Human factor illumination adaptive adjustment method based on multi-biological feature fusion and dynamic learning

Abstract

The invention relates to the technical field of intelligent lighting control and human factor engineering, in particular to a human factor lighting self-adaptive adjusting method based on multi-biological-feature fusion and dynamic learning, and the method comprises the steps: synchronously collecting multi-source biological features of a user, such as electroencephalogram, eye movement, heart rate variability and facial thermal imaging; accurate user state estimation is generated through dynamic weighting of a fusion network based on an attention mechanism. On the basis, the system maintains and continuously updates a personal illumination response model library for each user, and online exploration optimization is carried out while a historical optimal illumination strategy is utilized in combination with a reinforcement learning framework, so that highly personalized illumination adjustment parameters are generated. After adjustment, the model is continuously improved through closed-loop verification, and rapid starting and long-term rhythm adaptation of a new user are achieved through meta-learning. According to the invention, the problems of single perception and lack of individuation and self-evolution ability in the prior art are solved, and intelligent light environment adjustment with comprehensive perception, continuous optimization and deep fitting of a human body is realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent lighting control and human factors engineering, and in particular to a human factors-based adaptive lighting adjustment method based on multi-biological feature fusion and dynamic learning. Background Technology

[0002] With the development of healthy and intelligent lighting, human-centered lighting aims to improve health, comfort, and efficiency by dynamically adjusting the light environment (such as color temperature, illuminance, and spectrum) to match people's physiological rhythms, psychological state, and work needs.

[0003] Existing technologies mostly rely on adjusting a single or simple parameter, for example: Preset modes based on time / scene: Lighting modes switch according to a schedule or fixed scene (such as reading or rest), lacking response to the individual's real-time status.

[0004] Adjustment based on environmental sensors: The illuminance is automatically adjusted according to the ambient light sensor, but the actual physiological and psychological feelings of people are not taken into account.

[0005] Preliminary feedback based on a single biological signal: for example, using heart rate variability (HRV) or skin conductance response (GSR) to roughly determine "stress" or "relaxation" state, and then adjusting the light accordingly. This type of method has significant shortcomings: the signal is singular, easily affected by interference, and the state judgment is coarse; it lacks personalized learning and cannot adapt to the differences in light response among different individuals; the adjustment strategy is static and fixed, and cannot evolve with user habits or long-term rhythm changes.

[0006] Therefore, there is an urgent need for an intelligent lighting adjustment method that can more accurately and comprehensively perceive the user's status and continuously learn and evolve in a personalized way. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, this invention provides a human-caused adaptive lighting adjustment method based on multi-biological feature fusion and dynamic learning, in order to solve the problems existing in the background art.

[0008] This invention provides the following technical solution: a human-caused adaptive lighting adjustment method based on multi-biological feature fusion and dynamic learning, comprising the following steps: S1: Simultaneously collect raw biometric data of at least three different types of target users, and perform preprocessing and feature extraction to obtain time-aligned multimodal biometric vectors; S2: Perform multi-level dynamic feature fusion on the multimodal biometric feature vector to generate a dynamically weighted user state vector, and map it to a quantitative estimate of the user's current multidimensional physiological and psychological state. S3: Generate personalized lighting strategies based on dynamic learning mechanisms: Based on the current user status, query the user's personal lighting response model library to obtain the historical best strategy basis, and introduce an exploration-trade-off mechanism based on reinforcement learning for online optimization to generate the final lighting parameter instructions; the personal lighting response model library dynamically records and updates the user's status, lighting parameters, and the correlation between feedback / performance. S4: Adjust the lighting equipment according to the lighting parameter instructions, and collect biometric data after adjustment for closed-loop verification, and feed back the verification results to update the personal lighting response model library; S5: Based on a meta-learning model, it analyzes common patterns across users and long-term circadian rhythm changes, which can be used for cold start recommendations for new users and fine-tuning of long-term lighting adjustment targets.

[0009] Furthermore, in step S1, the biometric types include at least three of the following: physiological signals reflecting the activity of the autonomic nervous system, EEG signals reflecting the state of the central nervous system, eye movement features reflecting visual fatigue and cognitive load, and facial thermal imaging features.

