LED lighting scene control method and system fusing user habit recognition

Abstract

The application discloses a kind of fusion user habit recognition's LED lighting scene control method and system, comprising: first determine the LED lighting target area to be handled and extract multiple categories of perception information, including environmental state and user behavior information.Second, using the preset user habit mode benchmark library, the extracted perception information is compared with the standard perception characteristics in the library, so as to match multiple source mode identifiers for each type of information. Then, at least one pending control mode is determined based on these multiple source mode identifiers. Finally, the final LED lighting operating state parameters are generated by combining the pending control mode with real-time perception information. Through multi-source information fusion and habit pattern recognition, the application realizes adaptive, personalized and precise control of the lighting scene, significantly improving user experience and energy efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an LED lighting scene control method and system that integrates user habit recognition. Background Technology

[0002] Existing LED lighting scene control methods typically rely on preset fixed patterns or simple environmental sensor feedback. These methods lack a deep understanding and integration of users' long-term usage habits and real-time behavioral intentions. For example, the system may only switch on / off or adjust brightness based on light intensity or the presence of a human body, unable to distinguish whether the user is working, reading, or relaxing, let alone adaptively learn the personalized preferences of different users in the same environment. This results in insufficiently intelligent and personalized lighting control, making it difficult to provide a truly comfortable and appropriate scene lighting environment, and also limiting further improvements in energy efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide an LED lighting scene control method and system that integrates user habit recognition.

[0004] In a first aspect, embodiments of the present invention provide an LED lighting scene control method integrating user habit recognition, the method comprising:

[0005] The target area of ​​LED lighting to be processed is determined, and various types of perception information are extracted from the target area of ​​LED lighting.

[0006] A user habit pattern benchmark library is established; the user habit pattern benchmark library includes standard perception features of standard LED lighting target areas under the various categories and standard pattern identifiers corresponding to each standard perception feature;

[0007] Based on the consistency comparison results between the multiple categories of perception information and the standard perception features of the standard LED lighting target area under the multiple categories, the multi-source mode identifiers of the multiple categories of perception information are obtained from the user habit pattern benchmark library.

[0008] Based on the multi-source mode identifiers of the various types of perception information, at least one undetermined control mode is determined for the LED lighting target area;

[0009] Based on the at least one undetermined control mode and the multiple types of perception information, control mode inference is performed to obtain LED lighting operation status parameters for the LED lighting target area.

[0010] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0011] Compared to existing technologies, the beneficial effects of this invention include: The LED lighting scene control method and system disclosed in this invention, which integrates user habit recognition, determines the target area of ​​the LED lighting to be processed and extracts multi-category perceptual information, including environmental state and user behavior information. Secondly, using a pre-set user habit pattern benchmark library, the extracted perceptual information is compared with the standard perceptual features in the library to match multi-source pattern identifiers for each type of information. Next, based on these multi-source pattern identifiers, at least one pending control mode is determined. Finally, intelligent inference is performed by combining the pending control mode with real-time perceptual information to generate the final LED lighting operating state parameters. This invention, through multi-source information fusion and habit pattern recognition, achieves adaptive, personalized, and precise control of lighting scenes, significantly improving user experience and energy efficiency. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a specific implementation of the LED lighting scene control method that integrates user habit recognition provided in the embodiments of the present invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0016] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the LED lighting scene control method integrating user habit recognition provided in this embodiment of the disclosure. The following is a detailed description of the LED lighting scene control method integrating user habit recognition.

[0018] Step S201: Determine the target area of ​​LED lighting to be processed, and extract various types of perception information for the target area of ​​LED lighting;

[0019] Step S202: Determine the user habit pattern benchmark library; the user habit pattern benchmark library includes standard perception features of the standard LED lighting target area under the various categories and the standard pattern identifiers corresponding to each of the standard perception features;

[0020] Step S203: Based on the consistency comparison results between the multiple categories of perception information and the standard perception features of the standard LED lighting target area under the multiple categories, the multi-source pattern identifiers of the multiple categories of perception information are obtained from the user habit pattern benchmark library.

[0021] Step S204: Based on the multi-source mode identifiers of the various types of perception information, determine at least one pending control mode for the LED lighting target area;

[0022] Step S205: Based on the at least one pending control mode and the multiple types of perception information, control mode inference is performed to obtain LED lighting operation status parameters for the LED lighting target area.

[0023] In this embodiment of the invention, for example, the server first determines the target area for LED lighting to be processed. This area can be a specific physical space, such as a family living room, office workstation, hotel room, or shopping mall display area. The server collects raw data in real time through a sensor network deployed in this area (such as light sensors, temperature and humidity sensors, human infrared sensors, cameras, sound acquisition devices, etc.). For the target area of ​​LED lighting, the server extracts multiple categories of perception information, which are mainly divided into two categories: environmental state perception information and user behavior perception information.

[0024] Environmental status perception information includes light intensity, color temperature, ambient temperature, humidity, time information (such as daytime, nighttime, and specific time), and spatial layout features (such as area size and furniture position). For example, in a family living room scenario, the server uses a light sensor to obtain the current natural light intensity as 300 lux, a temperature and humidity sensor to obtain the ambient temperature as 25 degrees Celsius and humidity as 50%, a clock module to obtain the current time as 19:30, and a camera to assist in identifying the layout of furniture such as sofas and televisions in the living room. User behavior perception information includes the user's activity type, location, action frequency, and voice commands. For example, the server uses a human infrared sensor to detect that the user is sitting still in the living room sofa area, uses a voice acquisition device to recognize that the user said "watch a movie," and uses a motion sensor to record that the user got up relatively few times in the past hour. The server preprocesses this raw perception information (such as noise reduction, normalization, and feature extraction) to form structured perception data for subsequent analysis.

[0025] The server maintains a user habit pattern benchmark library, which is built through historical data learning or pre-configuration. The benchmark library stores the standard perceptual characteristics of standard LED lighting target areas under various categories and their corresponding standard pattern identifiers. Standard perceptual characteristics are abstract descriptions of typical scenarios, and standard pattern identifiers are unique labels corresponding to these characteristics, used to characterize specific user habits or scenario requirements.

[0026] For example, under the category of environmental state perception, standard perception features might include "low-light environment at night" (feature vector includes light intensity < 50 lux, time 20:00-6:00), "high-light environment during the day" (light intensity > 500 lux, time 8:00-18:00), and "comfortable temperature and humidity environment" (temperature 22-26 degrees Celsius, humidity 40%-60%), with corresponding standard pattern identifiers of "night mode," "day mode," and "comfortable mode," respectively. Under the category of user behavior expression, standard perception features might include "leisure movie-watching behavior" (features include user location on a sofa, still action, voice command "watch movie"), "reading and learning behavior" (user location at a desk, action of turning pages, voice command "read"), and "party entertainment behavior" (multi-user detection, high-frequency action, noisy sound), with corresponding standard pattern identifiers of "movie-watching preference," "reading preference," and "party preference," respectively. Each entry in the benchmark library is associated with a standard perception feature vector and a standard pattern identifier, allowing the server to perform pattern matching based on these entries.

[0027] The server performs a consistency comparison between the extracted perceptual information of various categories and the standard perceptual features in the user habit pattern benchmark library, thereby obtaining multi-source pattern identifiers for each category of perceptual information. Specifically, the server first performs feature extraction operations on environmental state perceptual information and user behavior perceptual information respectively to obtain environmental state features and behavioral preference features. This step is implemented through a dual-branch feature extractor, which is trained to extract discriminative feature vectors from environmental data and behavioral data respectively.

[0028] Next, the server performs consistency judgment. For environmental state perception information, the server performs environmental consistency judgment by comparing its environmental state features with standard environmental perception features and standard behavioral perception features in the benchmark library. Environmental consistency judgment is achieved by calculating the similarity between feature vectors (such as cosine similarity, Euclidean distance). For example, the server calculates that the current environmental state features (light intensity 300 lux, time 19:30, temperature 25 degrees Celsius) have a low similarity to the standard environmental perception feature "nighttime low light environment" (light intensity <50 lux, time 20:00-6:00), but a high similarity to "comfortable temperature and humidity environment" (temperature 22-26 degrees Celsius, humidity 40%-60%). Therefore, the matching comparison result points to "comfortable mode". At the same time, the server also compares the environmental state features with standard behavioral perception features. For example, it finds that the current environmental state (night, comfortable temperature) has a certain correlation with the standard behavioral perception features of "leisure movie watching behavior" (usually occurs at night, in a comfortable environment), which may lead to a "movie watching preference" judgment label. Based on these comparison results, the server obtains multi-source pattern identifiers for environmental state perception information, such as "comfort mode" and "viewing preferences".

[0029] For user behavior perception information, the server performs text consistency judgment (here, "text" refers to the semantic matching of behavior features) on its behavior preference features against standard environment perception features and standard behavior perception features. For example, the server calculates that the current behavior preference features (location: sofa, still, command: "watch a movie") have a high similarity to the standard behavior perception feature "leisurely movie watching behavior," and the matching result points to "movie watching preference." Simultaneously, the server also compares the behavior preference features with standard environment perception features; for example, it finds that "watching a movie" behavior often occurs in "low-light environments at night," thus potentially obtaining a "night mode" identification. Combining these comparison results, the server obtains multi-source pattern identifiers for user behavior perception information, including, for example, "movie watching preference" and "night mode." In this way, the server matches multiple possible multi-source pattern identifiers for each category of perception information, reflecting different perspectives on the current scene.

[0030] Based on the multi-source pattern identifiers obtained in step three, the server determines at least one pending control mode for the LED lighting target area. The pending control mode is a preliminary control strategy for the lighting scene, integrating multi-source pattern identifiers from environmental conditions and user behavior. The server generates the pending control mode by fusing and analyzing the multi-source pattern identifiers using a rule engine or machine learning model (such as a classifier).

[0031] For example, the server obtains multi-source pattern identifiers for environmental state perception information as "Comfort Mode" and "Viewing Preference," and multi-source pattern identifiers for user behavior perception information as "Viewing Preference" and "Night Mode." Analyzing these identifiers, the server finds that "Viewing Preference" appears in both, and that "Night Mode" matches the current time, while "Comfort Mode" matches the current temperature and humidity. Therefore, the server might generate a pending control mode, "Home Theater Mode," which initially defines lighting requirements: low overall illuminance, high contrast, and warm color tone to avoid screen glare. Simultaneously, the server might also generate another pending control mode, "Night Comfort Mode," as an alternative, emphasizing low but uniform lighting to maintain comfort. The server represents these pending control modes as a structured set of control parameters, including dimensions such as target light intensity, color temperature, and dynamic effects, but the specific parameter values ​​have not yet been finalized.

[0032] The server performs final control mode inference based on at least one pending control mode and multiple categories of perception information to obtain accurate LED lighting operating state parameters. This step is achieved through a control mode inference model that considers the semantics of the pending control modes as well as the detailed features of the current environment.

