels such as L1, L2, L3, and L4. L1 is a standard dialogue, where the user provides information item by item according to the normal consultation rhythm; L2 is a simplified dialogue, where the user provides all key information in one round of dialogue; L3 is a detailed dialogue, where the information provided by the user is incomplete, and the system needs to ask follow-up questions multiple times to fill in the missing information; L4 is a silent dialogue, where the user does not explicitly answer key information throughout the process. In another embodiment of the invention, L1 means the information provided by the user is complete, and the system does not need to ask follow-up questions; L2 means the information provided by the user is missing one type of key information, and the system needs to ask follow-up questions once; L3 means the information provided by the user is missing two or more types of key information, and the system needs to ask follow-up questions at least twice; L4 means the user does not provide key information throughout the process, and the system needs to trigger the default value.
[0056] The intelligent function scope dimension reflects the range of user's intelligent function needs. Its multiple parameter values represent different scope types. In one embodiment of the invention, the parameter values of the intelligent function scope dimension include multiple different scope types such as S1, S2, S3, S4, S5, and S6. Specifically, S1 represents whole-house intelligence, where the user's intelligent function needs are designed for all rooms; S2 represents single-room intelligence, where the user's intelligent function needs are designed for only one room; S3 represents partial whole-house intelligence, where the user's intelligent function needs are designed for multiple rooms; S4 represents cross-room linkage, where the user's intelligent function needs are to achieve linkage between multiple rooms; S5 represents specific function configuration, where the user's intelligent function needs specify certain specific intelligent functions; and S6 represents customized function linkage, where the user's intelligent function needs specify the linkage logic for certain specific intelligent functions.
[0057] The family member dimension is used to reflect the composition of family members, and its multiple parameter values take different member composition types. In one embodiment of the present invention, the parameter values of the family member dimension include different member composition types such as M0, M1, M2, M3, and M4. Among them, M0 means no special members and all family members are ordinary adults, M1 means including elderly members, M2 means including child members, M3 means including both elderly and child members, and M4 means including domestic service personnel.
[0058] The boundary condition dimension reflects specific constraints in user requirements, and its multiple parameter values represent different types of specific constraints. In one embodiment of the present invention, the parameter values of the boundary condition dimension include different types of specific constraints such as E0, E1, E2, E3, E4, and E5. Among them, E0 represents no boundary condition, where there are no specific constraints in the user requirements, and the configuration is processed according to standard smart functions; E1 represents excluding or deleting specific smart functions, i.e., the user requirements indicate that a certain smart function is not needed, such as the user explicitly stating that security functions are not needed; E2 represents the user specifying a specific device or brand, such as the user requiring a specific router model; E3 represents the user simultaneously requesting mutually exclusive boundary conditions, i.e., there are rule conflicts between the user's requirements; E4 represents the user only requiring the design of certain rooms, i.e., the user specifies that only certain rooms should be designed; E5 represents the user inquiring about the physical compatibility of smart functions with rooms, i.e., the user inquiring whether a certain smart function is suitable for installation in a certain room.
[0059] The intelligent function density dimension is used to reflect the richness of intelligent functions in a smart home scenario. Its parameter values take different richness levels. In one embodiment of the present invention, the parameter values of the intelligent function density dimension include different richness levels such as T1 and T2. T1 is the basic density, that is, each room is equipped with a basic number of intelligent function items; T2 is the flagship density, that is, each room is equipped with a larger number of intelligent function items and additional value-added functions.
[0060] The description dimensions described in this invention can be set according to actual business needs, and usually include two or more description dimensions. It can include one or more of the above six dimensions, or other dimensions other than the above six dimensions. Each description dimension can also be set with one or more parameter values according to actual business needs. This application does not make specific limitations in this regard.
[0061] In one embodiment of the present invention, all six descriptive dimensions are used to define a smart home scene. The six descriptive dimensions are independent of each other. Each combination of parameter values for the multiple descriptive dimensions corresponds to a smart home scene. The smart home scene is a total scene composed of one or more executable smart functions, such as "security" and "emergency call". Each smart home scene can be represented by a combination of parameter values that includes the above six descriptive dimensions, thereby realizing the accurate definition and indexing of smart home scenes. By decomposing smart home scenes into multiple independent descriptive dimensions and defining smart home scenes by a combination of parameter values, the system does not rely on human experience, making the scene coverage of the training data systematic and comprehensive.