[0010] Furthermore, in step S2, the multi-level dynamic feature fusion includes: performing primary feature-level fusion on features of the same type; using an attention-based network to perform intermediate decision-level fusion on features of different types, dynamically learning the contribution weight of each feature to the inference of a specific state; and performing high-order state mapping on the fused feature vector to output a quantized state estimate.

[0011] Furthermore, the attention mechanism network can adaptively adjust the feature weight allocation based on environmental context information.

[0012] Furthermore, in step S3, the reinforcement learning framework uses the user's state vector as the state space, the adjustment amount of the lighting parameters as the action space, and the user's subjective feedback or automatic evaluation based on biometric state improvement as the reward signal.

[0013] Furthermore, the reinforcement learning framework employs proximal policy optimization or deep Q-network algorithms.

[0014] Furthermore, step S6 is also included: when the fused biometrics indicate that the user is in a preset abnormal state, the safety protection lighting mode is activated, the light is adjusted to a preset soothing mode and a reminder is issued.

[0015] A system for implementing the method as described in any one of the above, characterized in that it comprises: The multi-biometric sensing module is used to collect raw data of various biological characteristics; The data processing and fusion computing module is used to perform feature extraction, multi-level dynamic feature fusion, and user state mapping. The dynamic learning and policy engine module is used to maintain and update the personal lighting response model library and perform reinforcement learning-based lighting policy generation and optimization. The lighting control execution module is used to control lighting equipment according to lighting strategy instructions; The user interaction and feedback module is used to receive subjective feedback from users. The closed-loop verification module is used to verify the lighting adjustment effect and feed it back to the dynamic learning and strategy engine module.

[0016] Furthermore, it also includes a meta-learning server for performing cross-user commonality analysis and long-term rhythm prediction, and for interacting with the dynamic learning and policy engine module.

[0017] A computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the steps of the method as described in any one of the preceding claims.

[0018] The technical effects and advantages of this invention are as follows: This invention integrates multi-source heterogeneous biological features (physiological + neural + visual) and dynamically focuses on key signals using attention mechanisms, making it more resistant to interference than a single signal and able to more comprehensively and accurately depict the user's complex internal state. The system establishes and continuously updates a personal lighting response model for each user, and optimizes it online through reinforcement learning, enabling the lighting strategy to adapt to the user's unique preferences, habits, and changes over time. This invention not only responds to real-time states but also considers group commonalities and long-term rhythms through meta-learning, achieving multi-timescale human-centric fit from immediate adjustment to periodic adaptation. The system has an "exploration" capability, which can proactively discover better lighting solutions that the user may not have been aware of, possessing the potential for autonomous evolution rather than being limited to presets or historical records. This invention sets up a protection mechanism for abnormal states, improving the system's robustness and practicality. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a mind map for the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, the human-caused adaptive lighting adjustment method based on multi-biometric feature fusion and dynamic learning includes the following steps: S1: Simultaneously collect raw biometric data of at least three different types of target users, and perform preprocessing and feature extraction to obtain time-aligned multimodal biometric vectors; S2: Perform multi-level dynamic feature fusion on the multimodal biometric feature vector to generate a dynamically weighted user state vector, and map it to a quantitative estimate of the user's current multidimensional physiological and psychological state. S3: Generate personalized lighting strategies based on dynamic learning mechanisms: Based on the current user status, query the user's personal lighting response model library to obtain the historical best strategy basis, and introduce an exploration-trade-off mechanism based on reinforcement learning for online optimization to generate the final lighting parameter instructions; the personal lighting response model library dynamically records and updates the user's status, lighting parameters, and the correlation between feedback / performance. S4: Adjust the lighting equipment according to the lighting parameter instructions, and collect biometric data after adjustment for closed-loop verification, and feed back the verification results to update the personal lighting response model library; S5: Based on a meta-learning model, it analyzes common patterns across users and long-term circadian rhythm changes, which can be used for cold start recommendations for new users and fine-tuning of long-term lighting adjustment targets.