[0033] First, the server combines user behavior perception information and pending control modes to construct a control semantic representation. For example, for the pending control mode "home theater mode", the server combines the specific instruction "watch a movie" and the location "sofa" in the user behavior perception information to construct a control semantic vector that represents "the user is located in the sofa area, intends to watch a movie, and needs theater-style lighting".

[0034] Secondly, the server performs feature mapping operations on the environmental state features corresponding to the environmental state perception information through a feature space mapping model to obtain environmental state mapping features. This mapping model transforms the original environmental features (such as 300 lux of light intensity and 25 degrees Celsius of temperature) into a higher-level feature space that is more suitable for control decisions. For example, it maps the current light intensity to "moderate ambient light but with potential for screen interference" and the temperature to "suitable for long-term stay".

[0035] Finally, the server uses a control mode inference model based on control semantic representations and environmental state mapping features for inference. This model is a deep learning network (such as a multilayer perceptron or Transformer) that receives control semantic representations and environmental state mapping features as input and outputs the final LED lighting operating state parameters. For example, the model might output specific parameters: overall illuminance reduced to 30 lux, wall washer lights on the TV background wall turned on with a color temperature of 2700K, weak contour lighting (5 lux) around the sofa, and lights in other areas turned off. These parameters comprehensively consider the requirements of the pending control modes, the actual situation of the current ambient light (avoiding excessive darkness that could cause eye fatigue), and the user's specific location (ensuring sufficient but not glaring auxiliary light in the sofa area). If multiple pending control modes exist, the model will evaluate the applicability of each mode and may fuse them to generate an optimal parameter set, or select one mode based on confidence level.

[0036] The server sends the final LED lighting operating status parameters to the LED lighting system's controller. The controller then drives the luminaires to adjust brightness, color, and distribution, thereby achieving adaptive and personalized lighting scene control. The entire process runs in real time, with the server continuously monitoring changes in the sensed information and dynamically updating control parameters to adapt to changes in user habits and environmental conditions.

[0037] Through the above steps, the LED lighting scene control method of this invention can deeply integrate user habit recognition to achieve intelligent and scenario-based lighting control, thereby improving energy efficiency and user experience.

[0038] Specifically, the step of determining at least one pending control mode can be implemented through a mode fusion decision module. This module maintains a "mode identifier-control template" mapping table, which is part of or associated with a user habit mode benchmark library. Each standard mode identifier (such as "viewing preference" or "night mode") is mapped to one or more preset, parameterized "control strategy templates" (such as "home theater mode templates," which define the adjustment range and priority of brightness, color temperature, and dynamic effects).

[0039] The fusion analysis follows these rules:

[0040] 1. Voting and Weighting: All multi-source pattern identifiers from environmental state perception information and user behavior perception information are statistically analyzed. Identifiers with higher frequency of occurrence receive higher weights. For example, if "movie viewing preference" appears in both environmental and behavioral information identifiers, its weight is doubled.

[0041] 2. Conflict resolution: If seemingly conflicting labels appear (such as "bright environment" and "sleep preference"), the label with a higher direct match to the current user's perceived behavior information will be given priority, or the label with a higher probability of co-occurrence with the current time and environmental context in historical data will be given priority.

[0042] 3. Template Retrieval and Instantiation: Based on the final set of identifiers after weighted summation and resolution, the corresponding control strategy template is retrieved from the "Pattern Identifier-Control Template" mapping table. If multiple identifiers correspond to the same template, the template is strengthened; if they correspond to different templates, the template parameters can be weighted and averaged, or multiple pending control pattern instances can be generated for selection in the subsequent inference stage.

[0043] 4. Context Injection: The current original sensing information (such as specific illuminance values) is used as context parameters and injected into the retrieved control strategy template to form a preliminary description of the "pending control mode" containing specific parameter ranges, which is then passed to the next stage.

[0044] In this embodiment of the invention, the step of matching and obtaining the multi-source pattern identifiers of the various categories of perception information from the user habit pattern benchmark library based on the consistency comparison results between the multiple categories of perception information and the standard perception features of the standard LED lighting target area under the multiple categories can be implemented through the following example.

[0045] Each type of perception information in the multiple categories is compared with the standard perception features of the standard LED lighting target area under the multiple categories to obtain the consistency comparison result of each type of perception information.

[0046] From the consistency comparison results of each type of sensing information, determine the matching comparison results of each type of sensing information.

[0047] Based on the standard perceptual features corresponding to the adaptation ratio results of each type of perceptual information, the multi-source pattern identifiers of each type of perceptual information are obtained by matching from the user habit pattern benchmark library.

[0048] In an embodiment of the invention, for example, the server first separates different categories of data streams from the perceived information. In a home bedroom scenario, the perceived information extracted by the server mainly includes two categories: environmental state perceived information (category A) and user behavior perceived information (category B).

[0049] Category A: Environmental State Awareness Information. The server obtains the current time as 22:15 through the sensor network. The outdoor light sensor reading is 0 lux (complete darkness), the indoor central light sensor reading is 100 lux (the brightness of a small night light), and the temperature and humidity sensors show a temperature of 23 degrees Celsius and a humidity of 55%. The server transforms this raw data into a structured environmental state feature vector, such as [Time: 22:15, Ambient Light: 0 lux, Indoor Light: 100 lux, Temperature: 23°C, Humidity: 55%].

[0050] Category B: User Behavior Perception Information. The server detects that the user is lying down through the mattress pressure sensor, detects that the user's heart rate has dropped to the resting level through the heart rate monitoring module of the wearable device, and detects that no speech or significant activity sounds have been detected in the previous 5 minutes through the sound sensor. The server converts this data into a behavioral preference feature vector, such as [posture: lying down, physiological state: resting heart rate, sound activity: low].

[0051] Next, the server compares the perceptual information feature vector of each category with the standard perceptual features of all categories in the user habit pattern benchmark library. The benchmark library stores, for example:

[0052] Environmental state category standard feature 1 (standard perception feature A1): [Time: 21:00-6:00, ambient light: <5 lux, indoor light: 50-150 lux], its corresponding standard mode identifier is "nighttime bedtime environment".

[0053] Environmental condition standard feature 2 (standard perception feature A2): [Time: 18:00-21:00, Indoor lighting: 300-500 lux], corresponding to the label "Evening activity environment".

[0054] User behavior standard feature 1 (standard perception feature B1): [Posture: sitting / standing, sound activity: medium / high], corresponding to the label "active behavior preference".

[0055] User behavior standard feature 2 (standard perception feature B2): [Posture: lying down, physiological state: resting, sound activity: low], corresponding to the label "sleep preparation preference".

[0056] The server calculates the similarity (consistency comparison result) between the current environmental state feature vector and four standard feature vectors: A1, A2, B1, and B2. For example, the similarity score with A1 is high (matching in both time and lighting), the similarity score with A2 is low (partially matching in time but not in lighting), and the similarity scores with B1 and B2 are low and moderate, respectively (because B1 / B2 mainly describe behavior, and their direct comparability with environmental features is weak, but they can still be calculated through shared dimensions such as "time" or implicit associations). Similarly, the server also calculates the similarity between the current behavioral preference feature vector and these four standard features.

[0057] The server analyzes all comparison results for each type of perceived information. Instead of simply selecting the highest score, it determines which comparison result(s) is "fit" based on preset adaptation rules (such as threshold judgment and cross-category correlation weighting).

[0058] Regarding environmental state perception information, its similarity score with "Standard Perception Feature A1" (nighttime bedtime environment) is significantly higher than others, exceeding the adaptation threshold. Therefore, the server determines this comparison result as an adaptation comparison result. Although it also has some similarity with "Standard Perception Feature B2" (sleep preparation preferences), this score does not reach the threshold for being considered as the primary adaptation result for environmental information.

[0059] For user behavior perception information, it has the highest similarity score with "Standard Perception Feature B2" (sleep preparation preference), and the core dimensions (posture, physiological state) are completely matched. Therefore, this comparison result is determined to be the fit comparison result. At the same time, the server found that this behavior information also has a strong logical correlation with "Standard Perception Feature A1" (nighttime bedtime environment) (sleep behavior usually occurs in a specific nighttime environment), and its similarity score may also exceed a minor fit threshold. Therefore, it may be determined as a minor fit comparison result.

[0060] Finally, based on the adaptation comparison results determined in the previous step, the server locates the corresponding standard perceptual features in the user habit pattern benchmark library and reads the standard pattern identifiers bound to these features. These identifiers are the "multi-source pattern identifiers" obtained by matching the perceptual information.

[0061] For the environmental state perception information, the master adaptation comparison result points to "standard perception feature A1", so the server obtains its multi-source mode identifier as "nighttime bedtime environment". This identifier is an interpretation of the current scene from the perspective of environmental state category.

[0062] For user behavior perception information, the primary adaptation comparison result points to "standard perception feature B2," thus the server obtains a multi-source pattern identifier as "sleep preparation preference." Simultaneously, since the secondary adaptation comparison result points to "standard perception feature A1," the server may also obtain another multi-source pattern identifier, "nighttime bedtime environment." This means that from user behavior, the server not only interprets the user's preference for "preparing to sleep," but also infers the contextual information of "the current possible nighttime bedtime environment."

[0063] At this point, the server has completed the transformation from raw perceived information to high-level, understandable multi-source pattern identifiers. Environmental information is identified as "pre-bedtime environment," and behavioral information is identified as "sleep preparation preferences" and "pre-bedtime environment." These multi-source, multi-perspective identifiers provide rich decision-making basis for subsequent fusion judgments and the generation of precise lighting control instructions (e.g., turning off the main light, adjusting the night light to the lowest amber light level, and turning it off completely after 30 minutes).

[0064] In this embodiment of the invention, the multiple categories include environmental state perception categories and user behavior expression categories; the perception information of the multiple categories includes environmental state perception information and user behavior perception information of the LED lighting target area; the user habit pattern benchmark library includes standard environmental perception features of the standard LED lighting target area under the environmental state perception category and environmental state pattern identifiers corresponding to the standard environmental perception features, as well as standard behavior perception features of the standard LED lighting target area under the user behavior expression category and behavior preference pattern identifiers corresponding to the standard behavior perception features;

[0065] The step of matching the multi-source pattern identifiers of the various categories of perception information from the user habit pattern benchmark library based on the consistency comparison results between the multiple categories of perception information and the standard perception features of the standard LED lighting target area under the multiple categories can be implemented through the following example.

[0066] Based on the environmental state perception information and the user behavior perception information, feature extraction operations are performed respectively to obtain environmental state features and behavioral preference features;

[0067] The environmental state features are compared with the standard environmental perception features and the standard behavioral perception features to determine environmental consistency. Based on the environmental consistency comparison results, the multi-source pattern identifier of the environmental state perception information is obtained from the environmental state pattern identifier and the behavioral preference pattern identifier.