[0062] Step 102: Construct a dialogue dataset, which includes multiple dialogue data, each corresponding to a smart home scene. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scene and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scene.
[0063] In one embodiment of the present invention, such as Figure 2 As shown, step 102 includes:
[0064] Step 1021: Calculate all parameter value combinations for multiple description dimensions to obtain the theoretical full set of scenarios, which includes all smart home scenarios;
[0065] Step 1022: Select a preset number of basic scenarios from the theoretical full set of scenarios;
[0066] Specifically, firstly, all parameter value combinations of multiple description dimensions are calculated, and the smart home scenarios represented by all parameter value combinations are aggregated to obtain the theoretical full set of scenarios. The number of scenarios in the theoretical full set of scenarios is the Cartesian product of the number of parameter values of each description dimension. According to the parameter values described in the above embodiment, the number of smart home scenarios in the theoretical full set of scenarios can reach thousands to tens of thousands. In order to achieve maximum effective coverage with limited resources, this embodiment introduces a priority hierarchical sampling strategy, that is, to select a preset number of basic scenarios from the theoretical full set of scenarios.
[0067] In one embodiment of the present invention, such as Figure 3 As shown, step 1022 includes:
[0068] Select a first preset number of first basic scenarios from the theoretical full set of scenarios, wherein the first basic scenarios are smart home scenarios with the highest business priority;
[0069] Select a second preset number of second basic scenarios from the theoretical full set of scenarios. The second basic scenarios are smart home scenarios with medium business priority.
[0070] Select a third preset number of third basic scenarios from the theoretical full set of scenarios. The third basic scenarios are smart home scenarios with the lowest business priority.
[0071] Among them, the first preset quantity is greater than the second preset quantity, and the second preset quantity is greater than the third preset quantity.
[0072] Specifically, based on business priority, all smart home scenarios are divided into three priority levels. The business priority is determined by the frequency of occurrence and necessity of coverage of the smart home scenario in actual business. The first level has the highest priority and is the core high-frequency scenario. The smart home scenario with the highest frequency of occurrence and the most representative in business is selected as the first basic scenario. The specific sampling rules include covering all combinations of house types and all product grades; giving priority to covering high-frequency family member combinations such as M0, M1, and M2; and focusing on covering core boundary condition types such as E0, E1, E2, and E4. The sampling quantity of the first basic scenario accounts for a first preset quantity of the total number of basic scenarios, for example, it can be 60% to 70%.
[0073] The second level is of medium priority. It selects boundary scenarios that are not covered or insufficiently covered by the first level but have a certain probability of occurrence in actual business as the second basic scenarios. The specific sampling rules include: non-whole-house type scenarios with supplementary boundary condition dimension parameter value of E0 and smart function range dimension parameter value of S2 or S3; refined scenarios when different functions are excluded in the supplementary boundary condition dimension parameter value of E1; typical conflict scenarios such as supplementary boundary condition dimension parameter value of E5. The sampling quantity of the second basic scenario accounts for the second preset quantity of the total number of basic scenarios, for example, it can be 20% to 25%.
[0074] The third level is the lowest priority, selecting extreme superimposed scenarios with extremely low probability of occurrence but valuable for testing the model's generalization ability, as the third basic scenario. For example, a scenario with multiple members and multiple boundary conditions, such as a 4R four-bedroom, two-living room apartment in the apartment type dimension, a flagship density in the smart function density dimension, and M3 or M4 in the family member dimension. The boundary condition dimension is E3, which is a conflict scenario where users simultaneously require mutually exclusive boundary conditions. The sampling quantity of the third basic scenario accounts for the third preset quantity of the total number of basic scenarios, for example, it can be 10% to 15%.