[0022] In practice: First, the user's biosignals are collected simultaneously using multiple sensor devices. For example, the user wears an integrated head-mounted device with a built-in single-channel electroencephalogram (EEG) sensor and a miniature eye-tracking camera; simultaneously, a smartwatch is worn on the wrist to collect photoplethysmography (PPG) signals to calculate heart rate and heart rate variability, and to measure electrical skin activity (EDA); a small infrared thermal imager module is also installed above the computer on the desk for non-contact facial thermal imaging. All sensors are time-synchronized with a central processing unit (such as a local computer or edge server) via wireless or wired connections to ensure data timestamp alignment. After receiving the raw data, the central processing unit performs preprocessing: bandpass filtering of the EEG signal from 1-45Hz is applied to remove power frequency interference and baseline drift; peak detection is performed on the PPG signal and the interpeak interval between adjacent peaks is calculated to obtain the time-domain features (such as RMSSD) and frequency-domain features (LF and HF power) of HRV; pupil localization and diameter calculation are performed on the eye-tracking video, and the number of blinks per minute is counted; face region recognition is performed on the thermal imaging image, and the average temperature of the area around the eye socket is extracted. After preprocessing, a data frame is generated every 30 seconds, containing various feature values ​​extracted within that time period, which together constitute a multimodal feature vector.

[0023] Next, state inference is performed. The system runs a pre-trained neural network model to process this multimodal feature vector. The first part of this model is a feature attention fusion layer, which automatically calculates the importance weights of four features at the current moment for judging the two core states of "cognitive focus" and "visual fatigue," namely, the relative power of the EEG beta wave (13-30Hz), the rate of change of pupil diameter, the LF / HF ratio of HRV, and the temperature change in the supraorbital region. These features are then weighted and summed to form a comprehensive state encoding vector. The second part of this vector is a fully connected classification layer, which maps the state encoding vector to specific state probabilities or values. For example, the output might be "cognitive focus: 0.75 (range 0-1)" or "visual fatigue: 0.6." The system uses this as a quantitative estimate of the user's current state.

[0024] Then, the lighting strategy is generated and optimized. The system maintains an independent SQLite database for each user as a personal lighting response model library. The database stores historical records, each containing a timestamp, a state estimate (e.g., focus level, fatigue level), the lighting parameters executed at that time (e.g., color temperature 4000K, illuminance 500lx), and subsequent subjective feedback scores (1-5 points) or performance indicators automatically evaluated by the system (e.g., the improvement in focus level within the following 10 minutes). When a new state estimate is obtained, the system queries the historical records in the database to find the top 5 records where "focus level" and "fatigue level" are closest to the current state (e.g., the smallest Euclidean distance), and selects the lighting parameters corresponding to the record with the highest subjective feedback score as the initial suggested values ​​for this adjustment. A reinforcement learning agent also runs within the system, employing the Proximal Policy Optimization (PPO) algorithm. It uses the current state estimate as the "state" and the adjustment amount of the lighting parameters (e.g., color temperature ±200K, illuminance ±50lx) as the "action." In 90% of cases, the agent adopts the historical best recommendation (i.e., the action is "fine-tuning to the recommended value"). In 10% of cases, the agent will attempt a random but small-scale "exploratory" action, such as further reducing the color temperature by 50K from the recommended value. After executing the action, the new lighting parameters are sent to the smart luminaire.

[0025] After the lighting is adjusted, the system enters the evaluation phase. Over the next 5-10 minutes, the system continuously collects biometric data and calculates state indicators. The average "attention level" during this period is compared to the baseline value before adjustment to obtain the performance change ΔP. Simultaneously, a simple feedback pop-up appears on the user's mobile app: "Was the light comfortable? Please rate it (1-5 stars)." The system then performs a weighted average of the performance change ΔP (normalized) and the user's rating (e.g., 5 stars equals 1.0), serving as the "reward" signal for this adjustment action. This "reward" signal, the pre-adjustment "state," and the taken "action" together constitute a training data point used to update the policy and value networks of the PPO agent. Furthermore, this complete "state-action-reward" record is also stored in the personal lighting response model library, enriching historical data.

[0026] Finally, the system also includes a background meta-learning process. This process periodically (e.g., weekly) analyzes anonymized group data after desensitization, using clustering algorithms to discover optimal lighting parameter preference patterns for different professional groups (e.g., programmers, designers). When a new user uses the system for the first time, if their personal model library is empty (cold start), the system will use the corresponding group preference pattern as the initial lighting strategy based on the professional information they filled in during registration. Simultaneously, this process analyzes the user's long-term circadian rhythm, for example, by tracking the HRV characteristic changes one hour before rest each night to predict their diurnal rhythm phase, and based on this, fine-tunes the switching time between "awake" and "relaxed" modes of daily lighting adjustments within the following week to better match the user's individual circadian rhythm.

