Dual-model learning-based full-active smart home scene adaptation method

Abstract

The application discloses a full active intelligent home scene adaptation method based on double model learning, comprising the following steps: S1: system initialization configuration: guiding the user to complete a simple preference questionnaire and generating an initial behavior model; configuring a family portrait and obtaining family structure information, and loading a corresponding personalized care rule set. The application innovates a full active adaptation system of "initial configuration-data collection-model learning-scene generation-optimization feedback" through a double model collaborative learning mechanism, accurately captures stable habits and temporary behavior changes of users, and avoids scene rule lag or failure; through privacy hierarchical processing and multi-source data cross verification, the safety of user data is ensured while reducing the false alarm rate of abnormal early warning; the special population care logic and special event adaptation mechanism are deeply integrated, realizing the leap from "passive response instruction" to "active demand prediction", improving the adaptation accuracy and user experience of the intelligent home, and having strong practicality and promotion value.

Description

Technical Field

[0001] This invention relates to the field of smart home technology, and in particular to a fully proactive smart home scene adaptation method based on dual-model learning. Background Technology

[0002] As the wave of artificial intelligence technology sweeps the globe, a new round of technological revolution and industrial transformation is advancing in depth. Smart homes, as the core scenario in the residential field of "food, clothing, housing, and transportation," are undergoing a profound transformation from single-point intelligence to proactive services across all scenarios. Currently, most mainstream smart home scenario adaptation solutions adopt an "AI-assisted scenario customization mode," which is essentially still a semi-proactive form of "user guidance and system assistance," and has the following limitations in practical applications: 1. Passive and fragmented intelligent experience: The generation, switching and optimization of scenes rely heavily on users to manually input or confirm commands through devices such as the central control screen. The system cannot autonomously complete decision-making and closed-loop execution based on environmental perception and user behavior. Users need to continuously intervene and operate, resulting in poor experience continuity.

[0003] 2. Insufficient dynamic adaptation capability: The use of a single static learning model makes it difficult to effectively cope with dynamic changes in user behavior, such as unexpected situations like temporary business trips, holiday schedule adjustments, or the arrival of visitors. This can easily lead to the lag or even failure of scene rules, making it impossible to accurately match real-time needs.

[0004] 3. Shortcomings in functional coverage: There is a general lack of detailed care scenario design for special groups such as the elderly and children, and no targeted personalized abnormal behavior early warning and device linkage mechanism has been established; In summary, this application proposes a fully proactive smart home scene adaptation method based on dual-model learning. Summary of the Invention

[0005] Based on the technical problems existing in the background technology, this invention proposes a fully proactive smart home scene adaptation method based on dual-model learning.

[0006] The fully proactive smart home scene adaptation method based on dual-model learning proposed in this invention includes the following steps: S1: System initialization configuration: Guide users to complete a simple preference questionnaire to generate an initial behavior model; configure family profiles and obtain family structure information, and load the corresponding personalized care rule set; S2: Multimodal data synchronous acquisition: Through multiple types of collectors in the perception layer, data is collected in real time at a preset frequency, including user location, actions, physiological data, environmental parameters, device operating status, etc. A unified timestamp is added to all collected data to achieve spatiotemporal alignment; S3: Data Preprocessing and Hierarchical Management: Initial screening of collected data to remove invalid data; differentiated processing according to privacy level, sensitive data is anonymized and encrypted locally, and non-sensitive data can be selectively uploaded to the cloud after cleaning and standardization, generating standardized data packages containing timestamps, data sources, data types, and data values; S4: AI Dual-Branch Model Collaborative Learning: Based on a dual-branch learning architecture consisting of a short-term habit model and a long-term habit model, it captures temporary changes in user behavior and stable behavior patterns respectively; based on the AI ​​algorithm layer, it judges the deviation between new data and the long-term model in real time. When the deviation is significant, the short-term model is activated and given high weight, and the scenario decision is output through weighted calculation. S5: Scene generation, classification and execution: The AI ​​engine automatically generates scene instruction sets based on dual-model status and real-time environmental data; the scenes are divided into daily scenes, care scenes and emergency scenes, and executed according to priority; after the application layer device executes the instructions, the execution results are fed back to the data layer to form a closed loop; S6: Real-time optimization with user fine-tuning: Capture user manual intervention as reinforcement learning signals, collect operation context information synchronously, adjust model parameters in real time through incremental learning algorithms, assign higher weights to manually fine-tuned data, and achieve rapid model optimization. S7: Anomaly and Care Scenarios Judgment and Early Warning: Based on the user's long-term habit model, personalized anomaly thresholds are dynamically generated, and multi-source data cross-validation is used for key early warning scenarios; the anomalies are divided into three levels: prompt, reminder, and alarm, and corresponding early warning and linkage mechanisms are activated. S8: Special Event Adaptation and Recovery: The system detects special events of visitors through the device and automatically activates the corresponding scene mode; after the special event ends, the system automatically restores the normal scene to ensure that the personalized experience is not disturbed.

