A method and system for adaptive adjustment of color temperature of LED light

Abstract

The application relates to the technical field of intelligent lighting, and discloses an adaptive adjustment method and system for the color temperature of an LED lamp, wherein the adaptive adjustment method for the color temperature of the LED lamp comprises the following steps: acquiring original spectrum signal data and ambient light color temperature values and performing pretreatment; extracting user physiological states and behavior characteristics, and identifying user state categories; performing standardized treatment on an ambient perception data set and user state category labels, adopting an adaptive weight distribution strategy and an attention mechanism to perform feature fusion; acquiring corresponding optimal color temperature decisions, and continuously optimizing the optimal color temperature decisions; determining a transition level according to an optimal target color temperature value and a current color temperature value, adjusting a transition time according to a decision confidence, and adopting an S-shaped curve interpolation to generate a color temperature gradient track sequence; acquiring a driving control signal, and modulating and calibrating the driving control signal; and the application realizes intelligent matching of the color temperature of LED lighting with environmental conditions, time rhythms and user states.

Description

Technical Field

[0001] This invention relates to the field of intelligent lighting technology, and more specifically, to an adaptive adjustment method and system for the color temperature of LED lamps. Background Technology

[0002] With the rapid development of LED lighting technology and the increasing demand for intelligent solutions, color-temperature adjustable LED lighting systems have been widely used in smart homes, smart offices, and healthcare. Color temperature, as an important parameter describing the color characteristics of a light source, has a significant impact on human visual comfort, emotional state, and circadian rhythms. Studies have shown that high color temperature cool white light can enhance alertness and work efficiency, while low color temperature warm white light is beneficial for relaxation and improving sleep quality. Therefore, dynamically adjusting the color temperature of LED lighting according to environmental conditions and user needs has significant theoretical and practical value for improving lighting quality, protecting user visual health, and maintaining the balance of the human circadian rhythm.

[0003] Existing LED color temperature adjustment technologies primarily employ a single-dimensional adjustment method based on time-based timing or ambient light perception, controlling color temperature through preset color temperature-time mapping tables or color temperature-illuminance mapping rules. While these methods can achieve basic color temperature adjustment, they suffer from significant shortcomings, including a single perception dimension, rigid decision-making mechanisms, and an inability to adapt to personalized user needs. Particularly in applications such as medical rehabilitation lighting, where high color temperature accuracy and smoothness are required, existing technologies struggle to meet the specific needs of visually sensitive patients. Inappropriate color temperature settings or sudden color temperature changes may adversely affect the patient's recovery process.

[0004] To address the aforementioned issues, an intelligent color temperature adjustment method is needed that can comprehensively sense ambient lighting conditions, circadian rhythm factors, and user physiological states. This method would achieve precise color temperature matching through multi-source data fusion and adaptive decision-making, while ensuring the smoothness and long-term stability of the color temperature adjustment process, in order to meet the high-quality lighting needs of different application scenarios, especially medical rehabilitation lighting scenarios. Summary of the Invention

[0005] This invention provides an adaptive adjustment method and system for LED light color temperature, solving the technical problems of single perception dimension, rigid decision-making mechanism, and inability to adapt to personalized user needs in related technologies.

[0006] This invention provides an adaptive adjustment method for the color temperature of LED lamps, comprising: The raw spectral signal data and ambient light color temperature values ​​are acquired and preprocessed to obtain the environmental perception dataset. Based on the environmental perception dataset, user physiological state and behavioral features are extracted, and multimodal fusion is used to identify user state categories to obtain user state category labels; The environmental perception dataset and user state category labels are standardized, and feature fusion is performed using an adaptive weight allocation strategy and an attention mechanism to obtain a comprehensive feature vector. Color temperature decision is made based on comprehensive feature vectors, and the optimal color temperature decision is continuously optimized to obtain the optimal target color temperature value and decision confidence. Obtain the current color temperature value, determine the transition level based on the optimal target color temperature value and the current color temperature value, adjust the transition time according to the decision confidence level, and generate a color temperature gradient trajectory sequence using S-curve interpolation; The target driving duty cycle of warm and cool LEDs is calculated based on the color temperature gradient trajectory sequence. A driving control signal is generated by pulse width modulation. The driving control signal is then modulated and calibrated to obtain the LED color temperature adjustment strategy.

[0007] In a preferred embodiment, the step of obtaining the environmental perception dataset includes: Deploy multispectral sensor nodes and adopt a distributed acquisition architecture to synchronously measure ambient light and obtain raw spectral signal datasets; Dark current correction and sliding window mid-range filtering are performed on the data of each channel to remove sampling points whose rate of change exceeds the preset anomaly judgment threshold; The spectral power distribution curve is estimated by spectral reconstruction algorithm, the tristimulus values ​​and chromaticity coordinates are calculated, and the blackbody radiation trajectory point closest to the chromaticity coordinates of the light source is found on the CIE chromaticity diagram to obtain the ambient light color temperature value. Calculate the solar altitude angle, and obtain the time rhythm reference color temperature value by looking up the mapping table between the solar altitude angle and the natural light color temperature; The ambient light color temperature value, spectral signal intensity value, and time rhythm reference color temperature value are encapsulated to obtain the environmental perception dataset.

[0008] In a preferred embodiment, the step of obtaining the user status category label includes: Infrared thermal imaging sensors are used to acquire thermal images of the user's facial region. Non-uniformity correction, temperature calibration, and spatial filtering are performed on the thermal image sequence. Background modeling and foreground segmentation methods are used to extract the user region. A cascaded classifier is used to extract facial thermal distribution features and eye state features to obtain the infrared modal user state feature vector. Frequency-modulated continuous wave millimeter-wave radar is used to detect the target area. The radar echo signal is mixed and processed by fast Fourier transform. The constant false alarm rate detection algorithm is used to extract motion features and micro-motion features to obtain the user state feature vector of the radar mode. The data quality of infrared and radar modes is evaluated, confidence scores for each mode are generated, contribution weights for each mode are determined, and feature vectors of each mode are concatenated to obtain the fused user state feature vector and mode fusion confidence. The fused user state feature vector is input into the pre-trained classification model, and the category with the highest probability is selected to obtain the user state category label and classification confidence.

[0009] In a preferred embodiment, the step of obtaining the comprehensive feature vector includes: The environmental perception dataset and user state category labels are resampled onto a unified time grid, and each feature dimension is standardized. Reliability assessments are conducted based on the quality metrics of each data source and the characteristics of the current scenario to obtain a reliability score for each data source. Based on the reliability score and the judgment results of the dominant factors in the scenario, the fusion weight values ​​of environmental characteristics, time characteristics and user characteristics are determined; The various standardized features are scaled and concatenated to form the input feature matrix. The attention weight matrix is ​​calculated, and the attention weight matrix is ​​multiplied by the value vector to obtain the fused high-dimensional feature representation. Principal component analysis is used to reduce the dimensionality of high-dimensional features, resulting in a comprehensive feature vector.

[0010] In a preferred embodiment, the step of obtaining the optimal target color temperature value and decision confidence level includes: Determine whether the number of samples in the experience replay buffer, the number of pre-training rounds of the reinforcement learning model, and the decision performance on the validation set have reached the preset thresholds. If all conditions are met, switch to the reinforcement learning model decision mode; otherwise, maintain the rule engine decision mode. When the system is in the rule engine decision mode, it makes color temperature decisions. The rule engine includes basic color temperature rules, environmental compensation rules, user state adjustment rules, and sensitivity protection rules. The output of each rule is weighted and summed to obtain the optimal target color temperature value. When the system is in the decision-making mode of the reinforcement learning model, a state space is defined that includes the comprehensive feature vector, the current LED output color temperature value, the recent color temperature adjustment history and the reliability scores of each data source, and a continuous action space is defined to represent the target color temperature adjustment amount. A reinforcement learning model is designed using an actor-critic architecture. The actor network outputs action values ​​based on the current state, and the critic network evaluates the value of the state-action pair. A priority experience replay mechanism is used for sample storage and sampling, and the parameters of the actor network and critic network are optimized by gradient descent. The current state vector is input into the actor network to obtain the action value. The action value is then added to the current color temperature value and the amplitude is limited to obtain the optimal target color temperature value. The decision confidence is obtained by comprehensively considering the reliability score of the input data, the Q-value evaluation of the critic network, and the degree of matching between the current state and the training sample distribution.