[0068] The behavioral preference features are compared with the standard environmental perception features and the standard behavioral perception features for text consistency judgment. Based on the text consistency comparison results, the multi-source pattern identifier of the user behavior perception information is obtained from the environmental state pattern identifier and the behavioral preference pattern identifier.

[0069] In an embodiment of the invention, for example, the server determines that the target area for LED lighting to be processed is a study. The server collects raw data through a sensor network in that area.

[0070] The server receives environmental status information: the light sensor shows an illuminance of 200 lux in the desk area, the color temperature sensor shows a current light color temperature of 4000K, the clock module provides the time as Saturday 14:30, and the outdoor weather interface returns "cloudy". The server preprocesses this information and extracts features to generate a structured environmental status feature vector. This vector might be represented as [Time label: weekend afternoon, Location illuminance: 200 lux, Current color temperature: 4000K, Natural light conditions: cloudy, low light]. This feature extraction process may be accomplished by a trained environmental feature encoder, which maps the raw sensor readings into a dense vector containing semantic information.

[0071] Simultaneously, the server receives user behavior perception information: the camera (after privacy anonymization) detects that the user is sitting at a desk, leaning forward; the desktop pressure sensor detects the placement of a book; the microphone array captures continuous page-turning sounds and intermittent writing sounds, and no voice commands or leaving actions have been detected in the past 15 minutes. The server extracts features from this behavioral data to generate a behavioral preference feature vector. This vector might be represented as [User Posture: Sitting posture, Desktop object interaction: Book, Audio features: Continuous page-turning / writing sounds, Activity persistence: High]. This process is achieved through a behavioral feature encoder, which extracts abstract features representing user intent and preferences from multimodal behavioral data.

[0072] The server performs an "environmental consistency judgment" by comparing the environmental state feature vector obtained in the previous step with two types of standard features in the user habit pattern benchmark library. This judgment focuses on measuring similarity at the physical environment level.

[0073] First, the server compares the environmental state characteristics with standard environmental awareness characteristics in the benchmark library. The benchmark library stores features such as:

[0074] Standard environmental perception feature 1 (corresponding to the environmental state mode identifier "Focus on work environment"): The feature vector is described as [Time: Weekday daytime / Weekend afternoon, Location illuminance requirement: 300-500 lux, Color temperature preference: 5000K cool white light, Natural light supplement: Required].

[0075] Standard environmental perception feature 2 (corresponding to the environmental state mode identifier "leisure reading environment"): The feature vector is described as [time: evening, location illuminance requirement: 150-250 lux, color temperature preference: 2700K-4000K warm yellow light, natural light supplement: not required].

[0076] The server calculates the similarity between the current environmental state characteristics and these standard characteristics. Since the current time is a weekend afternoon, the illuminance of 200 lux is slightly lower than the ideal value for a "focused work environment," and the color temperature is 4000K, it is closer to a "leisure reading environment" in terms of illuminance and color temperature, but matches the "focused work environment" in terms of time attribute. After weighted calculation, assuming a higher similarity score with the "leisure reading environment," the server determines this as the appropriate comparison result, thus obtaining a multi-source pattern identifier: "leisure reading environment."

[0077] Next, the server compares the same environmental state features with standard behavior-aware features in the benchmark library to determine environmental consistency. While standard behavior-aware features primarily describe behavior, their vector space also contains contextual information typically associated with it. For example:

[0078] Standard behavior perception feature A (corresponding to the behavior preference pattern identifier "deep reading preference"): its vector may be implicitly associated with context such as [ambient lighting: stable and moderate, ambient noise: low].

[0079] Standard behavioral perception feature B (corresponding to the behavioral preference pattern identifier "creative writing preference"): its vector may be implicitly associated with context such as [ambient lighting: bright, ambient color temperature: cool].

[0080] The server calculates the similarity between the current environmental state characteristics (200 lux, 4000 K, cloudy) and the environmental context implied by these behavioral characteristics. Assuming a high match is found between the context of "stable moderate lighting" associated with "deep reading preference," the server obtains a second multi-source pattern identifier: "deep reading preference." Therefore, from the environmental state information, the server ultimately parses two multi-source pattern identifiers: "leisure reading environment" and "deep reading preference."

[0081] The server performs "textual consistency judgment" on the behavioral preference feature vector obtained in step one and the two types of standard features in the benchmark library. Here, "text" is used in a broad sense, referring to the matching of the semantics, intent, and contextual logic of behavioral patterns, with an emphasis on the classification of behavioral patterns.

[0082] First, the server compares the behavioral preference features with standard behavioral perception features in a benchmark library. It calculates the semantic similarity between the current behavioral features (sitting posture, book interaction, continuous page-turning and writing sounds) and standard behavioral features such as "deep reading preference" and "creative writing preference." Clearly, the match with "deep reading preference" is the highest, as this standard feature describes prolonged, focused interaction with a book. Based on this, the server determines an appropriate comparison result, obtaining a core multi-source pattern identifier for the user's behavioral perception information: "deep reading preference."

[0083] Then, the server performs text consistency judgment between the same behavioral preference feature and standard environment-aware features in the benchmark library. This judgment determines which standard environment mode the currently observed user behavior typically occurs in. The server analyzes the behavior of "deep reading" to determine the most frequently corresponding environment mode in historical habit data. By querying association models or calculating similarity, the server finds that the "deep reading preference" behavior is strongly associated with both "leisure reading environment" and "focused work environment," but in the context of a weekend afternoon, the association weight with "leisure reading environment" is higher. Therefore, the server derives another multi-source pattern identifier from the behavioral information: "leisure reading environment."

[0084] Ultimately, from the user behavior perception information, the server parsed out two multi-source pattern identifiers: "deep reading preference" and "leisure reading environment".

[0085] Through the above process, the server derives a set of mutually corroborating yet slightly emphasizing pattern identifiers from two independent data sources: environment and behavior. (Environmental information emphasizes the physical state of "leisure reading environment," while behavioral information emphasizes the user's intent of "deep reading preference.") This multi-source, cross-cutting identifier acquisition method enhances the robustness and accuracy of pattern recognition.

[0086] In this embodiment of the invention, the step of performing environmental consistency judgment between the environmental state features and the standard environmental perception features and the standard behavioral perception features, and obtaining the multi-source pattern identifier of the environmental state perception information from the environmental state pattern identifier and the behavioral preference pattern identifier based on the environmental consistency comparison result, can be implemented through the following example.

[0087] The environmental state features are compared with the standard environmental perception features to determine environmental consistency, and a first environmental consistency comparison result is obtained.

[0088] Based on the standard environmental perception features corresponding to the first environmental consistency comparison result of the adaptation, the first environmental state perception category determination identifier is determined from the environmental state mode identifier;

[0089] The environmental state features are compared with the standard behavioral perception features to determine environmental consistency, and a second environmental consistency comparison result is obtained.

[0090] Based on the standard behavior perception features corresponding to the second environment consistency comparison results, a first user behavior expression category determination identifier is determined from the behavior preference pattern identifier;

[0091] Based on the first environmental state perception category determination identifier and the first user behavior expression category determination identifier, the multi-source pattern identifier of the environmental state perception information is obtained.

[0092] In this embodiment of the invention, for example, the server determines the target area for LED lighting to be the workstation area of ​​employee Zhang San. Through sensors in this area, the server obtains environmental status perception information: the desktop illuminance sensor reading is 400 lux, the color temperature sensor shows 5000K, the ambient noise sensor detects a background noise value of 45 dB (typical office background noise), the indoor carbon dioxide sensor reading is 800 ppm (good air quality), and the public area sensors indicate that the lighting in the entire office area is in "standard office mode". The server extracts features from this information to generate an environmental status feature vector, for example, [desktop illuminance: 400 lux, light color temperature: 5000K, background noise: medium office level, air quality: good, overall lighting mode: standard office].

[0093] The server first performs the first stage of "environmental consistency judgment": comparing this environmental state feature vector with the standard environmental awareness feature set in the user habit pattern benchmark library. The benchmark library pre-stores various standard environmental awareness features and their identifiers for office scenarios, such as:

[0094] Standard environmental perception feature E1 (vector description): [Desktop illuminance range: 450-600 lux, color temperature range: 5000-6000K, expected background noise: low to medium, core requirement: high-focus visual task], the corresponding environmental state mode is identified as "high-intensity focused office environment".

[0095] Standard environmental perception feature E2 (vector description): [Desktop illuminance range: 300-450 lux, color temperature range: 4000-5000K, expected background noise: moderate, core requirement: routine document processing], corresponding to the label "routine office environment".

[0096] Standard environmental perception feature E3 (vector description): [desktop illuminance range: 200-300 lux, color temperature range: 2700-4000K, expected background noise: low, core need: short rest or thinking], corresponding to the label "relaxation / rest environment".

[0097] The server calculates the similarity between the current environmental characteristics and E1, E2, and E3. The current illuminance of 400 lux and color temperature of 5000K fall within the typical range of E2 (a typical office environment), and are close to the lower limit of E1 (high-intensity focus) but not its ideal value, while differing significantly from E3 (a relaxing environment). Meanwhile, background noise and air quality conditions are compatible with both E1 and E2. After comprehensive similarity calculation (possibly using cosine similarity or rule-based weighted scoring), the server determines that the match with the standard environmental perception characteristics E2 is the highest, exceeding the preset adaptation threshold. Therefore, the first environmental consistency comparison result points to E2.

[0098] Based on this matching comparison result, the server searches the benchmark library and determines the environmental state mode identifier corresponding to E2, namely "normal office environment". This identifier is recorded as the first environmental state perception category determination identifier. This means that the server first classifies the current workstation state as a typical normal office lighting environment from the perspective of purely physical environmental parameters.

[0099] Next, the server performs the second stage of "environmental consistency judgment": comparing the same environmental state feature vector ([desktop illuminance: 400 lux, light color temperature: 5000K, background noise: medium...]) with the standard behavioral perception feature set in the user habit pattern benchmark library. Here, the judgment is no longer about the similarity between environments, but rather "which habitual behavioral pattern of the user typically accompanies this current environmental state." The benchmark library stores standard behavioral features, such as:

[0100] Standard behavioral perception feature B1 (vector description): This feature contains the environmental context in which it typically occurs, such as [associated ambient lighting: medium to high brightness cool white light, associated environmental state: low interference], and the corresponding behavioral preference pattern is identified as "immersive programming / analysis preference".

[0101] Standard behavioral perception feature B2 (vector description): implies context such as [associated ambient lighting: moderate brightness neutral light, associated ambient state: acceptable background activity], the corresponding label is "document processing and communication preferences".

[0102] Standard behavioral perception feature B3 (vector description): implies context such as [associated ambient lighting: low brightness warm light, associated ambient state: quiet], the corresponding label is "creative thinking or reading preference".