[0007] Preferably, in S1, the initial behavior model is constructed based on questionnaire information filled in by the user regarding their daily routine, temperature preferences, and device usage habits, in order to solve the adaptation problem in the AI ​​cold start phase; the personalized care rule set includes, but is not limited to, targeted function configurations for elderly fall detection, child locks on children's devices, and activity monitoring for special groups.

[0008] Preferably, in S2, the sensing layer multi-type collectors include millimeter-wave radar, environmental sensors, device status collectors, and optional wearable devices; during the data acquisition process, preliminary verification is performed simultaneously to detect whether the sensors are offline and whether the data is within a reasonable physical range. Invalid data is marked and temporarily stored in the cache area and does not enter the main processing flow.

[0009] Preferably, in S3, sensitive data includes human body contour trajectory and physiological data, which are encrypted and stored using the AES algorithm, and the data does not leave the local central control gateway; non-sensitive data includes environmental temperature and humidity, illuminance, equipment energy consumption, etc., which are standardized and then used by the AI ​​algorithm layer.

[0010] Preferably, in S4, the long-term habit model focuses on learning stable and periodic user behavior patterns, with a long update cycle and stable weights; the short-term habit model focuses on capturing temporary and sudden behavioral changes, with a fast learning rate and a fast decay rate; the dual-model collaborative logic includes feature association and dynamic optimization, constructing a "behavior-state" causal relationship chain through timestamp alignment, and when the confidence of a certain branch is lower than the threshold, the context information of the other branch is called to assist in correction.

[0011] Preferably, in S5, emergency scenarios have the highest interruption priority, including fall detection, gas leak alarm, and long-term static warning; daily scenarios include routine scenarios such as getting up, sleeping, leaving home, and returning home; and care scenarios include nighttime lighting for the elderly and adjustment of children's sleep environment.

[0012] Preferably, in step S6, user manual intervention operations include, but are not limited to, temporarily adjusting device parameters via the APP, directly shutting down the device, and modifying scene configuration; the system records the operation itself and the corresponding time, environmental parameters, and user location context information, and achieves real-time adjustment of model parameters through incremental learning to ensure that the model quickly responds to the user's explicit intent.

[0013] Preferably, in step S7, the personalized anomaly threshold is dynamically calculated based on the user's historical behavior data to adapt to the behavioral characteristics of different users; multi-source data cross-validation is used to reduce the false alarm rate, such as the "fall" judgment needs to simultaneously satisfy the falling posture detected by millimeter-wave radar and the impact acceleration detected by wearable device; in the three-level warning, the prompt only records the log, the reminder pushes the notification to the user's APP, and the alarm, in addition to the APP push, also links the sound and light alarm and notifies the emergency contact.

[0014] Preferably, in S8, special event detection includes, but is not limited to, detection via doorbell or microphone devices. For example, if a doorbell is triggered and multiple people are speaking, it is determined to be a visitor event. The corresponding scene mode includes, but is not limited to, adjusting the lighting atmosphere, pausing privacy-related automated functions, and automatically restoring the normal scene after the visitor leaves.