[0075] By employing the aforementioned three-tiered priority selection, the total number of scenarios is significantly reduced from the theoretical maximum while ensuring full coverage of high-frequency core scenarios, necessary supplementation of boundary scenarios, and effective backup for extreme scenarios. To facilitate data indexing, version tracking, and subsequent processing, a unique identifier is assigned to each sampled basic scenario. In one embodiment of this invention, the identifier is formatted as: "Scene Number - Apartment Type Dimension Parameter Value - Dialogue Level Dimension Parameter Value - Smart Function Range Dimension Parameter Value - Family Member Dimension Parameter Value - Boundary Condition Dimension Parameter Value - Smart Function Density Dimension Parameter Value". For example... For example, the identifier "003-3R-L2-S2-M1-E1-T2" indicates: the 003rd basic scenario, which is a three-bedroom, one-living room apartment. The parameter value for the dialogue level dimension is L2 (simplified dialogue), the parameter value for the intelligent function range dimension is S2 (single-room intelligence), the parameter value for the intelligent function range dimension is M1 (including elderly members), the parameter value for the boundary condition dimension is E1 (excluding or deleting specific intelligent functions), and the parameter value for the intelligent function density dimension is T2 (flagship density). This identifier itself carries information about all the descriptive dimensions of the scenario, which is convenient for subsequent programmatic processing and quality sampling.
[0076] Step 1023: Generate one original dialogue data for each basic scenario, and generate multiple variant dialogue data based on the original dialogue data; both the original dialogue data and the variant dialogue data include at least one input field and at least one output field;
[0077] In one embodiment of the present invention, step 1023 includes:
[0078] Step 10231: Construct the input fields of the original dialogue data. The input fields include multi-turn dialogue text, which includes at least apartment type confirmation information, family member confirmation information, and special needs expression information. The apartment type confirmation information is an apartment type description text corresponding to the apartment type dimension of the basic scenario. The family member confirmation information is a family member description text corresponding to the family member dimension of the basic scenario. The special needs expression information is a special needs description text corresponding to the boundary condition dimension of the basic scenario. The interaction rounds of the multi-turn dialogue text and the way the apartment type confirmation information, family member confirmation information, and special needs expression information are provided are determined by the parameter values of the dialogue level dimension of the basic scenario.
[0079] Step 10232: Construct the output fields of the original dialogue data. The output fields include multiple room configuration schemes. Each room configuration scheme includes at least a room name, member tags, and a smart function configuration list. The room name is determined by the parameter values of the apartment type dimension of the basic scene. The member tags are determined by the parameter values of the family member dimension of the basic scene. The smart function configuration list includes multiple smart function items, and the number of smart function items is determined by the parameter values of the smart function density dimension of the basic scene. Multiple room configuration schemes constitute the design scheme.
[0080] Specifically, after sampling the basic scenarios, a raw dialogue data is generated for each basic scenario. Each raw dialogue data includes at least one input field and at least one output field. The input field includes multi-turn dialogue text to simulate the multi-turn interaction process between the user and the agent. The multi-turn dialogue text includes at least apartment type confirmation information, family member confirmation information, and special needs expression information. The apartment type confirmation information is the apartment type description text corresponding to the apartment type dimension of the basic scenario. For example, in the scenario of a three-bedroom, one-living room apartment, the multi-turn dialogue text includes expressions such as "My apartment has three bedrooms and one living room." The family member confirmation information is the family member description text corresponding to the family member dimension of the basic scenario. For example, in the scenario M2, which includes child members, the multi-turn dialogue text includes "..." The phrase "there are children at home" and other similar expressions refer to specific needs described in text corresponding to the boundary conditions of the basic scenario. For example, in scenario E1, the multi-turn dialogue text includes exclusionary statements such as "no security needed," while scenario E4 includes partial room design statements such as "only the living room and master bedroom." The interaction rounds of the multi-turn dialogue text, as well as the methods for providing information on the apartment layout, family members, and specific needs, are determined by the parameter values of the dialogue hierarchy dimension of the basic scenario. For example, in L1 standard dialogue, the interaction rounds are moderate, and the user provides information item by item; in L2 simplified dialogue, the user provides all information in one round; in L3 detailed dialogue, the system needs to ask follow-up questions multiple times to complete the information; in L4 silent dialogue, the user responds with ambiguity or silence throughout, and the system triggers the default value rule.