[0011] In a preferred embodiment, the step of generating the color temperature gradient trajectory sequence includes: Calculate the difference between the optimal target color temperature value and the current color temperature value. When the absolute value of the difference is less than the first threshold, it is determined to be a fine adjustment level. When the absolute value of the difference is between the first threshold and the second threshold, it is determined to be a standard transition level. When the absolute value of the difference is greater than the second threshold, it is determined to be a large adjustment level. The transition level and the corresponding transition time parameter are obtained. The decision confidence level is mapped to an adjustment factor for the transition time. The base transition time corresponding to the transition level is multiplied by the adjustment factor to obtain the actual transition time. The number of trajectory points is calculated based on the actual transition time and system control cycle. The normalized time parameter of each trajectory point is calculated, and the normalized color temperature value is calculated. The normalized color temperature value is then converted into the actual color temperature value through linear mapping to obtain the color temperature gradient trajectory sequence. The smoothness of the color temperature gradient trajectory sequence is verified by checking the color temperature change rate of adjacent trajectory points, the abrupt change point of the second derivative of the trajectory, and the monotonicity of the overall trajectory. Trajectory segments that do not meet the requirements are corrected to obtain the final trajectory sequence.

[0012] In a preferred embodiment, the step of generating the color temperature gradient trajectory sequence using S-curve interpolation further includes: A generalized S-curve with adjustable parameters is used. The generalized S-curve introduces shape control parameters on the basis of the standard logistic function. The value of the shape parameter is determined according to the user's visual sensitivity level.

[0013] In a preferred embodiment, the step of obtaining accurate and stable LED color temperature output includes: Based on the color temperature gradient trajectory sequence, the theoretical luminous flux ratio of warm and cool LEDs is calculated using a color temperature driving parameter mapping model, and the driving current ratio is calculated by combining the luminous flux driving current characteristic curve; the LED driving control signal is generated using a pulse width modulation method. A composite control structure of feedforward and feedback is adopted to achieve fast response control in the inner loop. The feedforward channel predicts the changing trend of LED output based on the change of the target duty cycle, and the feedback channel uses a proportional-integral control algorithm to adjust the duty cycle according to the deviation. A color temperature feedback sensor is used to periodically collect the actual output color temperature value and compare it with the color temperature value of the trajectory target to calculate the color temperature deviation value. The color temperature driving parameter mapping model is corrected online. The parameter correction amount is calculated, and the parameter correction amount is superimposed on the current mapping parameter after being limited to obtain the calibrated mapping parameter.

[0014] In a preferred embodiment, the inner-loop fast response control further includes: Model predictive control is adopted, and the system output of the LED driving system is calculated based on the discrete state-space prediction model of the LED driving system for the future preset number of steps. The control sequence is obtained by solving the objective function. The objective function includes the deviation term between the predicted output and the target trajectory and the penalty term for the change of control quantity.

[0015] This invention provides an adaptive color temperature adjustment system for LED lamps, used to execute the aforementioned adaptive color temperature adjustment method for LED lamps, comprising: The environmental perception module is used to acquire raw spectral signal data and ambient light color temperature values ​​and preprocess them to obtain an environmental perception dataset. The user status recognition module extracts user physiological status and behavioral features based on the environmental perception dataset, and uses multimodal fusion to identify user status categories and obtain user status category labels. The feature fusion module is used to standardize the environmental perception dataset and user state category labels, and uses an adaptive weight allocation strategy and attention mechanism to fuse features to obtain a comprehensive feature vector. The decision optimization module makes color temperature decisions based on comprehensive feature vectors, continuously optimizes the optimal color temperature decision, and obtains the optimal target color temperature value and decision confidence. The trajectory generation module is used to obtain the current color temperature value, determine the transition level based on the optimal target color temperature value and the current color temperature value, adjust the transition time based on the decision confidence level, and generate a color temperature gradient trajectory sequence using S-curve interpolation; The drive control module calculates the target drive duty cycle of warm and cool LEDs based on the color temperature gradient trajectory sequence, generates a drive control signal through pulse width modulation, and modulates and calibrates the drive control signal to obtain the LED color temperature adjustment strategy.

[0016] The beneficial effects of this invention are as follows: By organically combining multi-dimensional perception, intelligent fusion, adaptive decision-making, and precise execution, it achieves deep matching between lighting color temperature and environmental conditions, time rhythm, and user status. It has advantages such as comprehensive perception, intelligent decision-making, smooth adjustment, and long-term stability, and can provide users, especially visually sensitive people, with a high-quality personalized lighting experience.

[0017] The deep reinforcement learning decision model with a preheating mechanism adopted in this invention effectively solves the cold start problem of traditional reinforcement learning methods. In the early stage of model training, the decision quality is guaranteed by an expert rule engine. As samples accumulate, the system gradually switches to a reinforcement learning model to achieve adaptive optimization. This enables the system to continuously improve decision strategies based on users' personalized preferences and usage habits, adapt to complex and ever-changing real-world application scenarios, and improve the intelligence level of color temperature adjustment and user satisfaction. Attached Figure Description

[0018] Figure 1 This is a flowchart of the main process of an adaptive adjustment method for LED lamp color temperature according to the present invention; Figure 2 This is a detailed flowchart of an adaptive adjustment method for LED lamp color temperature according to the present invention; Figure 3 This is a block diagram of an adaptive color temperature adjustment system for LED lamps according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses an adaptive adjustment method for the color temperature of an LED lamp, such as... Figures 1 to 2 As shown, it includes: Step 1: Obtain the raw spectral signal data and ambient light color temperature values ​​and perform preprocessing to obtain the environmental perception dataset; Specifically, the following steps are included: Step 1.1, Deployment of multispectral sensor array and acquisition of light signals; Based on the spatial layout characteristics and lighting coverage requirements of the rehabilitation ward, multispectral sensor nodes were deployed at multiple key locations in the target lighting area. A distributed acquisition architecture was used to synchronously measure the ambient light, thereby obtaining the original spectral signal dataset covering the target area.

[0021] Specifically, the multispectral sensor uses a photodetector with full visible light band coverage, encompassing the entire visible light spectrum from short-wavelength blue light to long-wavelength red light. The number of bands is determined based on the accuracy requirements of color temperature calculation, typically set to at least several independent photosensitive channels. Sensor nodes are arranged according to a pre-defined spatial grid, with the node spacing determined based on the area of ​​the illuminated region and the requirements for illumination uniformity, ensuring that the collected data reflects the overall illumination distribution of the target area. Sensor nodes are connected to the data acquisition host via wired or wireless means. The acquisition host synchronously triggers each node to acquire spectral signals according to a pre-defined sampling period. The sampling period is determined based on the rate of change in ambient light and system response requirements, avoiding data redundancy and computational burden caused by excessively high sampling frequencies while ensuring real-time performance.

[0022] Step 1.2, Spectral signal preprocessing and quality assessment; Based on the acquired raw spectral signal dataset, noise suppression and outlier removal are performed on the data of each channel. Combined with the signal quality assessment mechanism, valid data is screened to obtain a quality-screened spectral signal dataset.

[0023] Specifically, dark current correction is performed on the raw signals of each channel of each sensor node to subtract the sensor's background output value under no-light conditions, thus eliminating the influence of inherent device bias. A sliding window mid-range filtering method is used to smooth the time-series signal. The length of the filtering window is determined based on the signal noise characteristics and response speed requirements, effectively suppressing random noise while preserving the true trend of signal change. For the filtered signal, its rate of change compared to the signal at adjacent time points is calculated. When the rate of change exceeds a preset anomaly detection threshold, the sampling point is marked as a potential anomaly, and the interpolation result of adjacent valid sampling points is used to replace it.