[0103] The server calculates the similarity between the current environmental state characteristics and the "associated environment" contexts implied by B1, B2, and B3. The current environment of 400 lux and 5000 K highly matches the associated environment description of B2 (document processing and communication) as "moderate brightness and neutral light," and "moderate office background noise" also aligns with the description of "acceptable background activity." Compared to the associated environment of B1 (immersive programming) as "medium to high brightness cool white light," the current illuminance is slightly lower; and it differs significantly from the associated environment of B3 (creative thinking). Therefore, the second environment consistency comparison result points to B2.

[0104] Based on this matching comparison result, the server determines the behavioral preference pattern identifier corresponding to B2 from the benchmark library, namely "document processing and communication preference". This identifier is recorded as the first user behavior expression category determination identifier. Essentially, this step involves the server inferring the type of behavior the user is most likely to be performing or about to perform under the observed environmental conditions.

[0105] The server now has two judgment indicators from different perspectives:

[0106] The first environmental state perception category identification label is "routine office environment". This is a direct classification of the physical state of the environment.

[0107] The first user behavior expression category is identified by the label: "Document processing and communication preferences." This is an indirect inference based on the user's potential behavioral intentions according to the environmental state.

[0108] The server combines these two identifiers to form a multi-source pattern identifier for environmental state perception information. In other words, when interpreting environmental information, the server draws a dual conclusion: "The current environment is a typical office environment, and this environment suggests that the user may be in a work preference mode of document processing and communication."

[0109] This multi-source identifier set contains richer semantic information than a single "typical office environment." It not only describes "how the lighting is" but also hints at "what the user might be doing." This provides more refined clues for subsequent lighting control decisions. For example, if subsequent directly received user behavior information (such as detecting prolonged sitting in front of a screen or frequent keyboard input) aligns with the inferred "document processing" preference, the system can maintain the current lighting parameters. If the behavior information deviates from this inference (such as detecting a user starting frequent conference calls), the system can combine the new behavior identifiers to consider whether fine-tuning the lighting is needed to suit the "communication" scenario (such as slightly reducing the center illuminance of the desktop to avoid facial overexposure during video conferences). In this way, starting only from environmental information, the system has already completed a preliminary habit-based pattern recognition and intent prediction.

[0110] In this embodiment of the invention, the step of performing text consistency judgment between the behavioral preference features and the standard environment perception features and the standard behavioral perception features, and obtaining the multi-source pattern identifier of the user behavior perception information from the environment state pattern identifier and the behavioral preference pattern identifier based on the text consistency comparison result, can be implemented through the following example.

[0111] The behavioral preference features are compared with the standard environmental perception features to determine text consistency, and a first text consistency comparison result is obtained.

[0112] Based on the standard environmental perception features corresponding to the first text consistency comparison result, a second environmental state perception category determination identifier is determined from the environmental state pattern identifier.

[0113] The behavioral preference features are compared with the standard behavioral perception features to determine text consistency, and a second text consistency comparison result is obtained.

[0114] Based on the standard behavior perception features corresponding to the adapted second text consistency comparison result, a second user behavior expression category determination identifier is determined from the behavior preference pattern identifier;

[0115] Based on the second environmental state perception category determination identifier and the second user behavior expression category determination identifier, the multi-source pattern identifier of the user behavior perception information is obtained.

[0116] In this embodiment of the invention, for example, the server determines the target area for LED lighting to be the living area of ​​a hotel room. Through sensors in this area, the server acquires user behavior perception information: a human infrared sensor detects that the user is seated on the sofa; the smart TV's status interface returns information indicating that the TV is on and playing a movie; a sound sensor analyzes ambient audio, confirming the presence of continuous, cinematic surround sound effects without any human voice dialogue; and a motion sensor detects no large-scale user movement in the past 30 minutes. The server extracts features from this behavioral data to generate a behavioral preference feature vector, such as [User position: sofa stationary, media device status: TV playing, audio scene: cinematic sound effects dominant, activity level: extremely low].

[0117] The server first performs the first stage of "text consistency judgment": comparing this behavioral preference feature vector with the standard environment-aware feature set in the user habit pattern benchmark library. The core of this "text consistency judgment" lies in logical association and contextual reasoning, that is, determining "in which standard environment mode the observed user behavior typically occurs." The benchmark library pre-stores standard environment-aware features and their identifiers for hotel room scenarios, for example:

[0118] Standard environmental perception feature E1 (vector description): [Overall illuminance: extremely low (<20 lux), accent lighting: soft fill light for TV background wall, color temperature: warm yellow light below 2700K, core objective: reduce screen reflection and create an immersive experience], the corresponding environmental state mode is identified as "cinema viewing environment".

[0119] Standard environmental perception feature E2 (vector description): [Overall illuminance: moderate (150-300 lux), lighting distribution: uniform, color temperature: neutral light around 4000K, core objective: provide comfortable general activity lighting], corresponding to the label "leisure and living environment".

[0120] Standard environmental perception feature E3 (vector description): [Overall illuminance: bright (>300 lux), lighting distribution: concentrated on the desk, color temperature: 5000K cool white light, core objective: to meet the visual needs of reading or working], corresponding to the label "reading / working environment".

[0121] The server does not directly calculate the numerical similarity between behavioral feature vectors and environmental feature vectors. Instead, it analyzes based on a pre-trained logical model or association rules. The model determines that the behavioral feature "sofa stationary, TV playing cinematic sound effects, extremely low activity level" has the strongest logical correlation with the "cinema viewing environment" described by the standard environmental perception feature E1, because this is a projection of typical movie-watching behavior onto the environment. This behavior has a weaker correlation with E2 (general leisure) and E3 (work reading). Therefore, the first text consistency comparison result points to E1.

[0122] Based on this matching comparison result, the server searches the benchmark library and determines the environmental state mode identifier corresponding to E1, namely "cinema viewing environment". This identifier is recorded as the second environmental state perception category determination identifier. This means that by analyzing user behavior, the server infers in reverse the standard mode that the lighting environment should be in or tend towards in order to support the current behavior. This is an environmental state prediction based on behavioral intent.

[0123] Next, the server performs the second stage of "text consistency judgment": directly comparing the same behavioral preference feature vector ([user location: sofa stationary, media device status: TV playing, audio scene: cinema sound effects...]) with the standard behavioral perception feature set in the user habit pattern benchmark library. This is a classification of the behavioral pattern itself. The benchmark library stores standard features at the behavioral level, such as:

[0124] Standard behavioral perception feature B1 (vector description): [Main activity: watching film and television content, typical posture: sitting / lying still, interactive object: TV / projector, sound features: mainly media sound effects], the corresponding behavioral preference mode is identified as "immersive viewing preference".

[0125] Standard behavioral perception feature B2 (vector description): [Main activity: reading or using a tablet, typical posture: sitting, interaction object: book / handheld device, sound characteristics: quiet or soft music], the corresponding label is "personal quiet reading preference".

[0126] Standard behavioral perception feature B3 (vector description): [Main activity: talking with multiple people, typical posture: diverse, interaction object: no specific, voice features: multiple people speaking], the corresponding label is "social conversation preference".

[0127] The server calculates the semantic similarity between the current behavioral preference features and standard behavioral feature vectors such as B1, B2, and B3. The current behavior perfectly matches the description of B1 (immersive viewing preference) across multiple dimensions, including "main activities," "typical postures," "interaction objects," and "voice features." Therefore, the second text consistency comparison result clearly points to B1.

[0128] Based on this matching comparison result, the server determines the behavioral preference pattern identifier corresponding to B1 from the benchmark library, namely "immersive viewing preference". This identifier is recorded as the second user behavior expression category determination identifier. This is a direct identification and classification of the user's current activity.

[0129] The server now has two judgment markers that originate from behavioral information analysis but point to different dimensions:

[0130] The second environmental state perception category determination label is: "Cinema viewing environment". This is the expected state of the environment inferred from user behavior.

[0131] The second user behavior category identification identifier is "Immersive Viewing Preference." This is a direct identification of the user's current actual behavior.

[0132] The server combines these two identifiers to form a multi-source pattern identifier for user behavior perception information. In other words, when interpreting user behavior information, the server concludes that "the user is exhibiting a preference for immersive movie viewing, therefore their environment should be adapted to the lighting pattern of a movie theater."

[0133] This multi-source identifier set is highly directional. It not only indicates what the user is "doing" (watching a movie), but also strongly suggests what the lighting system "should provide" (the cinema environment). This provides a direct and strong basis for lighting control decisions. When the server integrates environmental state perception information in subsequent steps (e.g., actually detecting normal lighting with a current room illuminance of 100 lux), it can immediately detect a significant discrepancy between the actual environmental condition (normal lighting) and the desired environmental condition (cinema viewing environment) inferred from behavioral information. This will drive the system to prioritize the "cinema viewing environment" identifier derived from behavioral information as the core basis for the pending control mode, and generate corresponding control commands (such as turning off the main light, turning on the warm, dim light on the TV background wall, and adjusting the color temperature of all lights to below 2700K), thereby quickly and accurately responding to the user's real-time behavioral intentions and creating a lighting scene that conforms to their habits.

[0134] In this embodiment of the invention, the step of performing feature extraction operations based on the environmental state perception information and the user behavior perception information to obtain environmental state features and behavioral preference features is implemented by a dual-branch feature extractor. The training steps of the dual-branch feature extractor can be implemented through the following example.

[0135] Obtain scene behavior association samples, which include environmental state records and user behavior records in response to the environmental state records;

[0136] The original dual-branch feature extractor performs feature extraction operations on the environmental state record and the user behavior record respectively to obtain the environmental state features of the scene behavior associated sample of the environmental state record and the behavior preference features of the scene behavior associated sample of the user behavior record.

[0137] Based on the environmental state features and behavioral preference features of the scene behavior-related samples, the original dual-branch feature extractor is optimized and iterative training is performed to obtain the dual-branch feature extractor.

[0138] In this embodiment of the invention, for example, the server initiates the training process of the dual-branch feature extractor. First, the server needs to obtain a large number of high-quality "scene behavior association samples" from historical databases or real-time labeled data streams. Each sample is a data pair that accurately records the correspondence between a specific environmental state and the user behavior that occurs in a certain LED lighting target area.

[0139] For example, the server extracts a training sample from the historical logs of a smart living room:

[0140] Environmental Status Record: This record is a multi-dimensional time-series data package containing complete sensor readings of the living room at a specific moment. Specific data includes: {Timestamp: 2023-10-26 20:15:00, Central Illuminance: 50 lux, TV Area Illuminance: 30 lux, Sofa Area Illuminance: 10 lux, Overall Lighting Color Temperature: 2200K, Ambient Noise Level: 35 dB (Movie Sound Effects), Human Infrared Signal: Single stationary target in the sofa area, Room Temperature: 23 degrees Celsius}. This raw data constitutes a snapshot of the environmental status.

[0141] User Behavior Log: This log contains user behavior data captured at the same time or within a short period of the aforementioned environmental status log. Specific data includes: {Smart TV Status: Playing (Movie ID: "Interstellar"), Voice Command Log: Most recent command was "Turn on the TV and dim the lights", Wearable Device Status: User's heart rate is stable, Motion Sensor: No movement events in the past 15 minutes}. This data collectively describes the user's behavior and intentions in this environment.