[0081] The output fields include multiple room configuration schemes, output line by line for each room. Each room configuration scheme includes at least a room name, member tags, and a smart function configuration list. In one embodiment of the invention, the format of the room configuration scheme is: "Room Name [Member Tags]: Smart Function Configuration List". The room name is determined by the parameter values of the apartment type dimension of the basic scenario. For example, a three-bedroom apartment corresponds to a living room, master bedroom, secondary bedroom one, secondary bedroom two, kitchen, and bathroom. The member tags are dynamically assigned by the parameter values of the family member dimension of the basic scenario. For example, in scenario M1, the bedroom where the elderly frequently reside is labeled "elderly"; in scenario M2, the children's room is labeled "children"; and in scenario M3, the elderly's room and... Children's rooms are labeled with corresponding tags, and the smart function configuration list includes multiple smart function items. Each smart function item corresponds to a type of smart home that can be triggered. The number of smart function items is determined by the parameter value of the smart function density dimension of the basic scenario. The T1 basic density configures approximately seven to eleven smart function items per room, while the T2 flagship density configures approximately ten to fifteen smart function items per room, plus value-added scenarios such as network, security, and fresh air. At the same time, the output fields are constrained according to the boundary condition dimension. For example, in E1, functions excluded by the user (such as "security") are deleted from the output field accordingly; in E4, only rooms designed by the user are retained; and in E5, reasonable suggestions are given for incompatible scenarios.
[0082] In one embodiment of the present invention, the family member confirmation information in the input field maintains a linked correspondence with the member tags and member-specific smart function items in the output field. Specifically, when the parameter value of the family member dimension indicates the presence of an elderly member, the input field includes an expression of age-friendly needs, such as difficulty getting up at night or needing simple operation, and the output field adds age-friendly smart function items such as night-time sensor, emergency call, and age-friendly lighting to the corresponding bedroom; when the parameter value of the family member dimension indicates the presence of a child member, the input field includes a statement on child safety concerns, and the output field adds child-related smart function items such as network control, safe electricity use, and timed sunshade to the corresponding room; when the parameter value of the family member dimension indicates the presence of both elderly and child members, the output field adds age-friendly smart function items and child safety smart function items to the corresponding room respectively; when the parameter value of the family member dimension indicates the presence of a domestic service worker, the input field includes a statement on permission management needs, and the output field adds smart function items such as permission management and entry / exit logs to the corresponding room.
[0083] In one embodiment of the present invention, when the parameter value of the boundary condition dimension of the basic scene causes the first room configuration scheme corresponding to the first member label in the output field to be excluded, the first member label is transferred to a room configuration scheme of the same type as the first room configuration scheme. For example, if the user only requires the design of the living room and the master bedroom, and the secondary bedroom is excluded, and if the elderly label was originally marked on the secondary bedroom, the elderly label is transferred to the master bedroom.
[0084] In one embodiment of the present invention, when there are multiple room configuration schemes of the same type as the first room configuration scheme, the member tag is transferred to the room configuration scheme with the highest priority according to a preset priority order. For example, the preset priority of the bedroom is master bedroom, which is higher than secondary bedroom one, which is higher than secondary bedroom two. When secondary bedroom one is excluded and master bedroom is retained, the elderly tag is transferred to master bedroom. When master bedroom is also excluded and secondary bedroom two is retained, the elderly tag is transferred to secondary bedroom two. Through the above tag transfer mechanism, it is ensured that no member tag without a corresponding room configuration scheme appears in the output field, thus ensuring the logical consistency of the design scheme output.
[0085] Step 10233: Adjust the original dialogue data based on the preset variant strategy to obtain multiple variant dialogue data; the variant strategy is used to change the expression style of the input fields in the original dialogue data, and / or to change the room configuration scheme of the output fields in the original dialogue data.