[0024] After preprocessing, the signal quality of each sensor node is comprehensively evaluated. Evaluation indicators include whether the signal amplitude is within the effective range, whether the response consistency between channels meets the requirements, and whether the signal stability meets the preset standards. Node data with quality evaluation results below the preset threshold are marked as low-confidence data, and their weight is reduced or they are removed in subsequent processing.

[0025] Step 1.3, Calculation of ambient light color temperature; Based on the quality-screened spectral signal dataset, the correlation color temperature calculation method is used to convert the multi-channel spectral signals into color temperature values, thereby obtaining the ambient light color temperature values ​​at the corresponding locations of each sensor node.

[0026] Specifically, based on the spectral response characteristics of each channel and the acquired signal intensity values, the continuous spectral power distribution curve of ambient light is estimated using a spectral reconstruction algorithm. Spectral reconstruction employs a basis function expansion method, representing the continuous spectrum as a linear combination of several preset basis functions. The basis functions are selected based on principal component analysis results of typical light source spectra, and the combination coefficients are obtained using a least-squares fitting method based on the measured channel signal values. After obtaining the spectral power distribution curve, the tristimulus values ​​of the light source are calculated according to the standard calculation procedure specified by the International Commission on Illumination (CIE). The tristimulus values ​​are obtained by integrating the spectral power distribution with the CIE standard chromaticity observer spectral tristimulus value function. Based on the tristimulus values, chromaticity coordinates are further calculated, representing the position of the light source color on the CIE chromaticity diagram. Using a correlated color temperature (CCT) calculation method, the blackbody radiation trajectory point closest to the light source's chromaticity coordinates is found on the CIE chromaticity diagram; the blackbody temperature corresponding to this point is the correlated color temperature value of the ambient light. The color temperature values ​​calculated by each sensor node are weighted and averaged according to the spatial location and quality assessment results of the nodes to obtain the comprehensive ambient light color temperature value of the target lighting area.

[0027] Step 1.4, Calculation of time rhythm reference color temperature value; Based on the current time information obtained from the system's real-time clock and the pre-stored geographical location parameters, the time rhythm reference color temperature value corresponding to the current moment is obtained by using the solar position calculation model and the natural light color temperature mapping relationship.

[0028] Specifically, the system reads the current date and time information from the real-time clock module, including complete time elements such as year, month, day, hour, minute, and second. Combined with pre-configured geographical location parameters of the lighting location, including longitude, latitude, and time zone information, astronomical calculation methods are used to calculate the sun's position relative to the lighting location at the current moment. The sun's position calculation uses the date to determine the day number and solar declination angle for the year; the declination angle reflects the north-south shift of the sun's direct point relative to the equator. The solar hour angle is calculated based on the local time and longitude; the hour angle reflects the east-west shift of the sun relative to the local meridian. Based on the declination angle, hour angle, and local latitude, the solar altitude angle is calculated using spherical trigonometry; the solar altitude angle represents the angle between sunlight and the ground plane. There is a corresponding relationship between the color temperature of natural light and the solar altitude angle. When the solar altitude angle is low, the path of sunlight through the atmosphere is longer, resulting in greater scattering loss of short-wave blue light, and natural light exhibits a warmer color temperature. When the solar altitude angle is high, the path of sunlight through the atmosphere is shorter, and natural light exhibits a cooler color temperature.

[0029] The system pre-stores a lookup table mapping solar altitude angle to natural light color temperature. This lookup table is built based on a large amount of measured data and atmospheric optical models, covering the entire solar altitude angle range from sunrise to sunset. Based on the calculated current solar altitude angle, the corresponding circadian rhythm reference color temperature value is obtained through table lookup and interpolation. For nighttime hours, since there is no natural light, the circadian rhythm reference color temperature value is set to a low color temperature suitable for rest, based on research findings on human circadian rhythms.

[0030] Based on the ambient light color temperature value, spectral signal intensity value of each channel, signal quality assessment result, and time rhythm reference color temperature value obtained from the aforementioned steps, the data is organized and encapsulated according to a predefined data structure to obtain a complete environmental perception dataset, which is then output to the subsequent processing module.

[0031] Specifically, the environmental perception dataset includes the following data fields: comprehensive ambient light color temperature, ambient light intensity, normalized signal intensity vectors for each spectral channel, spectral data quality confidence level, time-rhythm reference color temperature, current timestamp, and solar altitude angle. Each data field is encoded according to a unified data format, with numerical data represented using fixed-point or floating-point representation and accompanied by a data validity flag. The encapsulated environmental perception dataset is transmitted to the multi-source data fusion module via an internal data bus, serving as the environmental input information for color temperature decision-making.

[0032] Step 2: Based on the environmental perception dataset, extract the user's physiological state and behavioral features, and use multimodal fusion to identify the user state category to obtain the user state category label; Specifically, the following steps are included: Step 2.1, Infrared thermal imaging data acquisition and preprocessing; Based on the non-contact user monitoring needs in rehabilitation wards, infrared thermal imaging sensors are used to continuously acquire thermal images of the user's facial area. Effective facial thermal distribution information is extracted through image preprocessing to obtain a preprocessed facial thermal imaging sequence.

[0033] Specifically, the infrared thermal imaging sensor is deployed at an appropriate location in the illuminated area, with its field of view covering the main activity areas where the user is likely to be. The sensor's temperature resolution and spatial resolution are determined based on the accuracy requirements of facial feature recognition. The sensor continuously acquires infrared thermal images at a preset frame rate. The frame rate setting needs to balance the real-time recognition and data processing burden, and is usually set to several frames per second. For the acquired raw thermal image sequence, non-uniformity correction is performed to eliminate fixed pattern noise caused by inconsistent responses of pixels in the infrared detector array; temperature calibration is performed, converting the digital quantities output by the pixels into corresponding temperature values. Calibration parameters are obtained by periodically measuring a standard blackbody radiation source; spatial filtering is performed on the calibrated thermal images, using a Gaussian smoothing filter to suppress random noise. The filter kernel size is determined based on the noise level and detail preservation requirements; after preprocessing, background modeling and foreground segmentation methods are used to extract the user region from the thermal images. The background model is established using a Gaussian mixture model, which can adapt to slow changes in background temperature. Foreground segmentation is achieved by calculating the difference between the current frame and the background model and applying a threshold.

[0034] Step 2.2, Localization and feature extraction of key facial regions; Based on the preprocessed facial thermal imaging sequence, facial detection and key point localization algorithms are used to determine the location of the main facial regions, extract facial thermal distribution features and eye state features, and obtain the user state feature vector of the infrared mode.

[0035] Specifically, face detection is performed within the user region obtained from foreground segmentation. Face detection employs a cascaded classifier-based method, trained on a large number of infrared facial image samples, enabling accurate localization of the facial region's bounding box in the thermal image. Within the detected facial region, further facial keypoint localization is performed. Keypoints include locations with stable thermal features such as the center of the eye region, the tip of the nose, and the corners of the mouth. Keypoint localization uses a regression-based method, directly predicting the coordinates of each keypoint from the facial region image using a pre-trained regression model. Based on the localized keypoint locations, facial thermal distribution features are extracted, including the overall average facial temperature and the facial temperature distribution index. The data includes statistical features such as accuracy and temperature differences between the eye area and other facial areas; the extraction of eye state features focuses on the thermal characteristics of the eye area, and the opening and closing state of the eyes is determined by analyzing the temporal temperature change patterns in the eye area. When the eyes are open, the exposed area of ​​the eyeball is larger, and the temperature distribution in the eye area shows a specific pattern; when the eyes are closed, the eyelids cover the eyeball, and the temperature distribution in the eye area changes accordingly; through temporal analysis of eye features in consecutive frames, indicators such as blinking frequency and the proportion of eye closure time are calculated. These indicators are closely related to the user's fatigue level and concentration state; after normalization, the extracted features are spliced ​​to form an infrared modal user state feature vector.