[0015] Compared with existing technologies, the beneficial effects of this invention are: 1. By quickly building a user behavior baseline model through an optional initial preference questionnaire, the "cold start" problem in the early stages of AI systems is effectively solved, significantly shortening the adaptation cycle, avoiding early scenarios that deviate severely from user habits, and achieving a high-starting-point "fully proactive" experience without requiring users to manually set scenarios. 2. It adopts a dual-model incremental learning algorithm of "short-term + long-term" to grasp stable habits and respond quickly to temporary behavioral changes. Its outstanding advantage is that it can intelligently distinguish and handle special events, such as family gatherings and temporary visitors. The system can automatically or switch to the corresponding mode with one click and seamlessly restore the daily scene after the event ends, ensuring intelligent flexibility and avoiding rigidity. It has the advantages of intelligent dynamic adaptation and flexible event response. 3. Breaking through the linkage of basic equipment, it provides personalized proactive care scenarios for special groups such as the elderly and children, such as fall detection and abnormal activity warning, and integrates security and privacy protection into the data hierarchical processing architecture, which has the advantages of refined care and proactive safety protection; 4. The solution is lightweight and compatible with users' existing smart devices, reducing upgrade costs and deployment barriers, and has the advantages of high compatibility and ease of implementation; This invention innovates a fully proactive adaptation system through a dual-model collaborative learning mechanism, encompassing "initial configuration - data collection - model learning - scene generation - optimization feedback." This system can accurately capture stable user habits and temporary behavioral changes, avoiding lag or failure of scene rules. By employing privacy-graded processing and multi-source data cross-validation, it reduces the false alarm rate of anomaly warnings while ensuring user data security. Furthermore, by deeply integrating special population care logic and special event adaptation mechanisms, it achieves a leap from "passively responding to instructions" to "proactively anticipating needs," significantly improving the adaptation accuracy and user experience of smart homes. This invention possesses strong practicality and promotional value. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the fully proactive smart home scene adaptation method based on dual-model learning proposed in this invention; Figure 2 This is a flowchart illustrating the workflow of the fully proactive smart home scene adaptation method based on dual-model learning proposed in this invention. Figure 3 This is a flowchart illustrating the short-term and long-term dual-model graph combination of the fully proactive smart home scene adaptation method based on dual-model learning proposed in this invention. Detailed Implementation

[0017] The present invention will be further explained below with reference to specific embodiments. Example