[0027] Furthermore, in step S1, the biometric types include at least three of the following: physiological signals reflecting the activity of the autonomic nervous system, EEG signals reflecting the state of the central nervous system, eye movement features reflecting visual fatigue and cognitive load, and facial thermal imaging features.

[0028] The collected biometrics are not limited to a single type, but cover at least three of the following three categories to ensure robustness of state determination: The first category consists of signals reflecting the activity of the autonomic nervous system. This can be achieved using PPG and EDA sensors worn on the wrist. PPG signals are processed to obtain heart rate and heart rate variability indicators, such as SDNN, pNN50, LF, and HF power spectra. EDA signals are decomposed to obtain skin conductance level (SCL) and skin conductance response (SCR) frequency, which are effective indicators of stress and arousal.

[0029] The second category consists of signals reflecting central nervous system activity. Specifically, lightweight dry electrode EEG head-mounted devices can be used to acquire electrical signals from the prefrontal cortex. A Fast Fourier Transform (FFT) is performed on the raw EEG data to calculate the relative or absolute power in the Delta (1-4Hz), Theta (4-8Hz), Alpha (8-13Hz), and Beta (13-30Hz) frequency bands. Alpha power is associated with relaxation, while Beta power is associated with focus and cognitive load.

[0030] The third category reflects eye movement characteristics that indicate visual and cognitive states. This involves tracking the pupil and corneal reflective points using a miniature infrared camera, calculating the pupil diameter and its rate of change over time (pupil response), and detecting blinking events to statistically analyze blink frequency. Pupil dilation is often associated with cognitive effort, while frequent blinking may indicate visual fatigue.

[0031] The fourth category is facial thermal imaging features. An uncooled infrared thermal imager is used to capture images of the user's face at a frequency of approximately 1 Hz. Image processing algorithms are used to locate landmarks such as the inner corner of the eye and the tip of the nose, delineating regions of interest above the left and right eye sockets, and calculating the average temperature of these areas. Temperature fluctuations caused by changes in blood flow in specific facial areas can be correlated with emotional arousal and fatigue.

[0032] In practical deployments, a complete signal acquisition array can be constructed using a head-mounted device integrating EEG and eye tracking, a wrist-worn PPG / EDA device, and a standalone USB thermal imager.

[0033] Furthermore, in step S2, the multi-level dynamic feature fusion includes: performing primary feature-level fusion on features of the same type; using an attention-based network to perform intermediate decision-level fusion on features of different types, dynamically learning the contribution weight of each feature to the inference of a specific state; and performing high-order state mapping on the fused feature vector to output a quantized state estimate.

[0034] The specific implementation steps of multi-level dynamic feature fusion are as follows: Primary Feature-Level Fusion: For features from multiple similar sensors, such as two forehead electrodes Fp1 and Fp2 in an EEG headset, their respective Alpha and Beta powers are calculated. The system first performs principal component analysis on these four features (Fp1_Alpha, Fp1_Beta, Fp2_Alpha, Fp2_Beta), retaining the first two principal components, and fusing the original four related features into two unrelated, information-condensed new features.

[0035] Intermediate Decision-Level Fusion: The system constructs a neural network layer based on scaled dot product attention. The input to this layer is all the heterogeneous feature vectors after primary fusion, such as: [EEG principal component 1, EEG principal component 2, pupil diameter, blink frequency, HRV_LF / HF, supraorbital temperature]. This attention layer generates a "query," "key," and "value" vector for each feature. By calculating the similarity between all feature keys and queries in the current context (which may include time information), a set of attention weights is obtained. This set of weights determines which features are more important when inferring the current "cognitive focus." For example, in the afternoon of a weekday, the beta power of EEG and pupil diameter may receive higher weights; while in the evening when relaxed, the HF power of HRV and the alpha power of EEG may receive higher weights. The weighted summed "value" vector is the dynamically weighted comprehensive state code.