[0036] Step 2.3, Millimeter-wave radar signal acquisition and processing; Based on the need to perceive the user's motion state and micro-motion characteristics, millimeter-wave radar is used to continuously detect the target area. The user's motion feature information is extracted through radar signal processing to obtain the user state feature vector of the radar mode.

[0037] Specifically, millimeter-wave radar employs a frequency-modulated continuous wave (FMCW) system, transmitting electromagnetic waves whose frequency varies linearly with time and receiving the reflected echoes from the target to simultaneously measure the target's range and velocity. The radar deployment location and antenna pointing are determined based on the coverage requirements of the user's activity area, while parameters such as the radar's operating frequency band, bandwidth, and sweep period are configured according to the range resolution and velocity resolution requirements. The radar continuously detects targets according to a preset frame period, with each frame containing data from several frequency sweep cycles. The acquired radar echo signals are mixed to obtain an intermediate frequency (IF) signal, the frequency of which is proportional to the target distance. A Fast Fourier Transform (FFT) is performed on the IF signal to obtain the range-dimensional spectral distribution, with the peak positions corresponding to the target's range information. A second FFT is performed on the range-dimensional data from multiple frequency sweep cycles to obtain the velocity-dimensional spectral distribution, achieving range-Doppler two-dimensional imaging. Target detection is performed on the range-Doppler image using a Constant False Alarm Rate (CFAR) detection algorithm, which adaptively adjusts the detection threshold to detect targets while maintaining a constant false alarm probability. Detected targets are tracked and correlated, and the temporal variation curves of motion parameters such as distance, velocity, and acceleration are extracted. Based on the temporal data of these motion parameters, the user's macroscopic motion characteristics are extracted, including motion amplitude, motion frequency, and motion regularity.

[0038] Simultaneously, micro-motion analysis is performed on the radar signal. Micro-motion refers to small-amplitude periodic movements superimposed on the overall target motion, such as the subtle movements of the human body caused by breathing and heartbeat. By performing time-frequency analysis on the phase signal of a specific range cell, physiological parameters such as respiratory frequency and amplitude are extracted. These parameters can reflect the user's degree of relaxation and physiological state. The extracted motion features and micro-motion features are normalized and then concatenated to form the user state feature vector of the radar mode.

[0039] Step 2.4, Multimodal feature fusion and confidence assessment; Based on the user state feature vectors extracted from the infrared mode and the radar mode respectively, a confidence-weighted feature-level fusion method is used to effectively integrate the features of the two modes, resulting in the fused user state feature vector and the mode fusion confidence.

[0040] Specifically, the data quality of each modality is independently evaluated, generating modal confidence scores. The confidence assessment for the infrared modality considers the following factors: whether face detection was successful, whether the face occupies a sufficient proportion of the image, whether there is facial occlusion, and whether the signal-to-noise ratio of the thermal image meets the standards. The confidence assessment for the radar modality considers the following factors: whether the user target was successfully detected, whether the target echo intensity is sufficient, whether there is multi-target interference, and whether micro-motion signal extraction was successful.

[0041] Each evaluation factor is assigned a different weight according to its importance, and the weighted sum is used to obtain the confidence score for each modality. The confidence score is then normalized to the range of zero to one. Based on the confidence scores of each modality, an adaptive weight allocation strategy is used to determine the contribution weight of each modality during feature fusion. Modalities with higher confidence receive larger fusion weights, while modalities with lower confidence receive correspondingly smaller weights. When the confidence of a modality is lower than a preset minimum usable threshold, the data of that modality is deemed unreliable and relies entirely on the data of another modality during the fusion process. Feature fusion uses a weighted concatenation method, scaling the feature vectors of each modality according to the calculated weights before concatenation to form the fused user state feature vector. Simultaneously, the modality fusion confidence score is calculated based on the weighted average of the confidence scores of each modality. This confidence score reflects the overall reliability of the fused features and will be used in subsequent data fusion and decision-making processes.

[0042] Step 2.5, User Status Classification and Recognition; Based on the fused user state feature vector, a pre-trained classification model is used to identify the user's current state category, and the user state category label and classification confidence score are obtained.

[0043] Specifically, the classification of user states is determined based on the actual needs of medical rehabilitation lighting scenarios, and mainly includes the following categories: focused work state, characterized by the user remaining relatively still, eyes continuously gazing in a specific direction, and steady breathing; relaxed state, characterized by the user engaging in slight activity, blinking at a normal frequency, and slow breathing; rest and sleep state, characterized by the user remaining basically still, eyes closed for a long time, and deep and regular breathing; and active movement state, characterized by the user exhibiting obvious positional movement or limb movement.

[0044] The classification model employs a lightweight neural network architecture. The network input is the fused user state feature vector, and the output is the probability distribution of each state category. The network structure contains several fully connected layers, with non-linear activation functions between layers. The final layer uses a normalized exponential function (Softmax) to convert the output into a probability distribution. Network parameters are obtained through offline pre-training on a large number of labeled samples, covering various state categories and different users to ensure good generalization ability. During online recognition, the fused feature vector is input into the network for forward inference, and the category with the highest probability is taken as the recognition result. The probability value corresponding to this category is used as the classification confidence. The user state category label, classification confidence, fused user state feature vector, and modality fusion confidence are packaged and output as the user state input for subsequent multi-source data fusion.

[0045] Furthermore, since traditional fully connected neural networks struggle to effectively capture the temporal dependencies of state changes when processing temporal features, Long Short-Term Memory (LSTM) networks can be used instead of fully connected networks for user state classification. The aim is to enhance the model's ability to model the temporal evolution of user states and improve the temporal consistency and accuracy of state recognition. Specifically, LSTM networks employ a gating mechanism to effectively control the memorization and forgetting of long-term dependent information. The network includes three gating units: an input gate, a forget gate, and an output gate. The input gate controls the extent to which current input information is written into the memory unit, the forget gate controls the extent to which historical memory information is retained, and the output gate controls the extent to which information from the memory unit is output to the hidden state. The fused feature vectors from multiple consecutive frames are fed into the LSTM network as temporal input. When processing each frame, the network combines the memory information from previous frames for comprehensive judgment. The final output hidden state vector contains temporal context information. This vector is then passed through fully connected layers and a softmax layer to obtain the state category probability distribution. Using an LSTM network effectively avoids the jitter problem in single-frame recognition results, improving the stability of state recognition.

[0046] Step 3: Standardize the environmental perception dataset and user state category labels, and use an adaptive weight allocation strategy and attention mechanism to fuse features to obtain a comprehensive feature vector; Specifically, the following steps are included: Step 3.1, Multi-source data alignment and standardization; Based on the differences between environmental perception datasets and user status category labels in terms of timestamps, data formats, and numerical ranges, a data alignment and standardization process is adopted to convert various types of data into a unified format, resulting in aligned and standardized feature data.

[0047] Specifically, time alignment is performed because the sampling periods of the environmental perception module and the user state recognition module may differ, requiring the two types of data to be aligned to a unified time reference. Nearest neighbor interpolation or linear interpolation methods are used to resample various types of data onto a unified time grid based on data timestamps, ensuring strict temporal correspondence among the data participating in the fusion. Numerical standardization is then performed because the original numerical ranges and dimensions of various feature data differ; direct fusion would result in features with larger numerical ranges dominating the fusion result. Zero-mean unit variance standardization is used to independently standardize each feature dimension. Standardization parameters include the mean and standard deviation of each feature dimension, obtained through statistical analysis of historical data and updated periodically to adapt to changes in data distribution. Scalar features such as ambient light color temperature, light intensity, and circadian rhythm reference color temperature are directly standardized. Vector features such as spectral channel signal vectors and user state feature vectors are standardized for each component separately. The standardized feature data has a unified numerical range, facilitating subsequent weight allocation and feature fusion.

[0048] Step 3.2, Data source reliability assessment; Based on the quality indicators of each data source and the characteristics of the current scenario, a reliability assessment method is used to quantitatively evaluate the reliability of environmental perception data and user status data, and obtain the reliability score of each data source.