[0086] Specifically, after configuring the original dialogue data, the content of the input fields and / or output fields of the original dialogue data is changed through a preset variant strategy to obtain multiple variant dialogue data. This makes it possible for the original dialogue data and each variant dialogue data to have identifiable differences in input expression style or output scheme content under the same basic scenario. This enables the model to learn the possibility of multiple reasonable scheme configurations under the same needs in subsequent training, which can significantly improve the robustness of the model to different user expression styles, different needs emphasis, and different scheme preferences.
[0087] In one embodiment of the present invention, the multiple variant strategies include:
[0088] The standard variant strategy adjusts the user expression in the input field of the original dialogue data to a standard tone, assigns member rooms according to the default rules, and adjusts the value of the intelligent function density dimension parameter to the T1 base density.
[0089] The colloquial variation strategy adjusts the user expressions in the input fields of the original dialogue data to a more colloquial style, adjusts the member room allocation of the original dialogue data, and focuses on the configuration of network-related intelligent function items.
[0090] The technical variant strategy adjusts the user expressions in the input fields of the original dialogue data to a more technical style that includes technical terms, adjusts the member room allocation of the original dialogue data, and focuses on the configuration of security-related intelligent functions.
[0091] The density enhancement strategy maintains the input field expression of the original dialogue data unchanged, enhances the intelligent function density of the original dialogue data to a level higher than the basic level, and adds corresponding intelligent scene configurations. For example, while maintaining the tone of the original dialogue data, the intelligent scene density dimension is enhanced from T1 basic density to T2 flagship density, and value-added intelligent functions such as security, network, and fresh air are added to functional rooms such as balconies and master bedrooms.
[0092] In some other embodiments of the present invention, variant strategies can be added according to actual needs, such as a full configuration variant strategy, which adjusts the output fields with whole-house smart configuration, adding security to the balcony, adding network to the master bedroom, etc., and a low-frequency supplementary variant strategy, which is only generated for the third-level low-frequency scenario, and uses technical language to superimpose T2 flagship density to further test the performance of the intelligent agent in extreme scenarios. This embodiment does not impose any limitations here.
[0093] In one embodiment of the present invention, to avoid identical output fields of different variant dialogue data, this step introduces a difference control mechanism to control the output fields of the variant dialogue data, so that there are identifiable output differences among the variant dialogue data. Specifically, the differences are achieved by adjusting at least one of the following without changing the core intelligent function configuration logic: the coverage of intelligent function items in each functional room, that is, the emphasis of the output fields of different variant dialogue data on the intelligent function items is different, so that the coverage of intelligent function items such as audio-visual, fresh air, and security is different in each room; the combination method of member tags and rooms, that is, for families with multiple types of members, different The output fields of the variant dialogue data adjust the emphasis order of the needs of the elderly and children to make the combination of smart function items in the corresponding room configuration schemes different; the granularity of non-core rooms, that is, under the same house type, the output fields of different variant dialogue data differ in the number and specific configuration of smart function items in non-core functional rooms such as kitchen, bathroom, and porch; the handling of smart function item retention or exclusion under the same boundary condition, that is, for the same E1 or E5 boundary condition, different variant dialogue data have different perspectives on the understanding of user intent, resulting in different but reasonable processing results for whether a certain smart function item should be retained in a certain room in the output fields of different variant dialogue data.
[0094] Step 1024: Construct the dialogue dataset based on the original dialogue data and the variant dialogue data.
[0095] Specifically, all original dialogue data and variant dialogue data are merged to generate the final dialogue dataset. In one embodiment of the present invention, the dialogue dataset is stored in JSON format, and each record includes a scene identifier, variant number, input field, and output field. The total size of the dialogue dataset is approximately 2,300 to 2,550 records, and the distribution statistics of each descriptive dimension are recorded in the metadata file.
[0096] Step 103: Train the large language model based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on the user's interactive text;
[0097] Specifically, after constructing the dialogue dataset, a large language model is trained based on the dialogue dataset to obtain a designed intelligent agent. In one embodiment of the present invention, the dialogue dataset is divided into a training set and a validation set according to a preset ratio, for example, the training set accounts for 90% and the validation set accounts for 10%. A general large language model is selected as the base model (e.g., the Qwen series model), and a parameter-efficient fine-tuning method (e.g., LoRA) is used to train the large language model based on the training set, and the training effect is verified on the validation set. The training objective is to minimize the difference loss between the model output scheme and the standard output field in the dialogue dataset. Since the training data has achieved multi-dimensional scene coverage, complete boundary conditions, and rich variant differences through the aforementioned steps, the fine-tuned model can learn the complex business logic of smart home whole-house design.