[0142] The server collects tens of thousands of such related samples, covering a variety of typical scenarios such as "home theater," "reading," "party," "dinner," and "nighttime rest." Each sample ensures strict alignment of the environment and behavior in the temporal context, which is the basis for training a model that can understand the relationship between the two.

[0143] The server initializes a "raw dual-branch feature extractor". This network model contains two parallel subnetworks (branches) with potentially different structures: an environment feature encoding branch and a behavior feature encoding branch. Initially, the parameters of these two branches are randomized or preset, and they do not yet have effective feature extraction capabilities.

[0144] The server inputs the acquired scene behavior-related samples into this raw extractor for forward propagation calculation:

[0145] Environment Branch Processing: The server inputs the "environmental state record" (i.e., the raw sensor values) from the samples into the environment feature encoding branch. This branch may consist of fully connected layers, convolutional layers, or recurrent neural networks, and is used to process structured temporal data. The branch processes this record and outputs a fixed-dimensional vector, such as a 128-dimensional floating-point array, denoted as the scene behavior associated sample environment state feature. In the early stages of training, this feature vector is random and meaningless, and cannot effectively represent the environment state.

[0146] Behavior branch processing: Simultaneously, the server inputs the "user behavior records" (TV status, voice commands, sensor signals, etc.) from the samples into the behavior feature encoding branch. This branch may need to process multimodal data (such as text, status codes, and time-series signals), so its structure may be more complex. The branch also outputs a fixed-dimensional vector, such as another 128-dimensional floating-point array, denoted as the scene behavior association sample behavior preference feature. Initially, this feature vector cannot effectively represent the behavior pattern.

[0147] The server does not directly use these two feature vectors. Instead, it guides the two branches to learn meaningful feature representations by designing specific training objectives (loss functions). The core training principle is that environmental state features and behavioral preference features extracted from the same associated samples should be highly related or consistent semantically.

[0148] The server employs a training strategy based on contrastive learning or maximizing relevance. For example, for a sample pair (environmental record E, behavioral record B), the server extracts features Fe and Fb using an extractor. One training objective is to maximize the mutual information or cosine similarity between Fe and Fb, since they occur simultaneously and are correlated. Simultaneously, the server constructs "negative samples," such as pairing sample E with behavioral record B' from another random sample, with the training objective requiring the minimization of the similarity between Fe and Fb'.

[0149] The specific training loop is as follows:

[0150] The server retrieves a batch of scene behavior-related samples from the training set.

[0151] Inputting these samples into the original dual-branch feature extractor yields a batch of corresponding environmental state features and behavioral preference features.

[0152] The server calculates a pre-defined loss function (such as the contrastive loss InfoNCE). The value of this loss function measures the strength of the association between features of positive sample pairs and the degree of distinction between features of negative sample pairs under the current model.

[0153] The server uses the backpropagation algorithm to calculate the gradient of the loss function with respect to all parameters (weights and biases of the two branches) of the original dual-branch feature extractor.

[0154] The server uses an optimizer (such as Adam) to update the parameters of the extractor based on the gradient, which is called "tuning". The direction of tuning is to reduce the loss function, that is, to teach the model to extract features such that environmental features and behavioral features from the same scene are close to each other in the feature space, while those from different scenes are far apart.

[0155] The server repeats the above training loop, traversing the training dataset multiple times (multiple epochs) to perform iterative training.

[0156] Through iterative iterations with a preset number of loops, the parameters of the two branches are continuously adjusted and optimized. The environmental branch gradually learns to ignore sensor noise and extract advanced environmental semantic features such as "low illumination, warm color temperature, and active immersive media." The behavioral branch learns to extract advanced behavioral semantic features such as "focused media consumption, stillness, and intention for home entertainment" from a multitude of behavioral signals. More importantly, because the training objective forces the correlation between the two, the two feature vectors output by the finally trained dual-branch feature extractor, although from different branches, exist in a shared and comparable semantic space. This makes it feasible and meaningful to perform "consistency discrimination" (whether environmental consistency or textual consistency) between environmental features and behavioral features in subsequent steps. After training, the server can deploy this extractor into an online lighting control system for real-time processing of environmental state perception information and user behavior perception information from the target area.

[0157] Specifically, the dual-branch feature extractor includes a parallel environment encoding branch and a behavior encoding branch. The environment encoding branch can be designed as a multilayer perceptron (MLP) or a one-dimensional convolutional neural network (1D-CNN) to process structured environmental sensor time-series data. The behavior encoding branch can be designed as a recurrent neural network (RNN, such as LSTM or GRU) or a Transformer encoder to process user behavior sequence data (such as action sequences and embedded vectors of instruction text). Both branches are connected to fully connected layers at their ends, unifying the output dimension to d_model (e.g., 128 dimensions) and normalizing it so that the extracted environmental state features and behavioral preference features are located in a comparable vector space.

[0158] The feature space mapping model can be the encoder part of an encoder-decoder structure or a standalone deep autoencoder. In one configuration, it consists of 2-4 fully connected layers, with intermediate layers using the ReLU activation function. The intermediate layers nonlinearly map the d_model-dimensional environmental state features to another d_model-dimensional (or smaller) "environmental state mapping feature" space, whose features are considered more robust and relevant to control decisions. The control pattern inference model can be a multilayer perceptron regression model. Its input layer receives a concatenated vector: [control semantic representation; environmental state mapping features], where the control semantic representation is fused from user behavior perception information features and the code for the control pattern to be determined (e.g., through another small attention network). After passing through 2-3 hidden layers, the final output layer directly generates the LED lighting operating state parameters (such as illuminance value, color temperature value, etc.), and the output layer can use a linear activation function. Example of model training parameters: The optimizer uses Adam, the initial learning rate is set to 1e-4 to 1e-3, the batch size is set to 32 or 64 depending on the amount of data, and the training epochs are determined based on the convergence of the validation set loss, typically 50-200 epochs. During training, learning rate decay and early stopping strategies can be used to prevent overfitting.

[0159] In this embodiment of the invention, the step of performing feature extraction operations on the environmental state record and the user behavior record respectively through the original dual-branch feature extractor to obtain the environmental state features of the scene behavior associated sample of the environmental state record and the behavior preference features of the scene behavior associated sample of the user behavior record can be implemented through the following example.

[0160] When the environmental state record included in the scene behavior association sample belongs to the overall state of the LED lighting target area, the environmental state feature is extracted for at least one LED lighting scene state sampling unit in the overall state of the LED lighting target area by the original dual-branch feature extractor, and the scene behavior association sample environmental state feature of the overall state of the LED lighting target area is obtained according to the environmental state feature of the at least one LED lighting scene state sampling unit.

[0161] The original dual-branch feature extractor extracts behavioral preference features from the user behavior records included in the scene behavior association samples, thereby obtaining the scene behavior association sample behavioral preference features of the user behavior records.

[0162] In an embodiment of the invention, for example, the server is training a dual-branch feature extractor. It acquires a scene behavior association sample whose environmental state record describes the overall state of a "smart meeting room" at the start of a project debriefing meeting at 9:00 AM on a Monday. This record is a complex collection containing synchronized readings from multiple locations and types of sensors within the meeting room, collectively constituting the overall state of the LED-lit target area.

[0163] The server recognized that directly concatenating all the raw data from the sensors into a large, flat vector input to the environmental feature encoding branch would ignore spatial structure information and make it difficult for the model to focus on the local illumination conditions that are crucial for lighting control. Therefore, the server adopted a structured processing method.

[0164] First, based on the conference room's physical layout and lighting design, the server logically divides it into multiple LED lighting scene state sampling units. For example, these units include:

[0165] Unit A (Conference Table Core Area): Corresponds to the main lighting array above the long conference table. Relevant sensor data includes: readings from three illuminance sensors directly above this area (450, 460, and 440 lux respectively), a reading from a suspended color temperature sensor (5000K), and a signal from a wide-angle infrared sensor below this area (detecting multiple people sitting around).

[0166] Unit B (Lecture Hall and Projection Screen Area): Corresponds to the presentation area at the front of the conference room. Relevant data includes: the illuminance sensor reading in front of the screen (automatically adjusted to 80 lux to prevent glare), the color temperature sensor reading for the lectern's accent lighting (4800K), and the status signal of "someone standing on the lectern" detected by the camera.

[0167] Unit C (Surrounding Environment and Entrance Area): Corresponds to the walls, cabinets, and doorway area of ​​the conference room. Relevant data includes: wall washer light illuminance sensor readings (150 lux), ambient color temperature sensor readings (4000K), and the "door closed" status displayed by the door magnetic sensor.

[0168] The server extracts features from the data of each sampling unit using the environmental feature encoding branch of the original dual-branch feature extractor. Specifically, for unit A, the branch receives its corresponding illuminance, color temperature, human body signal, and other data. After calculation by the internal neural network layer, it outputs a local environmental state feature vector representing the state of that unit, denoted as Fe_A (e.g., a 64-dimensional vector). Similarly, the branch processes the data of units B and C respectively, obtaining local feature vectors Fe_B and Fe_C.

[0169] Next, the server needs to aggregate these local features to represent the overall state of the entire conference room. The server uses a feature aggregation layer (e.g., an attention mechanism layer or a pooling layer) to process Fe_A, Fe_B, and Fe_C. This aggregation layer learns to assign different weights to the features of different units. For example, in the current sample context of a "project debriefing meeting," unit A (the core area of ​​the conference table) is the core area for people to gather and discuss, and its lighting state is crucial to the meeting atmosphere; unit B (the presentation area) is also extremely important when someone is giving a presentation; while unit C (the surrounding area) mainly plays a supporting role in creating the environment. The aggregation layer may assign higher weights to Fe_A and Fe_B, and lower weights to Fe_C, and then fuse the weighted features (e.g., weighted summation or concatenation) to finally generate a unified scene behavior-related sample environment state feature that represents the overall state of the LED lighting target area, denoted as Fe_global. The information contained in this Fe_global feature vector is: "The core discussion area and presentation area of ​​the conference room are brightly lit with a cool color temperature, while the surrounding auxiliary light is moderate, creating an overall focused and formal collaborative lighting environment for multiple people."

[0170] Meanwhile, the server processes user behavior records from the same associated sample. These records contain user behavior data captured from the conference room system within the same time context (Monday at 9:00 AM, when the meeting began):

[0171] Meeting system log: The currently active meeting mode is "Project Review", the audio system is in "Omnidirectional Microphone Discussion Mode", and the projector is turned on and displaying charts.

[0172] Aggregated behavioral signals: Information fused from multiple sensors indicates that the current activity type is "multi-person meeting", the voice activity level is "medium-high" (multiple people are detected speaking alternately), and the main activity areas are the "meeting table area" and the "lectern area".

[0173] User instruction: History records show that the meeting initiator started the "Standard Meeting Mode" with a single click on the tablet at 08:55.