[0049] Specifically, the reliability assessment of environmental sensing data comprehensively considers the following factors: confidence level of spectral data quality, reflecting the signal-to-noise ratio and effectiveness of data collected by multispectral sensors; ambient light stability index, obtained by calculating the rate of change of recent ambient light color temperature values, a smaller rate of change indicates stable ambient light and higher data reliability; and sensor operating status indicator, reflecting whether the sensor is in normal working condition.

[0050] The reliability assessment of user status data comprehensively considers the following factors: modality fusion confidence, reflecting the overall quality of infrared and radar data; classification confidence, reflecting the certainty of user status identification results; and user presence indicators, where the reliability of user status data should decrease when no user is detected within the target area. Each assessment factor is assigned a corresponding weight based on its impact on data reliability, and a reliability score for each data source is calculated through weighted summation. The reliability score is normalized to a range of zero to one; a higher score indicates more reliable data and should receive greater weight during the fusion process.

[0051] Step 3.3, Adaptive fusion weight calculation; Based on the reliability scores of each data source and the judgment results of the dominant factors in the current scenario, an adaptive weight calculation method is used to determine the contribution weights of environmental features, time features, and user features in the fusion process, and obtain the fusion weight values ​​of each type of feature.

[0052] Specifically, the calculation of the fusion weights needs to consider both data reliability and scene relevance dimensions. The data reliability dimension directly uses the reliability score calculated above; data sources with higher reliability receive a larger base weight. The scene relevance dimension requires judging the importance of various data types to color temperature decisions based on the characteristics of the current scene. The judgment criteria include: when the rate of change in ambient light exceeds a preset threshold, it indicates that the ambient light is changing significantly, and environmental features have a more critical impact on color temperature decisions, thus increasing the weight of environmental features; when the user's state changes, it indicates that the user's color temperature needs may change, and the weight of user features should be increased accordingly; within specific time periods, such as near sunset or sunrise, the influence of time rhythm factors on color temperature is more significant, and the weight of time features should be increased. The combined data reliability weight and scene relevance weight are multiplied to obtain the comprehensive fusion weight of various features, and the comprehensive weight is normalized to ensure that the sum of the weights of all features is one.

[0053] Step 3.4, Feature interaction fusion based on attention mechanism; Based on the standardized feature data and the calculated fusion weights, a feature fusion network based on a self-attention mechanism is used to interactively learn and enhance the fusion of various features, resulting in a fused high-dimensional feature representation.

[0054] Specifically, the self-attention mechanism can automatically learn the relationships between features, enhance important features, and suppress redundant features. Various standardized features are scaled according to fusion weights to obtain weighted feature vectors. These weighted feature vectors are concatenated to form an input feature matrix, where each row corresponds to a vector representation of a feature class. The input feature matrix is ​​then fed into the self-attention calculation module, which includes three linear transformations that map the input features to query vectors, key vectors, and value vectors, respectively. The attention weight matrix is ​​obtained by calculating the dot product of the query vector and the key vector, followed by scaling and normalization. The attention weights reflect the strength of the relationships between features. Finally, the attention weight matrix is ​​multiplied by the value vectors to obtain the attention-weighted feature representation, which incorporates the interaction information between various features.

[0055] To enhance the expressive power of feature fusion, a multi-head attention mechanism is employed. Input features are fed into multiple parallel attention computation modules, each independently learning feature associations in different subspaces. Finally, the outputs of each module are concatenated and subjected to a linear transformation to obtain the final fused feature representation. This fused feature representation undergoes residual connections and layer normalization, preserving the original feature information while incorporating interactive enhancement information.

[0056] Step 3.5: Feature dimensionality reduction and output of the comprehensive feature vector; Based on the high-dimensional feature representation obtained by attention fusion, a feature dimensionality reduction method is used to compress the high-dimensional features into a comprehensive feature vector of fixed dimensions, resulting in the final output comprehensive feature vector and its associated information.

[0057] Specifically, while high-dimensional fused features contain rich information, excessive dimensionality increases the computational complexity of subsequent decision-making models and may contain redundant information. Principal Component Analysis (PCA) is used to reduce the dimensionality of high-dimensional features. PCA decomposes the feature covariance matrix into eigenvalues, identifying the orthogonal directions with the largest data variance as principal components. The original high-dimensional features are then projected onto these principal component directions to obtain the dimensionality-reduced feature representation. The number of principal components is determined based on the cumulative variance contribution rate, selecting the minimum number that retains a preset proportion of the original information, thus achieving a balance between dimensionality reduction and information preservation. The resulting fixed-dimensional comprehensive feature vector encapsulates comprehensive information on ambient lighting, temporal rhythm, and user status, effectively supporting subsequent color temperature decisions. The comprehensive feature vector, along with the fusion weights of each data source, data reliability scores, and other supplementary information, is output for use by the subsequent decision-making module.

[0058] Furthermore, since principal component analysis (PCA) is a linear dimensionality reduction method, it struggles to capture the nonlinear relationships between features. Therefore, an autoencoder can be used for nonlinear feature dimensionality reduction. The aim is to better preserve the nonlinear structural information in high-dimensional features and improve the expressive power of the comprehensive feature vector. Specifically, an autoencoder is an unsupervised learning neural network structure consisting of an encoder and a decoder. The encoder compresses high-dimensional input features into low-dimensional hidden layer representations, and the decoder reconstructs these low-dimensional hidden layer representations into high-dimensional outputs. The training objective is to minimize the difference between the input and the reconstructed output. The encoder uses a multi-layer fully connected network, with the number of neurons decreasing layer by layer. The output of the last layer is the dimensionality-reduced comprehensive feature vector. The decoder uses a network structure symmetrical to the encoder, with the number of neurons increasing layer by layer. The output dimension of the last layer is the same as the original input. The network training uses mean squared error as the loss function, and the network parameters are optimized through backpropagation. After training, only the encoder is used for online feature dimensionality reduction. The high-dimensional fused features are input into the encoder to obtain the low-dimensional comprehensive feature vector. The autoencoder can learn the nonlinear mapping relationships between features, and the dimensionality-reduced features retain more of the original information structure.

[0059] Step 4: Make color temperature decisions based on the comprehensive feature vector, continuously optimize the optimal color temperature decision, and obtain the optimal target color temperature value and decision confidence. Specifically, the following steps are included: Step 4.1, Decision-making mode judgment and preheating mechanism; Based on the system's operating status and historical sample accumulation, a decision-making mode judgment logic is used to determine whether to use a rule engine or a reinforcement learning model for decision-making, thereby obtaining the current decision-making mode indicator and the corresponding decision execution path.

[0060] Specifically, reinforcement learning models often exhibit unsatisfactory decision-making performance in the early stages of training due to a lack of sufficient experience samples, and direct deployment may lead to a decline in user experience. To address this cold start problem, the system introduces a warm-up mechanism, employing a rule engine based on expert knowledge for color temperature decision-making before the model matures. The decision-making mode is determined based on: whether the number of samples in the experience replay buffer reaches a preset minimum training threshold; whether the reinforcement learning model has completed a preset number of pre-training rounds; and whether the model's decision-making performance on the validation set reaches a preset performance threshold. When all of the above conditions are met, the system switches to the reinforcement learning model decision-making mode; otherwise, the system maintains the rule engine decision-making mode. In the rule engine decision-making mode, the system simultaneously collects state-action-reward samples for model training, enabling parallel execution of decision-making services and model training.

[0061] Step 4.2, rule engine color temperature decision; Based on the key feature components in the comprehensive feature vector, when the system is in the warm-up stage, a rule engine based on expert knowledge is used to make color temperature decisions and obtain the optimal target color temperature value output by the rule engine.