[0036] Higher-order state mapping: The comprehensive state encoding described above is input into a pre-trained multilayer perceptron classification / regression model. This model is trained using supervised learning, with training data from a laboratory environment where volunteers provided their biometric data and self-reported state labels (such as Likert scale scores for focus and fatigue) under known lighting conditions and task loads. The model ultimately outputs one or more continuous numerical values ​​(e.g., between 0 and 1), representing the system's state estimates for different dimensions of the user.

[0037] Furthermore, the attention mechanism network can adaptively adjust feature weight allocation based on environmental context information. This is achieved by concatenating an environmental context encoding vector into the input of the attention layer, in addition to the biometric vector. This encoding vector consists of the following information: the time of day (encoded as sine and cosine functions to reflect periodicity), and the current activity type obtained through a calendar or manual user input (e.g., "deep work," "meeting," "rest," encoded as a one-hot vector). This concatenated extended vector then participates in the calculation of attention weights. Therefore, during the learning process, the attention mechanism naturally associates the importance of environmental context with different biometric features. For example, when the context indicates "deep work" and the time is "3 PM," the network model, through training, learns to refer more to EEG Beta waves and pupillary response features related to cognitive effort when calculating attention weights, while relatively weakening EDA features related to emotion. Furthermore, in step S3, the reinforcement learning framework uses the user state vector as the state space, the adjustment amount of the lighting parameters as the action space, and the user's subjective feedback or automatic evaluation based on biometric state improvement as the reward signal. The specific construction of the reinforcement learning framework is as follows: State space: Defined as a quantized user state vector obtained after feature fusion and higher-order mapping. For example, a two-dimensional state space can be [cognitive attention estimate, visual fatigue estimate], with each dimension normalized to the interval [0,1].

[0038] Action space: Defined as the adjustment instructions for two main lighting parameters. For example, the action space is discrete: {color temperature increases by 200K, color temperature decreases by 200K, illuminance increases by 50lx, illuminance decreases by 50lx, and remains unchanged}. It can also be continuous, with the action output being two continuous values ​​between [-1, 1], mapped to the actual adjustment amounts of color temperature (-500K to +500K) and illuminance (-100lx to +100lx) respectively through linear transformation.

[0039] Reward Signal: Calculated after each action is performed and a waiting period (e.g., 5 minutes). The reward R_t consists of two parts: 1) Subjective reward Rs: If the user provides feedback rating (e.g., 1-5 stars) through the interface after adjustment, the rating is normalized to the [-1,1] interval. 2) Objective reward Ro: Calculated as the difference between the average value of key state indicators (e.g., "attention level") over a period after adjustment and the baseline value before adjustment, also normalized to the [-1,1] interval. The final reward R_t = α*Rs + (1-α)*Ro, where α is a weighting coefficient (e.g., 0.7, emphasizing subjective feelings). If the user does not provide subjective feedback, then R_t = Ro. The reward signal is used to evaluate the merits of the previous action and drive the agent's policy update.

[0040] Furthermore, the reinforcement learning framework employs proximal policy optimization or deep Q-network algorithms.

[0041] When employing the Proximal Policy Optimization (PPO) algorithm, the agent comprises an "actor" network and a "critic" network. The "actor" network takes the current state s_t as input and outputs an action probability distribution (for discrete actions) or action mean and variance (for continuous actions), from which the actual action a_t is sampled. The "critic" network takes the state s_t as input and outputs a state value estimate V(s_t). After the agent interacts with the environment and collects a series of trajectory data, it calculates the advantage function A_t. The core of the PPO algorithm is to construct a pruned alternative objective function to update the parameters of the "actor" network, limiting the difference between the new and old policies to a small range, thereby achieving stable and efficient policy updates. The specific loss function and update steps follow the standard PPO algorithm formula. The system implements the agent using the PyTorch or TensorFlow framework and continuously performs offline training in the background using newly collected interaction data (state, action, reward, next state), periodically updating the trained new policy network parameters to the online system.

[0042] Furthermore, step S6 is also included: when the fused biometrics indicate that the user is in a preset abnormal state, the safety protection lighting mode is activated, the light is adjusted to a preset soothing mode and a reminder is issued.