[0174] The server inputs all this multimodal behavioral recording data into the behavioral feature encoding branch of the original dual-branch feature extractor. This branch is designed to handle data that mixes discrete states (such as meeting patterns), continuous signals (such as speech activity), and text labels (such as activity type). The branch may contain structures such as embedding layers, recurrent neural networks, or Transformer encoders to understand and encode the semantics of behavioral sequences and states.

[0175] After branch calculations, a fixed-dimensional scene behavior association sample behavior preference feature vector, denoted as Fb, is finally output. This Fb feature vector aims to capture the collective intent and preferences behind user behavior. For the current sample, the semantics encoded by Fb might be: "The user group is conducting a formal, collaborative, discussion- and presentation-oriented working meeting, preferring a high-definition, low-interference lighting environment to support screen viewing and face-to-face communication."

[0176] At this point, the server has extracted two key features from a single training sample using a structured approach: Fe_global (overall environmental state features) and Fb (user behavior preference features). In subsequent training steps, the server will utilize a large number of such sample pairs (Fe_global, Fb) to adjust the parameters of the dual-branch feature extractor through optimization algorithms. This ensures that for each real-world scene behavior-related sample, the extracted Fe_global and Fb representations in the feature space are highly correlated. In this way, the environmental branch learns how to extract semantic features of the lighting scene closely related to the collective behavioral intentions of users from complex, spatially distributed environmental sensor data; the behavioral branch learns how to extract preference features indicating clear needs for the lighting environment from diverse user interactions and system state data. This lays a solid model foundation for ultimately achieving precise, habit-driven lighting scene control.

[0177] In this embodiment of the invention, the multiple categories include environmental state perception categories and user behavior expression categories; the perception information of the multiple categories includes environmental state perception information and user behavior perception information.

[0178] The method of inferring the control mode based on the at least one undetermined control mode and the multiple categories of perception information to obtain the LED lighting operation status parameters for the LED lighting target area can be implemented through the following example.

[0179] The user behavior perception information and the at least one pending control mode are jointly controlled as input to construct a control semantic representation for control mode inference.

[0180] Based on the control semantic representation and the environmental state characteristics corresponding to the environmental state perception information, control mode inference is performed to obtain the LED lighting operation state parameters for the LED lighting target area.

[0181] In this embodiment of the invention, for example, the server has completed the preliminary processing: for the LED lighting target area of ​​the study, it extracts environmental state perception information (cloudy afternoon, desktop illuminance 200 lux, color temperature 4000K) and user behavior perception information (user's sitting posture, continuous book turning / writing sound). By comparing with a user habit pattern benchmark library, the server has determined at least one pending control mode, such as "deep reading comfortable lighting mode".

[0182] The server now needs to generate the final, precise LED lighting operating status parameters down to specific parameter values. It first performs deep semantic fusion of the control intent.

[0183] The server uses real-time user behavior perception information. After feature extraction, the core semantics of this information have been identified as "the user is engaged in deep reading or writing." Simultaneously, the server uses at least one pending control mode generated in the previous step—a "comfortable lighting mode for deep reading." This mode is a relatively abstract strategy template matched from a habit library, which may contain some guiding principles, such as: "Provide uniform, glare-free desktop lighting, with a warm color temperature to alleviate visual fatigue, and automatic illuminance compensation based on ambient light."

[0184] The server's task is to combine specific, real-time user behavior semantics with relatively abstract, pending control pattern templates to construct a richer, more targeted control semantic representation. This is accomplished through a dedicated semantic fusion module. This module concatenates or performs cross-attention calculations on behavioral feature vectors (representing "working at a desk, focused, sustained") and encoded vectors of pending control patterns (representing "deep reading, comfortable, eye protection").

[0185] The fused output is a new, structured control semantic representation vector. This vector no longer merely describes "what happened" or "what should probably be done," but explicitly expresses the composite instruction: "Given the observed deep focus on reading, the system should execute a deep reading lighting mode with eye comfort as the core objective." In this representation, the user's real-time "focused" state reinforces the need for "undisturbed" lighting, while the "continuous" behavior emphasizes "eye protection" (such as appropriate color temperature). This control semantic representation provides high-level, strongly constrained intent guidance for the final parameter inference.

[0186] Having a clear control semantic representation, the server needs to "translate" it into specific parameter values ​​that can be executed on physical lighting fixtures. This translation process must fully consider the current, specific environmental conditions, because the same intention of "deep reading comfort lighting" requires different absolute illuminance values ​​on a "sunny afternoon" and a "cloudy afternoon."

[0187] The server extracts an environmental state feature vector (representing "cloudy day, weak natural light, current illuminance 200 lux, color temperature 4000K") from the current environmental state perception information. This feature describes the objective physical conditions that the lighting system needs to adapt to and compensate for.

[0188] Subsequently, the server inputs the control semantic representation and environmental state features into a pre-trained control pattern inference model (such as a deep neural network regression model). The role of this model is to calculate the optimal, executable set of parameters based on the high-level control intent and the low-level environmental status.

[0189] The model performs internal calculations. It understands that the core of the control semantics is "deep reading" and "comfort," thus excluding schemes with high color temperatures (such as 6000K) or high dynamic contrast. Simultaneously, the model receives environmental features indicating "insufficient natural light, and the existing lighting environment is dim and has a neutral color temperature." Combining these two factors, the model makes the following inferences:

[0190] Illuminance Parameter: To achieve the illuminance standard for "comfortable reading" (e.g., national standards recommend 300-500 lux) and compensate for insufficient natural light on cloudy days, the model decides to increase the illuminance from the existing 200 lux. However, it doesn't simply set it to 500 lux, as an excessively high, sudden increase in illuminance could cause discomfort. The model may refer to the user's historical adaptation curve and output a gradual increase to 320 lux. This value satisfies the basic reading needs while also considering a smooth transition from the current environment to the target environment.

[0191] Color temperature parameter: To enhance "comfort" and "eye protection," the model needs to adjust the color temperature from the current 4000K towards a warmer direction. However, directly adjusting it to 2700K might appear too yellowish in cloudy environments. The model weighs the control semantics (warmth) against the ambient tone (whiteness in cloudy conditions) and infers an optimized value of 3800K. This color temperature is slightly warmer than the current one, increasing the sense of warmth and reducing potential blue light effects, but without causing a severe color temperature conflict with the ambient light in cloudy conditions.

[0192] Lighting distribution parameters: The "deep reading" in the control semantics implies a high level of user concentration. Therefore, the model may further infer that the brightness ratio of ambient background light (such as bookshelf lights) needs to be slightly reduced, while ensuring that the beam angle of the main desktop light is concentrated to reduce interference from the surrounding field of vision.

[0193] Finally, the model outputs a complete set of specific LED lighting operation status parameters, such as: {"Main lamp illuminance target value": 320 lux, "Main lamp color temperature target value": 3800K, "Background light brightness ratio": 30%, "Transition time": 5 seconds}.

[0194] The server sends these parameters to the gateway or controller of the study lighting system. The controller then drives the smart lights on the desktop, smoothly adjusting the light level from 200 lux, 4000K to 320 lux, 3800K within 5 seconds, while dimming the surrounding ambient lights. Thus, the system not only responds to the user's real-time focused reading behavior but also creates a truly personalized, comfortable, and scene-appropriate lighting environment based on the habitual pattern of "comfortable lighting for deep reading" and the specific environmental conditions of a cloudy afternoon. The entire inference process represents a precise combination of high-level intent and underlying physical reality.

[0195] In this embodiment of the invention, the step of inferring the control mode based on the environmental state features corresponding to the control semantic representation and the environmental state perception information to obtain the LED lighting operation state parameters for the LED lighting target area can be implemented through the following example.

[0196] The environmental state features corresponding to the environmental state perception information are obtained by performing feature mapping operation on the environmental state features through the feature space mapping model.

[0197] The control mode inference model infers the control mode based on the control semantic representation and the environmental state mapping features to obtain the LED lighting operation state parameters for the LED lighting target area.

[0198] In this embodiment of the invention, for example, the server has constructed a control semantic representation (deep reading comfort intent) for the study room scenario and holds the original environmental state features (vectors containing information such as "cloudy day, 200 lux, 4000K") extracted from the current environmental perception information. Before making the final inference, the server first optimizes the original environmental features.

[0199] The server invokes a pre-trained feature space mapping model. This model is a neural network whose function is to map the raw environmental features extracted directly from sensors, which may contain redundancy or dimensional differences, to a purer feature subspace that is more relevant to lighting control decisions. The model receives the raw environmental feature vector, performs an internal nonlinear transformation, and outputs a new vector of the same or different dimensions, i.e., the environmental state mapping feature.

[0200] This mapping process purifies and focuses the information. For example, the original features "cloudy" and "illuminance 200 lux" are highly collinear (because cloudy weather leads to low illuminance). The mapping model learns this correlation and represents the core fact of "severely insufficient natural light" in a more compact way in the mapped features, while potentially weakening or integrating other minor fluctuations. The mapped features more directly reflect the strength of the environment's need for lighting compensation, providing standardized input for the accurate calculation of subsequent control variables.

[0201] The server inputs the refined environmental state mapping features obtained in the previous step, along with the previously constructed control semantic representation representing the user's intent, into another pre-trained control pattern inference model.

[0202] The inference model is a multi-output regressive neural network. It receives two inputs: one indicating "what to do" (control semantics), and the other quantifying "how the current basis is" (environmental mapping features). Internally, the model calculates the optimal control parameter values ​​required in the current environment to achieve this intention based on the complex mapping relationships learned during training.

[0203] The model performs calculations. It understands the control semantic requirement of "comfortable reading," which corresponds to an illuminance range and a warmer color temperature direction. Simultaneously, it reads the degree of "severe insufficiency of natural light" quantified by environmental mapping features. The model adds this insufficiency to the basic illuminance requirement for "comfortable reading," and, referring to color temperature coordination principles, ultimately outputs a set of precise LED lighting operating parameters: increasing the desktop main light illuminance to 320 lux and adjusting the color temperature to 3800K.

[0204] The server then sends out these specific parameters for execution, completing the precise conversion from the abstract intent and original environment to the specific light environment.

[0205] In this embodiment of the invention, the feature space mapping model and the control mode inference model are obtained through the following process, and can be implemented through the following examples.

[0206] Based on the first sample environmental state features of the first sample LED lighting target area and the sample environmental state information of the first sample LED lighting target area, a staged feature space mapping model is trained; the first sample environmental state features are extracted from the environmental state perception information of the first sample LED lighting target area.

[0207] The second sample environmental state features, sample user behavior perception information, and sample control mode of the second sample LED lighting target area are obtained; the second sample environmental state features are extracted from the environmental state perception information of the second sample LED lighting target area.

[0208] The second sample environment state features are obtained by performing feature mapping operation on the second sample environment state features through the stage feature space mapping model.