[0062] Specifically, the decision-making logic of the rule engine is built upon expert experience in the lighting field and medical lighting standards, and mainly includes the following decision rules: Base color temperature rule: A base target color temperature is set based on a time-rhythm reference color temperature value. The base color temperature is higher during the day and lower at night, with a smooth change during transition periods. Environmental compensation rule: Compensation adjustments are made based on the difference between the ambient light color temperature and the base color temperature. When the ambient light color temperature deviates significantly from the base color temperature, the LED output color temperature is adjusted in the opposite direction to maintain the stability of the overall lighting color temperature. User state adjustment rule: Color temperature is personalized according to the user's state category. The color temperature is appropriately increased to enhance alertness when focused on work, decreased to promote relaxation when resting or sleeping, and kept neutral when active. Sensitivity protection rule: For visually sensitive patients, the upper limit of the color temperature is further reduced based on the above rules to avoid eye irritation from high color temperature light. The outputs of each rule are weighted and summed to obtain the final optimal target color temperature value. The weights are determined based on the priority of each rule and the applicability of the current scenario.

[0063] Step 4.3, define the state space and action space for reinforcement learning; Based on the characteristics of the color temperature decision task and the structure of the comprehensive feature vector, the state space and action space of the reinforcement learning model are defined to provide a standardized input-output interface for the training and inference of the model.

[0064] Specifically, the state space definition needs to include all key information affecting color temperature decisions. The state vector consists of the following components: components of each dimension of the comprehensive feature vector, including fused information of ambient lighting, temporal rhythm, and user state; the current LED output color temperature value, reflecting the current state of the system; recent color temperature adjustment history, containing a sequence of color temperature values ​​over several recent time steps, used to capture trend information of color temperature changes; and reliability scores of each data source, used to indicate the confidence level of the current input information. The action space is defined in the form of a continuous action space. The action value represents the adjustment amount of the target color temperature. The range of the adjustment amount is determined based on the adjustable range of the LED light source's color temperature and the reasonable amplitude of a single adjustment, usually limited to a certain range to avoid drastic color temperature changes. A positive action value indicates an increase in color temperature, and a negative value indicates a decrease in color temperature; the absolute value represents the adjustment amplitude. The optimal target color temperature value is calculated by adding the adjustment amount to the current color temperature value and is limited to the adjustable range of the LED light source's color temperature.

[0065] Step 4.4, Actor-Critic Network Architecture Design; Based on the defined state space and action space, an actor-critic architecture is used to design the network structure of the reinforcement learning model, resulting in a neural network model for policy learning and value evaluation.

[0066] Specifically, the actor-critic architecture comprises two independent neural networks: the actor network and the critic network. The actor network is responsible for outputting actions based on the current state. Its input is a state vector, and its output is the action value. The network structure employs a multi-layer fully connected network. The hidden layers use the Rectified Linear Unit (ReLU) activation function, and the output layer uses the hyperbolic tangent function Tanh to constrain the output to a range of -1 to +1, then maps it to the actual action value range through linear scaling. The critic network is responsible for evaluating the value of the state-action pair. Its input is the concatenation of the state vector and the action value, and its output is the Q-value, which represents the cumulative expected reward obtained by performing the action in the current state. The critic network also uses a multi-layer fully connected structure, with the hidden layers using the Rectified Linear Unit activation function, and the output layer not using an activation function to allow the Q-value to take any real value. To improve training stability, the system maintains a target network copy of both the actor and critic networks. The parameters of the target network are slowly updated to track the parameters of the main network using a soft update method. The soft update coefficients are determined based on training stability requirements and are typically set to small values.

[0067] Step 4.5, Reward Function Design; Based on the optimization objective and constraints of color temperature adjustment, a reward function for the reinforcement learning model is designed to provide optimization guidance for model training.

[0068] Specifically, the reward function needs to comprehensively consider multiple optimization objectives for color temperature adjustment, including user comfort, rhythm matching, adjustment smoothness, and energy efficiency. The user comfort reward component is calculated based on explicit user feedback or implicit behavioral indicators. Explicit feedback is collected through the user interface, gathering user satisfaction ratings for the current lighting effect. Implicit behavioral indicators include whether the user actively adjusts the color temperature, changes in blink frequency, and changes in the user's activity level; these indicators indirectly reflect the user's acceptance of the current color temperature. The rhythm matching reward component is calculated based on the deviation between the current color temperature and the time-rhythm reference color temperature; the smaller the deviation, the higher the reward, encouraging the model to output a color temperature value that conforms to the human circadian rhythm. The adjustment smoothness reward component is calculated based on the amplitude of color temperature changes between adjacent time steps; the smaller the amplitude of change, the higher the reward, penalizing drastic fluctuations in color temperature. The energy efficiency reward component is calculated based on LED driver power, encouraging reduced energy consumption while meeting lighting requirements. Each reward component is assigned different weights according to its importance, and a weighted sum is used to obtain the comprehensive reward value. The weight parameters of the reward function can be adjusted according to actual application needs to balance the relationship between the various optimization objectives.

[0069] Step 4.6: Prioritize experience replay and model training; Based on the state-action-reward-next state samples collected during system operation, a priority experience replay mechanism is used for sample storage and sampling. The parameters of the actor network and the critic network are optimized by gradient descent to obtain the trained decision model.

[0070] Specifically, the experience replay mechanism stores samples generated during system operation in a replay buffer. During training, batches of samples are randomly sampled from the buffer for learning, breaking the temporal correlation between samples and improving training stability. Prioritized experience replay introduces a priority mechanism based on random sampling. The sampling probability of a sample is proportional to its temporal difference (TD) error. Samples with larger TD errors indicate that the model's prediction of that sample has a larger bias; prioritizing these samples accelerates model learning. To compensate for the bias introduced by priority sampling, importance sampling weights are used to correct the gradient. Model training employs batch gradient descent. Each time, a batch of samples is sampled from the replay buffer, updating the critic network parameters. The training objective of the critic network is to minimize the mean squared error between the predicted Q-value and the target Q-value, which is calculated by the target network. The actor network parameters are also updated; the training objective of the actor network is to maximize the critic network's Q-value evaluation of the actor's output action. The training process is performed at a preset frequency, continuously optimizing model performance without affecting real-time decision-making.

[0071] Step 4.7, color temperature decision output and confidence calculation; Based on the current decision-making model and the corresponding decision results, the decision confidence level is calculated and the final optimal target color temperature value is output, thus obtaining the decision results for subsequent execution control.

[0072] Specifically, when the system is in rule engine decision mode, the optimal target color temperature value is directly adopted from the rule engine output, and the decision confidence is calculated based on the reliability score of the input data and the rule matching degree. When the system is in reinforcement learning model decision mode, the current state vector is input into the actor network for forward inference to obtain the action output value. The action value is then added to the current color temperature value and limited to obtain the optimal target color temperature value. The calculation of decision confidence comprehensively considers the following factors: the reliability score of the input data, reflecting the quality of the decision basis; the Q-value evaluation of the current state-action pair by the critic network, where a higher Q-value indicates greater confidence in the decision; and the degree of matching between the current state and the training sample distribution, obtained by calculating the similarity between the current state vector and historical states, where a higher degree of matching indicates richer decision-making experience for that type of state. The decision confidence is obtained by weighted summation and normalization of each factor, with a value ranging from zero to one. The optimal target color temperature value and the decision confidence are packaged and output for use by the subsequent gradient transition control module.

[0073] Furthermore, since a single actor-critic network may suffer from Q-value overestimation in complex environments, leading to unstable decision-making strategies, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm can be used to replace the basic actor-critic algorithm. The aim is to alleviate the Q-value overestimation problem and improve the stability and final performance of the decision-making strategy through a dual-critic network and a delayed policy update mechanism. Specifically, the core improvements of the TD3 algorithm include three aspects: First, it adopts a dual-critic network architecture, training two independent critic networks simultaneously and taking the smaller of the two network outputs when calculating the target Q-value, effectively suppressing the tendency for Q-value overestimation. Second, it introduces a target policy smoothing mechanism, adding clipped Gaussian noise to the target action when calculating the target Q-value, making the Q-value estimation more robust to small changes in the action. Third, it employs a delayed policy update mechanism, updating the actor network only after the critic network has been updated several times, ensuring that the actor network is optimized based on a more accurate Q-value estimate. Using the TD3 algorithm can improve the training stability and final decision quality of reinforcement learning models on color temperature decision-making tasks.