[0098] Step 104: Construct a multi-agent system, which includes an interactive agent and a design agent; the interactive agent is used to interact with the user, obtain key design information corresponding to the parameter values of multiple description dimensions, and transmit the key design information to the design agent, which is used to generate a design scheme based on the key design information.
[0099] Specifically, the interactive agent can be implemented by combining prompt word engineering with a general large model. It is used to interact with users and collect key design information corresponding to parameter values of multiple descriptive dimensions, such as apartment type, family members, functional preferences, and boundary conditions, within no more than a preset number of interactions. When the user does not provide complete key design information, the interactive agent uses existing interaction content to ask the user follow-up questions to complete the missing key design information and transmits the key design information to the design agent. After receiving the structured key design information output by the interactive agent, the design agent generates a design scheme containing multiple room configuration options based on the business logic learned in the above embodiments. Compared with using a single general large model, the multi-agent system can separate the intent recognition and scheme design tasks. The fine-tuned model is dedicated to the scheme design task, with clear task boundaries and single responsibilities. This fundamentally reduces the risk of long context illusion caused by task coupling, and significantly improves the stability and reliability of scheme generation.
[0100] In one embodiment of the present invention, the multi-agent system further includes: a product matching agent;
[0101] The product matching agent receives the design scheme output by the design agent, matches specific products from the product database according to the design scheme, and outputs the final scheme containing specific product recommendations.
[0102] Specifically, after receiving the design scheme output by the design agent, the product matching agent can accurately match the corresponding equipment from the product database based on the intelligent function items, member tags and other multi-dimensional attributes in the configuration scheme of each room in the design scheme, generate a final scheme containing specific product recommendations and price references and return it to the user terminal for display.
[0103] In one embodiment of the present invention, the multi-agent system can further perform dynamic intent recognition, scheme comparison and iterative optimization based on user feedback on the scheme, so as to realize the continuous optimization of the design scheme in multiple rounds of interaction.
[0104] This embodiment also provides a smart home scene design intelligent agent construction device, such as... Figure 4 As shown, it includes:
[0105] The dimension construction module 201 is used to construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values, and each combination of parameter values of the multiple descriptive dimensions corresponds to a smart home scene.
[0106] The dataset construction module 202 is used to construct a dialogue dataset, which includes multiple dialogue data, each dialogue data corresponding to a smart home scene. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scene and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scene.
[0107] Training module 203 is used to train a large language model based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on user interaction text.
[0108] The multi-agent construction module 204 is used to construct a multi-agent system, which includes an interactive agent and a design agent. The interactive agent is used to interact with the user, obtain key design information corresponding to the parameter values of multiple description dimensions, and transmit the key design information to the design agent. The design agent is used to generate a design scheme based on the key design information.
[0109] In one embodiment of the present invention, the multi-agent system constructed by the multi-agent construction module 204 further includes a product matching agent, which is used to receive the design scheme output by the design agent, match specific products from the product database according to the design scheme, and output a final scheme containing specific product recommendations.
[0110] In one embodiment of the present invention, the dataset construction module 202 includes:
[0111] The theoretical full-scale calculation unit is used to calculate all parameter value combinations of multiple description dimensions to obtain the theoretical full-scale scene set, which includes all smart home scenes;
[0112] The basic scene selection unit is used to select a preset number of basic scenes from the theoretical full scene set;
[0113] The dialogue data generation unit is used to generate one original dialogue data for each basic scenario, and to generate multiple variant dialogue data based on the original dialogue data; both the original dialogue data and the variant dialogue data include at least one input field and at least one output field.
[0114] A dataset construction unit is used to construct the dialogue dataset based on the original dialogue data and the variant dialogue data.