[0209] The original control mode inference model infers the control mode based on the sample user behavior perception information and the second sample environmental state mapping features to obtain the predictive control mode for the second sample LED lighting target area;

[0210] Based on the predictive control mode and the sample control mode, the staged feature space mapping model and the original control mode inference model are optimized and integrated for training to obtain the trained feature space mapping model and control mode inference model.

[0211] In this embodiment of the invention, exemplarily, the server initiates the model training process. First, it needs to train a mapping model capable of understanding the essential characteristics of environmental states. To this end, the server collects a large amount of data on the target area of ​​the first sample LED lighting. For example, this data comes from the living rooms of hundreds of different homes, containing a massive amount of first sample environmental state features. These feature vectors are extracted from the raw sensor data (such as illuminance, color temperature, time, human presence, etc.) by the environmental branch (or its predecessor) of a dual-branch feature extractor, with each feature vector representing an environmental snapshot of a living room at a certain moment.

[0212] At the same time, the server possesses more original sample environmental state information corresponding to these environmental state characteristics. This information can be the sensor readings themselves at the same moment, or it can be descriptive labels that have been simply aggregated (such as "bright", "dim", "warm tone", "no one").

[0213] The server employs a self-supervised learning paradigm to train the original feature space mapping model. Specifically, the server inputs the first sample environmental state features into the original model, which outputs a first sample environmental state mapping feature. Then, the server allows an auxiliary original control mode inference model (used only as a decoder or reconstructor at this point) to perform "environment reconstruction" or "environment description" on this mapping feature, outputting a model control state representation (e.g., a predicted sensor reading distribution or environmental label).

[0214] The training objective is to compare this "model-controlled state representation" with the actual "sample environmental state information" and calculate the reconstruction error or classification error. For example, if the model predicts "the current illuminance is approximately 300 lux" based on the mapping features, while the actual value is 320 lux, an error occurs. The server uses a backpropagation algorithm to fine-tune only the parameters of the original feature space mapping model, aiming to ensure that the features extracted by the mapping model can reconstruct the original environmental information as accurately as possible through the decoder. After multiple rounds of iterative training, the server obtains a staged feature space mapping model. This model has learned to remove noise from the original environmental features and extract information-dense representations that are crucial to describing the essence of the environmental state (such as "overall brightness level," "light color atmosphere," and "space occupancy").

[0215] Next, the server needs to collect supervised data for training the final control task. This data comes from a second sample of LED lighting target areas, such as a group of office workstations deployed with a complete smart lighting system. For each workstation's record in a given scenario, the server obtains three pieces of information:

[0216] The second sample environmental state features are also obtained by the feature extractor, representing the initial environmental state of the workstation (e.g., [desktop illuminance: 200 lux, color temperature: 6000 K, time: afternoon]).

[0217] Sample user behavior perception information: User behavior data recorded in this scenario (such as [posture: leaning forward, application status: programming IDE full screen, high frequency of mouse and keyboard activity]).

[0218] Sample control mode: This is the actual optimal control result, that is, the final executed lighting operation state parameters (such as {target illuminance: 500 lux, target color temperature: 5000 K}) determined by the expert system or historical best records under this environment and behavior.

[0219] The server begins ensemble training. For a single data point in the second sample set, the server first invokes the staged feature space mapping model to process the environmental state features of the second sample, obtaining the environmental state mapping features of the second sample. These features are already more representative than the original features.

[0220] Next, the server inputs the sample user behavior perception information (or its extracted behavioral features) along with the second sample environmental state mapping features into an initial control mode inference model. This inference model is now used as the actual controller, and its structure is designed to output lighting parameters. The model calculates based on the input and outputs a predictive control mode, i.e., the lighting parameters it believes should be executed.

[0221] The server then compares the predicted control pattern output by the model with the actual sample control pattern (true parameters) and calculates the difference between them as a loss (such as mean squared error). This loss measures the decision accuracy of the current concatenated mapping model and inference model.

[0222] A crucial step is that the server simultaneously tunes the parameters of both the staged feature space mapping model and the original control mode inference model based on this loss. This means that the backpropagation gradient will flow through both models at the same time. By updating the parameters through optimization algorithms, the goal is to make the output of the entire cascaded system (environmental features, mapping model, mapping features, inference model, prediction parameters) as close as possible to the true optimal control parameters.

[0223] The server performs multiple rounds of iterative training on the second sample set. During this process:

[0224] The control pattern inference model has learned how to accurately calculate lighting parameters based on user behavioral intentions and the refined environmental state.

[0225] The training objective of the feature space mapping model has undergone a subtle but crucial shift: it is no longer simply about perfectly reconstructing environmental information, but rather tuned to extract the environmental feature representations most helpful for the final control decision. For example, it will place greater emphasis on those key dimensions that affect illuminance compensation calculations, while downplaying some environmental details that are irrelevant to control.

[0226] When the training loss converges to a satisfactory level, the server obtains the trained feature space mapping model and control pattern inference model. These two models are co-optimized; the mapping model provides the inference model with an optimal "environment summary," and the inference model uses this summary and user behavior information to make accurate decisions.

[0227] In this embodiment of the invention, the staged feature space mapping model trained based on the first sample environmental state features of the first sample LED lighting target area and the sample environmental state information of the first sample LED lighting target area can be implemented through the following example.

[0228] Acquire the first sample environmental state characteristics of the first sample LED lighting target area and the sample environmental state information for the first sample LED lighting target area;

[0229] The first sample environment state features are obtained by performing feature mapping operation on the first sample environment state features through the original feature space mapping model.

[0230] The original control mode inference model describes the LED lighting target area by mapping the environmental state features of the first sample to the LED lighting target area, and obtains the model control state representation for the LED lighting target area of ​​the first sample.

[0231] Based on the model control state representation and the sample environment state information, the original feature space mapping model is optimized and iteratively trained to obtain a staged feature space mapping model.

[0232] In an embodiment of the invention, exemplarily, the server initiates the training process of the phased feature space mapping model. It retrieves training samples for self-supervised learning from a historical database. These samples originate from a first sample LED lighting target area, such as a large number of smart meeting rooms with different configurations. For each meeting room's record at a given moment, the server acquires two parts of data:

[0233] The first sample environmental state features: These are feature vectors extracted from the raw sensor data in the conference room by the feature extraction network. For example, a feature vector may represent the compressed encoding of information such as "100 lux illuminance in the conference table area, 5000 K color temperature, 80 lux illuminance in the projection screen area, multiple people detected, and 55 dB ambient noise".

[0234] Sample environment state information: This is descriptive data that corresponds to the feature vectors mentioned above and is closer to the original perception. It can be a set of key sensor readings at the same time (e.g., {main illuminance: 450, color temperature: 5000, screen illuminance: 80, number of people: 5}), or it can be a simple environment label based on rules (e.g., “multi-person meeting in progress_bright”).

[0235] The server initializes a primitive feature space mapping model and a primitive control mode inference model. At this point, the parameters of both models are randomly initialized.

[0236] The server inputs the first sample's environmental state features into the original feature space mapping model. This model performs a non-linear transformation on these input features, outputting a first sample's environmental state mapping feature. Initially, this mapping feature is random and meaningless.

[0237] Next, the server inputs this mapped feature into the original control mode inference model. At this stage, the inference model is temporarily assigned a specific task: to act as an "environmental description decoder." Its goal is not to predict control parameters, but to attempt to reconstruct or describe the original environmental state based on the mapped feature. The model outputs a model control state representation that attempts to approximate the sample environmental state information in the sample. For example, it might output a set of predicted sensor values ​​{main illuminance: 420, color temperature: 5100, screen illuminance: 75, number of people: 4}, or an environmental label "meeting_generally bright".

[0238] The server compares the model's control state representation with the actual sample environment state information and calculates the difference between the two, i.e., the reconstruction error. For example, it calculates the squared difference between the predicted illuminance of 450 and the actual illuminance of 420, or it calculates the classification cross-entropy between the predicted label and the actual label.

[0239] Subsequently, the server executes the backpropagation algorithm. The key point is that the server only uses this error to update (optimize) the parameters of the original feature space mapping model, while keeping the parameters of the original control mode inference model frozen. The optimization direction is to modify the parameters of the mapping model so that the mapped features it generates can be more accurately reconstructed into the original environmental information by the subsequent fixed decoder (inference model).

[0240] The server repeats this process on a large amount of initial sample data: input environment features -> mapping model generates mapping features -> decoder reconstructs environment information -> calculates reconstruction error -> backpropagation updates the mapping model. After multiple rounds of iterative training, the mapping model gradually learns to extract the most essential and informative parts of the environment features, while discarding redundancy and noise. Because the features it extracts must be able to reconstruct the environment well through a fixed, limited-capacity decoder, it is forced to learn an efficient and general environment representation.

[0241] Finally, when the reconstruction error converges to a low level, the server obtains a staged feature space mapping model that has been trained. This model has the ability to map chaotic environmental perception features to a well-structured and information-condensed feature space, laying a solid foundation for subsequent joint supervised training with the control inference model.

[0242] In this embodiment of the invention, the step of determining at least one pending control mode for the LED lighting target area based on the multi-source mode identifiers of the various types of perception information can be implemented through the following example.

[0243] The multi-source pattern identifiers of the various categories of perceptual information are summarized to obtain the pattern integration result;

[0244] Based on the mode integration results, at least one pending control mode is determined for the target LED lighting area.

[0245] In this embodiment of the invention, for example, the server has matched multi-source pattern identifiers to the environmental state perception information and user behavior perception information of the study room scene, respectively. The environmental information identifiers are "leisure reading environment" and "deep reading preference", and the behavior information identifiers are "deep reading preference" and "leisure reading environment".

[0246] The server initiates a pattern aggregation process. It treats these identifiers from different data sources and categories as a set to be integrated. The server first performs identifier alignment and deduplication, discovering that "deep reading preference" and "leisure reading environment" appear repeatedly in both sets. This indicates a high degree of consistency between judgments derived from two independent channels: environment and behavior, reinforcing the confidence of these two patterns.

[0247] Next, the server resolves conflicts and assigns weights. In this case, there are no conflicts. The server assigns initial weights to each identifier based on its source reliability (e.g., behavioral identifiers are usually more direct about intent) and historical frequency. For example, "deep reading preference" receives the highest weight because it is directly derived from behavioral data and highly consistent with the current action; "leisure reading environment," as a direct description of the environmental state, receives the next highest weight. The server integrates all identifiers and their weights to form a structured pattern integration result. This result clearly indicates that the core feature of the current scenario is that the "deep reading" behavior occurs in a "leisure reading" lighting environment, and the two corroborate each other.

[0248] Based on the above pattern integration results, the server queries the "User Habit Pattern Baseline Library" for predefined, more specific lighting control pattern templates. These templates are a mapping from pattern identifiers to preliminary control strategies.