[0043] Please see Figure 1 As shown, a system for implementing the method as described in any of the above includes: The multi-biometric sensing module is used to collect raw data of various biological characteristics; The data processing and fusion computing module is used to perform feature extraction, multi-level dynamic feature fusion, and user state mapping. The dynamic learning and policy engine module is used to maintain and update the personal lighting response model library and perform reinforcement learning-based lighting policy generation and optimization. The lighting control execution module is used to control lighting equipment according to lighting strategy instructions; The user interaction and feedback module is used to receive subjective feedback from users. The closed-loop verification module is used to verify the lighting adjustment effect and feed it back to the dynamic learning and strategy engine module.

[0044] A system for implementing the aforementioned method comprises the following hardware and software components: The multi-biometric sensing module physically includes: 1) a modified head-mounted device integrating EEG (NeuroSky MindWave Mobile2) and eye tracking (Pupil Labs Core); 2) a wrist-worn PPG / EDA sensor (Empatica E4); and 3) a USB-interface infrared thermal imager (FLIR Lepton 3.5). These devices connect to the host computer via Bluetooth or USB.

[0045] Data Processing and Fusion Computation Module: A Python process runs on the main control computer. This process uses the PyAV library to read video streams, NeuroSky and Empatica SDKs to read biosignals, OpenCV and Dlib for face and pupil recognition, Scikit-learn and NumPy for feature extraction and PCA dimensionality reduction, and PyTorch to load and run pre-trained feature attention fusion networks and high-order state mapping models.

[0046] Dynamic Learning and Policy Engine Module: This module runs another Python process on the main control computer. This process uses SQLAlchemy to manipulate the SQLite personal database, enabling CRUD operations on the personal lighting response model library. Simultaneously, a PPO reinforcement learning agent based on the Stable-Baselines3 library runs within this process, responsible for policy matching, exploration, and updating.

[0047] The lighting control execution module is a smart light fixture (such as a Philips Hue bulb or a Yeelight smart ceiling light) that connects via Wi-Fi. The main control computer sends JSON commands containing parameters such as color temperature, brightness, and fading time to the light fixture via an HTTP REST API or a dedicated SDK (such as the phue library).

[0048] User Interaction and Feedback Module: Develop a simple mobile app (using the Flutter framework) or desktop widget. The interface primarily displays the current lighting mode and two buttons: "Good Review" (thumbs up icon) and "Bad Review" (thumbs down icon). User click events are sent in real-time via WebSocket back to the host computer's dynamic learning and strategy engine module as subjective feedback signals.

[0049] The closed-loop verification module is not a standalone hardware component; rather, it is a collaborative effort between the data processing and fusion computing module and the dynamic learning and policy engine module. The former continuously calculates the adjusted state indicators, while the latter compares them with the pre-adjustment indicators, generates objective rewards, and combines them with subjective rewards to evaluate actions and update the model.

[0050] All modules exchange data on the host computer (such as a mini PC or Raspberry Pi 4) through internal inter-process communication (such as ZeroMQ) or shared memory, forming a complete localized system.

[0051] Furthermore, it also includes a meta-learning server for performing cross-user commonality analysis and long-term rhythm prediction, and for interacting with the dynamic learning and policy engine module.

[0052] Building upon the aforementioned system, a meta-learning server deployed in the cloud is added. This server primarily consists of the following components: 1) Database: Stores aggregated data from multiple user terminals that has undergone complete anonymization (removing all personally identifiable information, retaining only grouping labels and feature parameters such as occupation and age group). 2) Meta-learning model training service: Regularly (e.g., weekly) runs Jupyter Notebook scripts to analyze the anonymized data using Scikit-learn clustering algorithms (such as K-Means), mining lighting preference patterns among different user groups (labels), and training a predictive model that can predict the initial lighting preference parameters of new users based on their demographic information. 3) API interface: Provides query services for local terminal systems. When a new user cold starts, the local dynamic learning and policy engine module calls the API via HTTPS, submitting user group labels (e.g., {"occupation":"software_engineer","age_group":"30-40"}). The server returns the predicted initial lighting parameters (e.g., {"color_temp":4500,"illuminance":400}). The local system periodically uploads long-term rhythm analysis data summaries to the server, allowing the server-side model to provide periodic rhythm phase shift correction suggestions.