[0115] In one embodiment of the present invention, the dialogue data generation unit includes:
[0116] An input field construction subunit is used to construct the input fields of the original dialogue data. The input fields include multi-turn dialogue text, which includes at least apartment type confirmation information, family member confirmation information, and special needs expression information. The apartment type confirmation information is an apartment type description text corresponding to the apartment type dimension of the basic scenario. The family member confirmation information is a family member description text corresponding to the family member dimension of the basic scenario. The special needs expression information is a special needs description text corresponding to the boundary condition dimension of the basic scenario. The interaction rounds of the multi-turn dialogue text and the way the apartment type confirmation information, family member confirmation information, and special needs expression information are provided are determined by the parameter values of the dialogue hierarchy dimension of the basic scenario.
[0117] The output field construction subunit is used to construct the output fields of the original dialogue data. The output fields include multiple room configuration schemes, each of which includes at least a room name, member tags, and a smart function configuration list. The room name is determined by the parameter values of the apartment type dimension of the basic scene. The member tags are determined by the parameter values of the family member dimension of the basic scene. The smart function configuration list includes multiple smart function items, the number of which is determined by the parameter values of the smart function density dimension of the basic scene. Multiple room configuration schemes constitute the design scheme.
[0118] The variant subunit is used to adjust the original dialogue data based on the preset variant strategy to obtain multiple variant dialogue data; the variant strategy is used to change the expression style of the input fields in the original dialogue data, and / or the room configuration scheme of the output fields, and to perform secondary adjustments on the output fields of the variant dialogue data through a preset difference control mechanism.
[0119] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the smart home scene design intelligent agent construction method in the above embodiment.
[0120] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the steps in the intelligent agent construction method for smart home scene design in the above embodiment.
[0121] This embodiment also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the steps in the smart home scene design intelligent agent construction method in the above embodiment.
[0122] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for constructing intelligent agents in smart home scene design, characterized in that, include: Construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values. Each combination of parameter values for multiple descriptive dimensions corresponds to a smart home scene. Construct a dialogue dataset, which includes multiple dialogue data, each corresponding to a smart home scenario. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scenario and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scenario. The large language model is trained based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on the user's interactive text. A multi-agent system is constructed, comprising an interactive agent and a design agent; the interactive agent interacts with the user to obtain key design information corresponding to parameter values of multiple descriptive dimensions, and transmits the key design information to the design agent, which generates a design scheme based on the key design information.
2. The method for constructing intelligent agents in smart home scene design according to claim 1, characterized in that, The multiple independent descriptive dimensions include at least two of the following: The floor plan dimension reflects the room and living room structure of the house, and its multiple parameter values represent different room and living room structure types; The dialogue hierarchy dimension reflects the completeness of information provided during user interaction, and its multiple parameters take different levels of completeness. The intelligent function scope dimension reflects the scope of a user's intelligent function needs, and its multiple parameter values take different scope types. The family member dimension reflects the composition of a household, and its multiple parameters can take different member composition types. The boundary condition dimension is used to reflect the special constraints in user requirements, and its multiple parameters can take different special constraint types. The smart function density dimension reflects the richness of smart functions in a smart home scenario, and its parameter values take different levels of richness.
3. The method for constructing intelligent agents in smart home scene design according to claim 2, characterized in that, The construction of the dialogue dataset includes: Calculate all parameter values for multiple description dimensions to obtain the theoretical full set of scenarios, which includes all smart home scenarios; Select a preset number of basic scenarios from the theoretical full set of scenarios; For each basic scenario, generate one original dialogue data, and generate multiple variant dialogue data based on the original dialogue data; both the original dialogue data and the variant dialogue data include at least one input field and at least one output field; The dialogue dataset is constructed based on the original dialogue data and the variant dialogue data.