[0249] In one specific implementation, the construction of the user habit pattern benchmark library can be achieved through the following semi-automatic process:

[0250] 1. Data Collection and Preprocessing: During the initial phase or daily operation, the system continuously collects historical sensing information data streams from one or more standard LED lighting target areas, including environmental status sensing information (such as illumination, temperature, and timestamps) and user behavior sensing information (such as sensor-triggered events, voice command text, and device interaction logs). The collected raw data is cleaned, denoised, normalized, and time-series aligned.

[0251] 2. Feature Extraction and Scene Slicing: The dual-branch feature extractor (or its pre-trained initialization network) is used to extract environmental state features and behavioral preference features from the preprocessed historical data. Simultaneously, based on time windows or event markers (such as "movie started" or "reading ended"), the continuous data stream is sliced ​​into independent "scene segments".

[0252] 3. Clustering Analysis and Pattern Discovery: The extracted feature vectors (environmental features and / or behavioral features) are input into an unsupervised clustering algorithm (e.g., K-Means, DBSCAN, or Gaussian mixture model) for analysis. The algorithm automatically aggregates scene fragments with similar features into several categories. The feature vector of the center point of each category forms the initial candidate standard perceptual features.

[0253] 4. Expert Annotation and Label Generation: The system provides an interactive interface for administrators or users to review and annotate the categories generated by the above clustering. Each category can be assigned a semantically clear standard pattern label (such as "late-night rest mode" or "afternoon focused work preference"). The annotation process can be confirmed by combining representative raw data of the category (such as average illumination value and typical user actions).

[0254] 5. Database Generation and Updates: The labeled pairs (candidate standard perceptual features, standard pattern identifiers) are stored in the database to form an initial user habit pattern benchmark database. In subsequent operation, the system can periodically use newly generated, high-confidence scene data and their recognition results as new samples, repeating steps 2-4 to incrementally update and optimize the benchmark database, achieving self-learning and expansion. Through this process, the benchmark database not only contains predefined patterns but can also adaptively learn and incorporate habit patterns specific to different users or environments.

[0255] The server discovered that the highly weighted "deep reading preference" tag in the integration results was directly associated with a control template called "Deep Reading Comfort Lighting Mode". This template contains a general description of lighting principles for this type of scenario, such as: "Ensure sufficient illuminance in the center area of ​​the desktop (reference value 300-500 lux), warm color temperature (3700-4200K), and reduce ambient light interference."

[0256] Since the integration results are clear and consistent, the server directly identifies this template as at least one pending control mode for the current study area. This pending mode is not a specific set of parameters, but rather a strategy framework that indicates the direction and general scope of control, awaiting final parameter inference based on precise environmental state information. If the integration results have multiple possible directions, the server may generate multiple pending modes for subsequent steps to select the best or merge.

[0257] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned LED lighting scene control method that integrates user habit recognition. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0258] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method of controlling an LED lighting scene with fusion of user habit recognition, characterized by, The method includes: The target area of ​​LED lighting to be processed is determined, and various types of perception information are extracted from the target area of ​​LED lighting. A user habit pattern benchmark library is established; the user habit pattern benchmark library includes standard perception features of standard LED lighting target areas under the various categories and standard pattern identifiers corresponding to each standard perception feature; Based on the consistency comparison results between the multiple categories of perception information and the standard perception features of the standard LED lighting target area under the multiple categories, the multi-source mode identifiers of the multiple categories of perception information are obtained from the user habit pattern benchmark library. Based on the multi-source mode identifiers of the various types of perception information, at least one undetermined control mode is determined for the LED lighting target area; Based on the at least one undetermined control mode and the multiple types of perception information, control mode inference is performed to obtain LED lighting operation status parameters for the LED lighting target area.

2. The method of claim 1, wherein, The step of matching the multi-source pattern identifiers of the various categories of perceptual information with the standard perceptual features of the standard LED lighting target area under the various categories based on the consistency comparison results includes: Each type of perception information in the multiple categories is compared with the standard perception features of the standard LED lighting target area under the multiple categories to obtain the consistency comparison result of each type of perception information. From the consistency comparison results of each type of sensing information, determine the matching comparison results of each type of sensing information. Based on the standard perceptual features corresponding to the adaptation ratio results of each type of perceptual information, the multi-source pattern identifiers of each type of perceptual information are obtained by matching from the user habit pattern benchmark library.

3. The method of claim 1, wherein, The multiple categories include environmental state perception categories and user behavior expression categories; the perception information of the multiple categories includes environmental state perception information and user behavior perception information of the LED lighting target area; the user habit pattern benchmark library includes standard environmental perception features of the standard LED lighting target area under the environmental state perception category and environmental state pattern identifiers corresponding to the standard environmental perception features, as well as standard behavior perception features of the standard LED lighting target area under the user behavior expression category and behavior preference pattern identifiers corresponding to the standard behavior perception features; The step of matching the multi-source pattern identifiers of the various categories of perceptual information with the standard perceptual features of the standard LED lighting target area under the various categories based on the consistency comparison results includes: Based on the environmental state perception information and the user behavior perception information, feature extraction operations are performed respectively to obtain environmental state features and behavioral preference features; The environmental state features are compared with the standard environmental perception features and the standard behavioral perception features to determine environmental consistency. Based on the environmental consistency comparison results, the multi-source pattern identifier of the environmental state perception information is obtained from the environmental state pattern identifier and the behavioral preference pattern identifier. The behavioral preference features are compared with the standard environmental perception features and the standard behavioral perception features for text consistency judgment. Based on the text consistency comparison results, the multi-source pattern identifier of the user behavior perception information is obtained from the environmental state pattern identifier and the behavioral preference pattern identifier.

4. The method according to claim 3, characterized in that, The step of performing environmental consistency judgment between the environmental state features and the standard environmental perception features and the standard behavioral perception features, and obtaining the multi-source pattern identifier of the environmental state perception information from the environmental state pattern identifier and the behavioral preference pattern identifier based on the environmental consistency comparison result, includes: The environmental state features are compared with the standard environmental perception features to determine environmental consistency, and a first environmental consistency comparison result is obtained. Based on the standard environmental perception features corresponding to the first environmental consistency comparison result of the adaptation, the first environmental state perception category determination identifier is determined from the environmental state mode identifier; The environmental state features are compared with the standard behavioral perception features to determine environmental consistency, and a second environmental consistency comparison result is obtained. Based on the standard behavior perception features corresponding to the second environment consistency comparison results, a first user behavior expression category determination identifier is determined from the behavior preference pattern identifier; Based on the first environmental state perception category determination identifier and the first user behavior expression category determination identifier, the multi-source pattern identifier of the environmental state perception information is obtained.

5. The method according to claim 3, characterized in that, The step of performing text consistency judgment between the behavioral preference features and the standard environment perception features and the standard behavioral perception features, and obtaining the multi-source pattern identifier of the user behavior perception information from the environment state pattern identifier and the behavioral preference pattern identifier based on the text consistency comparison result, includes: The behavioral preference features are compared with the standard environmental perception features to determine text consistency, and a first text consistency comparison result is obtained. Based on the standard environmental perception features corresponding to the first text consistency comparison result, a second environmental state perception category determination identifier is determined from the environmental state pattern identifier. The behavioral preference features are compared with the standard behavioral perception features to determine text consistency, and a second text consistency comparison result is obtained. Based on the standard behavior perception features corresponding to the adapted second text consistency comparison result, a second user behavior expression category determination identifier is determined from the behavior preference pattern identifier; Based on the second environmental state perception category determination identifier and the second user behavior expression category determination identifier, the multi-source pattern identifier of the user behavior perception information is obtained.

6. The method according to claim 3, characterized in that, The step of performing feature extraction operations based on the environmental state perception information and the user behavior perception information to obtain environmental state features and behavioral preference features is implemented through a dual-branch feature extractor. The training steps of the dual-branch feature extractor include: Obtain scene behavior association samples, which include environmental state records and user behavior records in response to the environmental state records; When the environmental state record included in the scene behavior association sample belongs to the overall state of the LED lighting target area, the environmental state feature is extracted for at least one LED lighting scene state sampling unit in the overall state of the LED lighting target area by the original dual-branch feature extractor, and the scene behavior association sample environmental state feature of the overall state of the LED lighting target area is obtained according to the environmental state feature of the at least one LED lighting scene state sampling unit. The original dual-branch feature extractor extracts behavioral preference features from the user behavior records included in the scene behavior association samples, thereby obtaining the scene behavior association sample behavioral preference features of the user behavior records. Based on the environmental state features and behavioral preference features of the scene behavior-related samples, the original dual-branch feature extractor is optimized and iterative training is performed to obtain the dual-branch feature extractor.

7. The method according to claim 1, characterized in that, The multiple categories include environmental state perception category and user behavior expression category; the perception information of the multiple categories includes environmental state perception information and user behavior perception information; The control mode inference based on the at least one undetermined control mode and the multiple categories of sensing information, to obtain LED lighting operation status parameters for the LED lighting target area, includes: The user behavior perception information and the at least one pending control mode are jointly controlled as input to construct a control semantic representation for control mode inference. The environmental state features corresponding to the environmental state perception information are obtained by performing feature mapping operation on the environmental state features through the feature space mapping model. The control mode inference model infers the control mode based on the control semantic representation and the environmental state mapping features to obtain the LED lighting operation state parameters for the LED lighting target area.

8. The method according to claim 7, characterized in that, The feature space mapping model and the control mode inference model are obtained through the following process, including: Based on the first sample environmental state features of the first sample LED lighting target area and the sample environmental state information of the first sample LED lighting target area, a staged feature space mapping model is trained; the first sample environmental state features are extracted from the environmental state perception information of the first sample LED lighting target area. The second sample environmental state features, sample user behavior perception information, and sample control mode of the second sample LED lighting target area are obtained; the second sample environmental state features are extracted from the environmental state perception information of the second sample LED lighting target area. The second sample environment state features are obtained by performing feature mapping operation on the second sample environment state features through the stage feature space mapping model. The original control mode inference model infers the control mode based on the sample user behavior perception information and the second sample environmental state mapping features to obtain the predictive control mode for the second sample LED lighting target area; Based on the predictive control mode and the sample control mode, the staged feature space mapping model and the original control mode inference model are optimized and integrated for training to obtain the trained feature space mapping model and control mode inference model.

9. The method according to claim 8, characterized in that, The staged feature space mapping model is trained based on the first sample environmental state features of the first sample LED lighting target area and the sample environmental state information of the first sample LED lighting target area, including: Acquire the first sample environmental state characteristics of the first sample LED lighting target area and the sample environmental state information for the first sample LED lighting target area; The first sample environment state features are obtained by performing feature mapping operation on the first sample environment state features through the original feature space mapping model. The original control mode inference model describes the LED lighting target area by mapping the environmental state features of the first sample to the LED lighting target area, and obtains the model control state representation for the LED lighting target area of ​​the first sample. Based on the model control state representation and the sample environment state information, the original feature space mapping model is optimized and iteratively trained to obtain a staged feature space mapping model.

10. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-9.

Images