[0018] Reference Figure 1-3 This embodiment proposes a fully proactive smart home scene adaptation method based on dual-model learning, including the following steps: S1: System initialization configuration: Guide users to complete a simple preference questionnaire to generate an initial behavior model; configure family profiles and obtain family structure information, and load the corresponding personalized care rule set; The initial behavior model is built based on questionnaire information filled in by users regarding their daily routines, temperature preferences, and device usage habits, and is used to solve the adaptation problem in the AI ​​cold start phase; the personalized care rule set includes, but is not limited to, targeted function configurations for elderly fall detection, child locks on children's devices, and activity monitoring for special groups; S2: Multimodal data synchronous acquisition: Through multiple types of collectors in the perception layer, data is collected in real time at a preset frequency, including user location, actions, physiological data, environmental parameters, device operating status, etc. A unified timestamp is added to all collected data to achieve spatiotemporal alignment; The sensing layer includes various types of data acquisition devices, such as millimeter-wave radar, environmental sensors, device status acquisition devices, and optional wearable devices. During the data acquisition process, preliminary verification is performed simultaneously to check whether the sensors are offline and whether the data is within a reasonable physical range. Invalid data is marked and temporarily stored in the cache and does not enter the main processing flow. S3: Data Preprocessing and Hierarchical Management: Initial screening of collected data to remove invalid data; differentiated processing according to privacy level, sensitive data is anonymized and encrypted locally, and non-sensitive data can be selectively uploaded to the cloud after cleaning and standardization, generating standardized data packages containing timestamps, data sources, data types, and data values; Sensitive data, including human body contour trajectory and physiological data, is encrypted and stored using the AES algorithm, and the data does not leave the local central control gateway; non-sensitive data, including environmental temperature and humidity, light intensity, equipment energy consumption, etc., are standardized and then used by the AI ​​algorithm layer. S4: AI Dual-Branch Model Collaborative Learning: Based on a dual-branch learning architecture consisting of a short-term habit model and a long-term habit model, it captures temporary changes in user behavior and stable behavior patterns respectively; based on the AI ​​algorithm layer, it judges the deviation between new data and the long-term model in real time. When the deviation is significant, the short-term model is activated and given high weight, and the scenario decision is output through weighted calculation. The long-term habit model focuses on learning stable and periodic user behavior patterns, with a long update cycle and stable weights; the short-term habit model focuses on capturing temporary and sudden behavioral changes, with a fast learning rate and a fast decay rate; the dual-model collaborative logic includes feature association and dynamic optimization, constructing a "behavior-state" causal relationship chain through timestamp alignment, and calling the context information of the other branch to assist in correction when the confidence of a certain branch is lower than the threshold. S5: Scene generation, classification and execution: The AI ​​engine automatically generates scene instruction sets based on dual-model status and real-time environmental data; the scenes are divided into daily scenes, care scenes and emergency scenes, and executed according to priority; after the application layer device executes the instructions, the execution results are fed back to the data layer to form a closed loop; Emergency scenarios have the highest interruption priority, including fall detection, gas leak alarm, and long-term static warning; daily scenarios include routine scenarios such as waking up, sleeping, leaving home, and returning home; and care scenarios include nighttime lighting for the elderly and adjusting the sleep environment for children. S6: Real-time optimization with user fine-tuning: Capture user manual intervention as reinforcement learning signals, collect operation context information synchronously, adjust model parameters in real time through incremental learning algorithms, assign higher weights to manually fine-tuned data, and achieve rapid model optimization. User manual intervention includes, but is not limited to, temporarily adjusting device parameters via the app, directly shutting down the device, and modifying scene configurations. The system records the operation itself, along with the corresponding time, environmental parameters, and user location context information. Through incremental learning, the system enables real-time adjustment of model parameters, ensuring that the model responds quickly to the user's explicit intent. S7: Anomaly and Care Scenarios Judgment and Early Warning: Based on the user's long-term habit model, personalized anomaly thresholds are dynamically generated, and multi-source data cross-validation is used for key early warning scenarios; the anomalies are divided into three levels: prompt, reminder, and alarm, and corresponding early warning and linkage mechanisms are activated. The personalized anomaly threshold is dynamically calculated based on the user's historical behavior data to adapt to the behavioral characteristics of different users; multi-source data cross-validation is used to reduce the false alarm rate. For example, the judgment of "fall" needs to simultaneously meet the falling posture detected by millimeter-wave radar and the impact acceleration detected by wearable device; in the three-level warning, the prompt only records the log, the reminder pushes the notification to the user's APP, and the alarm, in addition to the APP push, also links the sound and light alarm and notifies the emergency contact. S8: Special Event Adaptation and Recovery: The system detects special events of visitors through the device and automatically activates the corresponding scene mode; after the special event ends, the system automatically restores the normal scene to ensure that the personalized experience is not disturbed. Special event detection includes, but is not limited to, detection via doorbells and microphones. For example, if a doorbell is triggered and multiple people are talking, it is determined to be a visitor event. Corresponding scene modes include, but are not limited to, adjusting the lighting atmosphere, pausing privacy-related automated functions, and automatically restoring the normal scene after the visitor leaves.

[0019] This embodiment innovates a fully proactive adaptation system through a dual-model collaborative learning mechanism, encompassing "initial configuration - data collection - model learning - scene generation - optimization feedback." This system can accurately capture stable user habits and temporary behavioral changes, avoiding lag or failure of scene rules. By employing privacy-graded processing and multi-source data cross-validation, it reduces the false alarm rate of anomaly warnings while ensuring user data security. Furthermore, by deeply integrating special population care logic and special event adaptation mechanisms, it achieves a leap from "passively responding to instructions" to "proactively anticipating needs," significantly improving the adaptation accuracy and user experience of smart homes. This approach possesses strong practicality and promotional value.