4. The method for constructing intelligent agents in smart home scene design according to claim 3, characterized in that, A predetermined number of basic scenarios are selected from the theoretical full set of scenarios, including: Select a first preset number of first basic scenarios from the theoretical full set of scenarios, wherein the first basic scenarios are smart home scenarios with the highest business priority; Select a second preset number of second basic scenarios from the theoretical full set of scenarios. The second basic scenarios are smart home scenarios with medium business priority. Select a third preset number of third basic scenarios from the theoretical full set of scenarios. The third basic scenarios are smart home scenarios with the lowest business priority. Among them, the first preset quantity is greater than the second preset quantity, and the second preset quantity is greater than the third preset quantity.
5. The method for constructing intelligent agents in smart home scene design according to claim 3, characterized in that, The process involves generating one original dialogue data for each basic scenario, and generating multiple variant dialogue data based on the original dialogue data, including: The input fields for constructing the original dialogue data include multi-turn dialogue text, which at least includes apartment type confirmation information, family member confirmation information, and special needs expression information. The apartment type confirmation information is an apartment type description text corresponding to the apartment type dimension of the basic scenario; the family member confirmation information is a family member description text corresponding to the family member dimension of the basic scenario; and the special needs expression information is a special needs description text corresponding to the boundary condition dimension of the basic scenario. The interaction rounds of the multi-turn dialogue text and the methods of providing the apartment type confirmation information, family member confirmation information, and special needs expression information are determined by the parameter values of the dialogue hierarchy dimension of the basic scenario. The output fields of the original dialogue data are constructed, and the output fields include multiple room configuration schemes. Each room configuration scheme includes at least a room name, member tags, and a smart function configuration list. The room name is determined by the parameter value of the apartment type dimension of the basic scene. The member tags are determined by the parameter value of the family member dimension of the basic scene. The smart function configuration list includes multiple smart function items, and the number of smart function items is determined by the parameter value of the smart function density dimension of the basic scene. Multiple room configuration schemes constitute the design scheme. The original dialogue data is adjusted based on the preset variant strategy to obtain multiple variant dialogue data; the variant strategy is used to change the expression style of the input fields in the original dialogue data, and / or to change the room configuration scheme of the output fields in the original dialogue data.
6. The method for constructing intelligent agents in smart home scene design according to claim 5, characterized in that, When the parameter value of the boundary condition dimension of the basic scenario causes the first room configuration scheme corresponding to the first member label in the output field to be excluded, the first member label is transferred to a room configuration scheme of the same type as the first room configuration scheme.
7. The method for constructing intelligent agents in smart home scene design according to claim 6, characterized in that, When there are multiple room configuration schemes of the same type as the first room configuration scheme, the member tag is transferred to the room configuration scheme with the highest priority according to the preset priority order.
8. The method for constructing intelligent agents in smart home scene design according to claim 1, characterized in that, The multi-agent system also includes: a product matching agent; The product matching agent receives the design scheme output by the design agent, matches specific products from the product database according to the design scheme, and outputs the final scheme containing specific product recommendations.
9. A smart home scene design intelligent agent construction device, characterized in that, include: The dimension construction module is used to construct multiple independent descriptive dimensions of the home environment. Each descriptive dimension includes multiple parameter values, and each combination of parameter values of the multiple descriptive dimensions corresponds to a smart home scene. The dataset construction module is used to construct a dialogue dataset, which includes multiple dialogue data, each corresponding to a smart home scene. The dialogue data includes at least one input field and at least one output field. The input field is constructed based on the parameter values of multiple descriptive dimensions of the smart home scene and is used to simulate the interaction text between the user and the smart agent. The output field is used to describe the design scheme corresponding to the smart home scene. The training module is used to train the large language model based on the dialogue dataset to obtain a design agent that can generate design schemes for corresponding smart home scenarios based on the user's interactive text. A multi-agent construction module is used to construct a multi-agent system, which includes an interactive agent and a design agent. The interactive agent interacts with the user to obtain key design information corresponding to parameter values of multiple description dimensions, and transmits the key design information to the design agent. The design agent generates a design scheme based on the key design information.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent agent construction method for intelligent home scene design as described in any one of claims 1 to 8.
11. A computer-readable storage medium for storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps in the intelligent agent construction method for intelligent home scene design as described in any one of claims 1 to 8.
12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps in the smart home scene design intelligent agent construction method as described in any one of claims 1 to 8.