[0053] A computer-readable storage medium, such as an SD card, a solid-state drive, or cloud storage, stores an executable package. This package contains source code or a compiled executable file written in Python, along with associated configuration files, pre-trained model weight files (.pth format), and an SQLite database template.

[0054] When a person skilled in the art deploys the program in the storage medium onto a device with general computing capabilities (such as an x86 computer or an ARM development board) and runs it, the device will automatically perform the following operations: load the program, initialize the drivers for each sensor, and start all the software processes described in claims 8 and 9. These processes will work collaboratively according to the sequence of steps described in claims 1 to 7, from data acquisition, processing and fusion, learning and decision-making to controlling the lighting fixtures, forming a complete and operational human-caused adaptive lighting control system. A technician only needs to connect the specified hardware sensors and smart lighting fixtures according to the documentation and perform simple network configuration to get the entire system running, without needing to engage in creative development regarding algorithm principles and implementation details.

[0055] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A human-caused adaptive lighting adjustment method based on multi-biological feature fusion and dynamic learning, characterized in that, Includes the following steps: S1: Simultaneously collect raw biometric data of at least three different types of target users, and perform preprocessing and feature extraction to obtain time-aligned multimodal biometric vectors; S2: Perform multi-level dynamic feature fusion on the multimodal biometric feature vector to generate a dynamically weighted user state vector, and map it to a quantitative estimate of the user's current multidimensional physiological and psychological state. S3: Generate personalized lighting strategies based on dynamic learning mechanisms: Based on the current user status, query the user's personal lighting response model library to obtain the historical best strategy basis, and introduce an exploration-trade-off mechanism based on reinforcement learning for online optimization to generate the final lighting parameter instructions; the personal lighting response model library dynamically records and updates the user's status, lighting parameters, and the correlation between feedback / performance. S4: Adjust the lighting equipment according to the lighting parameter instructions, and collect biometric data after adjustment for closed-loop verification, and feed back the verification results to update the personal lighting response model library; S5: Based on a meta-learning model, it analyzes common patterns across users and long-term circadian rhythm changes, which can be used for cold start recommendations for new users and fine-tuning of long-term lighting adjustment targets.

2. The method according to claim 1, characterized in that, In step S1, the biometric types include at least three of the following: physiological signals reflecting the activity of the autonomic nervous system, EEG signals reflecting the state of the central nervous system, eye movement features reflecting visual fatigue and cognitive load, and facial thermal imaging features.

3. The method according to claim 1, characterized in that, In step S2, the multi-level dynamic feature fusion includes: performing primary feature-level fusion on features of the same type; using an attention-based network to perform intermediate decision-level fusion on features of different types, dynamically learning the contribution weight of each feature to the inference of a specific state; and performing high-order state mapping on the fused feature vector to output a quantized state estimate.

4. The method according to claim 3, characterized in that, The attention mechanism network can adaptively adjust the feature weight allocation based on environmental context information.

5. The method according to claim 1, characterized in that, In step S3, the reinforcement learning framework uses the user's state vector as the state space, the adjustment amount of the lighting parameters as the action space, and the user's subjective feedback or automatic evaluation based on biometric state improvement as the reward signal.

6. The method according to claim 1 or 5, characterized in that, The reinforcement learning framework employs proximal policy optimization or deep Q-network algorithms.

7. The method according to claim 1, characterized in that, It also includes step S6: when the fused biometrics indicate that the user is in a preset abnormal state, the safety protection lighting mode is activated, the light is adjusted to a preset soothing mode and a reminder is issued.

8. A system for implementing the method as described in any one of claims 1-7, characterized in that, include: The multi-biometric sensing module is used to collect raw data of various biological characteristics; The data processing and fusion computing module is used to perform feature extraction, multi-level dynamic feature fusion, and user state mapping. The dynamic learning and policy engine module is used to maintain and update the personal lighting response model library and perform reinforcement learning-based lighting policy generation and optimization. The lighting control execution module is used to control lighting equipment according to lighting strategy instructions; The user interaction and feedback module is used to receive subjective feedback from users. The closed-loop verification module is used to verify the lighting adjustment effect and feed it back to the dynamic learning and strategy engine module.

9. The system according to claim 8, characterized in that, It also includes a meta-learning server for performing cross-user commonality analysis and long-term rhythm prediction, and for interacting with the dynamic learning and policy engine module.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images