[0074] Step 5: Obtain the current color temperature value, determine the transition level based on the optimal target color temperature value and the current color temperature value, adjust the transition time based on the decision confidence level, and generate a color temperature gradient trajectory sequence using S-curve interpolation; Specifically, the following steps are included: Step 5.1, color temperature difference calculation and transition level determination; Based on the optimal target color temperature value and the current LED output color temperature value, the difference in color temperature adjustment is calculated. The transition level is determined according to the difference and the preset level division threshold, and the transition level flag and corresponding transition parameters are obtained.

[0075] Specifically, the current actual output color temperature value is obtained from the LED driver control module. This value is measured by a color temperature feedback sensor or calculated based on the current driving parameters. The difference between the optimal target color temperature value and the current color temperature value is calculated. The absolute value of the difference represents the required adjustment range of the color temperature, and the sign of the difference indicates the direction of adjustment: a positive value indicates that the color temperature needs to be increased, and a negative value indicates that the color temperature needs to be decreased. Based on the magnitude of the absolute value of the difference, the transition process is divided into three levels: when the absolute value of the difference is less than the first threshold, it is determined to be the fine adjustment level, indicating that the color temperature change range is very small, and a faster transition speed can be used; when the absolute value of the difference is between the first and second thresholds, it is determined to be the standard transition level, indicating that the color temperature change range is moderate, and a standard transition speed is used; when the absolute value of the difference is greater than the second threshold, it is determined to be the large adjustment level, indicating that the color temperature change range is large, and a slower transition speed is required to avoid visual discomfort. Each level corresponds to a different transition time parameter. The specific values ​​of the transition time parameter are determined based on the results of visual comfort research and actual application testing. The fine adjustment level has the shortest transition time, and the large adjustment level has the longest transition time. The preferred values ​​of the first and second thresholds are set based on the research results of the human eye's perception threshold for color temperature changes. For example, the first threshold can be set in the range of 100K to 300K, and the second threshold can be set in the range of 500K to 1000K. The specific values ​​can be adjusted according to the actual application scenario and the characteristics of the user group. For patients with high visual sensitivity, a smaller threshold can be used to provide more detailed graded control.

[0076] Step 5.2, Confidence-adaptive transition speed adjustment; Based on the decision confidence level output in step 4, the transition time parameter obtained from the transition level determination is adaptively adjusted to obtain the adjusted actual transition time, thereby realizing the linkage control between decision confidence level and transition speed.

[0077] Specifically, decision confidence reflects the reliability of the color temperature decision. A high confidence level indicates the system has strong confidence in the current decision and can use a faster transition speed to reach the target color temperature as quickly as possible. A low confidence level indicates some uncertainty in the decision result, requiring a slower transition speed to allow the system more time for decision correction and avoid drastic color temperature fluctuations caused by erroneous decisions. The transition time is adjusted using a linear mapping method, mapping the decision confidence level to an adjustment coefficient for the transition time. The range of the adjustment coefficient is determined based on the allowable adjustment range of the transition speed. When the confidence level reaches its highest value, the adjustment coefficient is at its minimum, resulting in the shortest transition time; when the confidence level reaches its lowest acceptable value, the adjustment coefficient is at its maximum, resulting in the longest transition time. The actual transition time is obtained by multiplying the base transition time corresponding to the transition level by the adjustment coefficient. For visually sensitive patients, the system can be configured with a more conservative adjustment strategy, maintaining a longer transition time even at high confidence levels to minimize the impact of color temperature changes on patients.

[0078] Step 5.3: Generation of S-curve interpolation trajectory; Based on the current color temperature value, the optimal target color temperature value, and the actual transition time, an S-curve interpolation algorithm is used to generate a color temperature gradient trajectory, resulting in a sequence of trajectory points that smoothly transition from the current color temperature to the optimal target color temperature.

[0079] Specifically, the S-curve, also known as the Sigmoid curve, is characterized by slow changes at the beginning and end, and faster changes in the middle. This characteristic conforms to the human eye's adaptation to changes in light, providing the most comfortable visual transition experience. The trajectory generation process calculates the number of trajectory points based on the actual transition time and the system control cycle. The number of trajectory points equals the transition time divided by the control cycle and rounded down. For each trajectory point, a normalized time parameter is calculated, ranging from zero to one, representing the current position relative to the entire transition process. The normalized time parameter is substituted into an S-curve function to calculate the normalized color temperature value. The S-curve function uses a logistic function, and its output is also a normalized value ranging from zero to one, representing the current color temperature relative to the initial color temperature and the target color temperature. The normalized color temperature value is then converted to the actual color temperature value through a linear mapping. The mapping relationship is: the actual color temperature equals the initial color temperature plus the normalized color temperature value multiplied by the color temperature difference. The color temperature values ​​of each trajectory point are organized into a trajectory sequence according to time order. The first point in the sequence is the current color temperature value, and the last point is the target color temperature value. The intermediate points transition smoothly according to the S-curve pattern.

[0080] Step 5.4, trajectory smoothness verification and correction; Based on the generated color temperature gradient trajectory sequence, a smoothness verification method is used to check whether the trajectory meets the visual comfort requirements. Trajectory segments that do not meet the requirements are corrected to obtain the final trajectory sequence that passes the verification.

[0081] Specifically, the smoothness verification mainly checks the following indicators: whether the rate of color temperature change between adjacent trajectory points exceeds the preset maximum rate of change threshold; an excessive rate of change will cause the color temperature change to be clearly perceived by the human eye; whether there are abrupt changes in the second derivative of the trajectory; abrupt changes in the second derivative indicate a jump in the acceleration of color temperature change, which may cause visual discomfort; whether the overall trajectory has non-monotonic changes, i.e., whether the color temperature first rises and then falls or first falls and then rises; non-monotonic changes increase the burden on visual adaptation; for trajectory segments with excessive rate of change, correction is made by increasing the density of trajectory points in that segment, decomposing the color temperature change that was originally completed in one step into multiple steps; for abrupt changes in the second derivative, correction is made by local smoothing filtering, using the weighted average of adjacent points to replace the original value of the abrupt change point; for non-monotonic changes, it is checked whether it is due to the target color temperature being updated during the transition; if so, the trajectory is regenerated; otherwise, it is considered an abnormal situation and an alarm is triggered; the verified trajectory sequence is used as the final output for the LED driver control module to execute.

[0082] Step 5.5: The trajectory sequence output is synchronized with the execution. Based on the verified color temperature gradient trajectory sequence, an execution synchronization mechanism with the LED driver control module is established. The trajectory data is output point by point according to the preset control cycle, resulting in a trajectory output stream that is strictly aligned with the execution timing.

[0083] Specifically, after the trajectory sequence is generated, it is stored in the trajectory buffer. The system reads trajectory points from the buffer sequentially according to the control cycle and outputs them to the LED driver control module. To ensure the timing accuracy of trajectory execution, a timer interrupt mechanism is used to trigger the output of trajectory points, with the timer period strictly consistent with the system control cycle. Each time a timer interrupt is triggered, the trajectory point to be output is read from the trajectory buffer, and its color temperature value is used as the target input for LED driver control. At the same time, the remaining number of points in the trajectory buffer is checked. When the remaining number of points is lower than a preset threshold, a notification is sent to the decision module that the trajectory is about to be completed, triggering a new round of decision-making and trajectory generation. During trajectory execution, if the decision module outputs a new optimal target color temperature value, the system needs to determine whether to interrupt the current trajectory execution to create a new trajectory. The judgment criteria include: whether the difference between the new optimal target color temperature and the current trajectory endpoint exceeds a preset threshold, and whether the current trajectory execution progress has exceeded a preset proportion. If it is determined that a trajectory switch is needed, a new trajectory is regenerated starting from the current actual color temperature value to ensure the smoothness of the trajectory switching process.