[0020] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fully proactive smart home scene adaptation method based on dual-model learning, characterized in that, Includes the following steps: S1: System initialization configuration: Guide users to complete a simple preference questionnaire to generate an initial behavior model; configure family profiles and obtain family structure information, and load the corresponding personalized care rule set; S2: Multimodal data synchronous acquisition: Through multiple types of collectors in the perception layer, data is collected in real time at a preset frequency, including user location, action, physiological data, environmental parameters, and device operating status. A unified timestamp is added to all collected data to achieve spatiotemporal alignment. S3: Data Preprocessing and Hierarchical Management: Initial screening of collected data to remove invalid data; differentiated processing according to privacy level, sensitive data is anonymized and encrypted locally, and non-sensitive data can be selectively uploaded to the cloud after cleaning and standardization, generating standardized data packages containing timestamps, data sources, data types, and data values; S4: AI Dual-Branch Model Collaborative Learning: Based on a dual-branch learning architecture consisting of a short-term habit model and a long-term habit model, it captures temporary changes in user behavior and stable behavior patterns respectively; based on the AI ​​algorithm layer, it judges the deviation between new data and the long-term model in real time. When the deviation is significant, the short-term model is activated and given high weight, and the scenario decision is output through weighted calculation. S5: Scene generation, classification and execution: The AI ​​engine automatically generates scene instruction sets based on dual-model status and real-time environmental data; the scenes are divided into daily scenes, care scenes and emergency scenes, and executed according to priority; after the application layer device executes the instructions, the execution results are fed back to the data layer to form a closed loop; S6: Real-time optimization with user fine-tuning: Capture user manual intervention as reinforcement learning signals, collect operation context information synchronously, adjust model parameters in real time through incremental learning algorithms, assign higher weights to manually fine-tuned data, and achieve rapid model optimization. S7: Anomaly and Care Scenarios Judgment and Early Warning: Based on the user's long-term habit model, personalized anomaly thresholds are dynamically generated, and multi-source data cross-validation is used for key early warning scenarios; the anomalies are divided into three levels: prompt, reminder, and alarm, and corresponding early warning and linkage mechanisms are activated. S8: Special Event Adaptation and Recovery: The system detects special events of visitors through the device and automatically activates the corresponding scene mode; after the special event ends, the system automatically restores the normal scene to ensure that the personalized experience is not disturbed.

2. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S1, the initial behavior model is constructed based on questionnaire information filled in by the user regarding their daily routine, temperature preferences, and device usage habits, and is used to solve the adaptation problem in the AI ​​cold start phase; the personalized care rule set includes, but is not limited to, targeted function configurations for elderly fall detection, child locks on children's devices, and activity monitoring for special groups.

3. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S2, the sensing layer includes various types of data collectors, such as millimeter-wave radar, environmental sensors, device status data collectors, and optional wearable devices. During data acquisition, preliminary verification is performed simultaneously to check whether the sensor is offline and whether the data is within a reasonable physical range. Invalid data is marked and temporarily stored in the cache and does not enter the main processing flow.

4. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S3, sensitive data includes human body contour trajectory and physiological data, which are encrypted and stored using the AES algorithm and do not leave the local central control gateway; non-sensitive data includes ambient temperature and humidity, illuminance, and equipment energy consumption, which are standardized and then used by the AI ​​algorithm layer.

5. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S4, the long-term habit model focuses on learning stable and periodic user behavior patterns, with a long update cycle and stable weights. The short-term habit model focuses on capturing temporary and sudden behavioral changes, with a fast learning rate and a fast decay rate; the dual-model collaborative logic includes feature association and dynamic optimization, constructing a "behavior-state" causal relationship chain through timestamp alignment, and calling the context information of the other branch to assist in correction when the confidence of a certain branch is lower than the threshold.

6. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S5, emergency scenarios have the highest interruption priority, including fall detection, gas leak alarm, and long-term static warning; daily scenarios include routine scenarios such as getting up, sleeping, leaving home, and returning home; and care scenarios include nighttime lighting for the elderly and adjusting the sleep environment for children.

7. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S6, user manual intervention operations include, but are not limited to, temporarily adjusting device parameters via the APP, directly shutting down the device, and modifying scene configuration; the system records the operation itself and the corresponding time, environmental parameters, and user location context information, and achieves real-time adjustment of model parameters through incremental learning to ensure that the model quickly responds to the user's explicit intent.

8. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S7, the personalized anomaly threshold is dynamically calculated based on the user's historical behavior data to adapt to the behavioral characteristics of different users; multi-source data cross-validation is used to reduce the false alarm rate. For example, the judgment of "falling" needs to simultaneously satisfy the falling posture detected by millimeter-wave radar and the impact acceleration detected by wearable devices; in the three-level warning, the prompt only records the log, the reminder pushes the notification to the user's APP, and the alarm, in addition to the APP push, also links the sound and light alarm and notifies the emergency contact.

9. The fully proactive smart home scene adaptation method based on dual-model learning according to claim 1, characterized in that, In S8, special event detection includes, but is not limited to, detection through doorbell and microphone devices. For example, if a doorbell is triggered and multiple people are talking, it is determined to be a visitor event. The corresponding scene mode includes, but is not limited to, adjusting the lighting atmosphere, pausing privacy-related automated functions, and automatically restoring the normal scene after the visitor leaves.

Images