[0084] Furthermore, since the standard S-curve has a fixed shape, it is difficult to personalize it for the visual sensitivity of different user groups. A generalized S-curve with adjustable parameters can be used instead of the standard S-curve. The aim is to personalize the transition trajectory by adjusting the curve shape parameters, providing the most suitable transition experience for users with different sensitivities. Specifically, the generalized S-curve introduces shape control parameters based on the standard logistic function. By adjusting the parameter values, the steepness and inflection point positions of the curve can be changed. The value of the shape parameter is determined according to the user's visual sensitivity level. For patients with high visual sensitivity, a smaller shape parameter value is used to make the curve smoother and the buffer between the start and end stages longer; for users with normal visual sensitivity, a larger shape parameter value is used to make the curve relatively steeper, shortening the transition time while ensuring comfort. The user's visual sensitivity level can be obtained from medical records or automatically adjusted based on the user's feedback on different transition speeds through the system's adaptive learning mechanism. The introduction of the generalized S-curve makes the generation of the transition trajectory more flexible and adaptable, meeting the personalized needs of different user groups.

[0085] Step 6: Calculate the target driving duty cycle of warm and cool LEDs based on the color temperature gradient trajectory sequence, generate a driving control signal through pulse width modulation, and modulate and calibrate the driving control signal to obtain the LED color temperature adjustment strategy. Specifically, the following steps are included: Step 6.1, color temperature-driving parameter mapping calculation; Based on the current target color temperature point in the color temperature gradient trajectory sequence, the theoretical driving parameters of cool and warm LEDs are calculated using the color temperature-driving parameter mapping model, and the target driving duty cycle of cool white LEDs and warm white LEDs is obtained.

[0086] Specifically, the system employs a dual-color-temperature LED light source architecture, achieving continuous adjustment of color temperature by mixing the outputs of cool white LEDs and warm white LEDs. Cool white LEDs have a higher nominal color temperature, while warm white LEDs have a lower nominal color temperature; mixing them in different proportions yields any color temperature value between the two. A color temperature-drive parameter mapping model describes the correspondence between the trajectory color temperature value and the drive duty cycle of the cool and warm LEDs. This model is established through factory calibration or online learning. The mapping calculation, based on the trajectory color temperature value and the nominal color temperature values ​​of the cool and warm LEDs, uses a color temperature mixing formula to calculate the theoretical luminous flux ratio of the cool and warm LEDs. Color temperature mixing follows a weighted average rule; the mixed color temperature equals the weighted average of the color temperatures of each light source according to luminous flux. After obtaining the luminous flux ratio, the required drive current ratio to achieve this ratio is calculated using the luminous flux-drive current characteristic curves of the cool and warm LEDs. Based on the linear relationship between drive current and pulse width modulation duty cycle, the drive current ratio is converted into a duty cycle parameter. The mapping calculation also needs to consider the total luminous flux constraint. While maintaining the target color temperature, the absolute driving power of the warm and cool LEDs is adjusted according to the ambient illuminance requirements to achieve coordinated control of color temperature and brightness.

[0087] Step 6.2, Pulse width modulation drive signal generation; Based on the calculated target driving duty cycle of the warm and cold LEDs, the LED driving control signal is generated using the pulse width modulation (PWM) method, resulting in the modulation signal waveform used to drive the LED light source.

[0088] Specifically, pulse width modulation (PWM) is a modulation method that controls average power by changing the duty cycle of a pulse signal. It boasts advantages such as high efficiency, good linearity, and ease of digital control, and is widely used in LED driving. The system configures independent PWM channels for both cool white and warm white LEDs. The modulation frequency of each channel is determined based on the LED's response characteristics and the visual flicker perception threshold. The modulation frequency needs to be sufficiently high to avoid the human eye perceiving flicker, and is typically set to a value much higher than the human eye's flicker fusion frequency. The generation of the PWM signal is accomplished by a dedicated timer hardware module. The timer generates periodic pulses according to a preset modulation period, and the high-level duration of the pulses is controlled by the duty cycle parameter. The calculated target duty cycle value is written to the timer's comparison register, and the timer automatically generates a PWM signal corresponding to the duty cycle. To avoid current surges caused by sudden changes in duty cycle, the duty cycle is updated gradually, allowing only a preset maximum increment change in duty cycle per modulation period, gradually reaching the target duty cycle over multiple periods. The generated PWM signal is amplified by the driver circuit and output to the LED light source to control the LED's luminous intensity.

[0089] Step 6.3, Inner Loop Fast Response Control; Based on the pulse width modulation drive signal and the real-time response of the LED light source, an inner loop control strategy is adopted to achieve rapid adjustment of the drive parameters, resulting in a fast-responding color temperature control effect.

[0090] Specifically, the goal of inner-loop control is to ensure that the actual output of the LED light source can quickly track changes in the driving signal, compensating for the effects of LED response delay and non-ideal characteristics of the driving circuit. Inner-loop control employs a composite control structure of feedforward and feedback. The feedforward channel predicts the changing trend of the LED output based on the change in the target duty cycle, adjusting the driving parameters in advance to accelerate the response speed. The feedback channel collects the actual value of the LED driving current, compares it with the target current value, and adjusts the duty cycle according to the deviation to eliminate steady-state error. Feedback control uses a proportional-integral (PI) control algorithm. The proportional element generates an immediate adjustment based on the current deviation, while the integral element generates an adjustment to eliminate steady-state error based on the accumulation of the deviation. The proportional and integral coefficients are tuned according to the system's response speed and stability requirements, using either trial-and-error methods or analytical design methods based on the system model. The execution cycle of inner-loop control is synchronized with the pulse width modulation cycle, ensuring that adjustments are made based on the latest feedback information in each modulation cycle.

[0091] Step 6.4, Color temperature feedback acquisition and deviation calculation; Based on the actual output of the LED light source, a color temperature feedback sensor is used to periodically collect the actual output color temperature value, compare it with the target color temperature value of the trajectory, calculate the deviation, and obtain the color temperature deviation value and deviation trend information.

[0092] Specifically, the color temperature feedback sensor is deployed in the light-emitting direction of the LED light source, enabling direct measurement of the actual color temperature value of the LED mixed light. The sensor's measurement principle is similar to that of ambient light color temperature measurement; it collects light intensity information of different wavelengths through a multi-channel photodetector and calculates the color temperature value through chromaticity calculation. The feedback acquisition cycle is determined based on the outer loop calibration requirements, and is typically much longer than the inner loop control cycle to avoid computational burden and measurement noise from frequent acquisitions. After each acquisition of the actual color temperature value, it is compared with the current trajectory target color temperature value to calculate the color temperature deviation, which is equal to the actual color temperature minus the trajectory target color temperature. Simultaneously, the deviation values ​​from recent acquisitions are recorded, and the trend of deviation changes is analyzed to determine whether the deviation is continuously increasing, remaining stable, or gradually decreasing. This deviation trend information guides the selection of the outer loop calibration strategy. A continuously increasing deviation trend indicates a system drift problem, requiring stronger calibration; a stable deviation indicates a fixed bias, which can be eliminated through a one-time calibration; a gradually decreasing deviation indicates that the system is converging on its own, and the calibration frequency can be appropriately reduced.

[0093] Step 6.5, Incremental calibration compensation for the outer loop; Based on the color temperature deviation value and deviation trend information, an incremental calibration strategy is adopted to correct the color temperature-driving parameter mapping model online, and the calibrated mapping parameters are obtained to compensate for the color temperature deviation caused by factors such as LED aging and temperature drift.

[0094] Specifically, LED light sources experience light decay and color drift during long-term use, meaning the luminous flux gradually decreases and the color temperature gradually shifts, causing the original color temperature-driving parameter mapping relationship to become inaccurate. Furthermore, the photoelectric characteristics of LEDs are affected by temperature; changes in ambient temperature or the LED's own heat generation can cause color temperature output drift. The purpose of outer-loop calibration is to compensate for the effects of these factors by correcting the mapping parameters online, ensuring the accuracy of color temperature output during long-term use. The core idea of ​​the incremental calibration strategy is to make only small corrections to the mapping parameters each time, avoiding the introduction of new instability factors through large corrections. The calibration calculation calculates the parameter correction amount required to eliminate the deviation based on the current color temperature deviation value and the sensitivity coefficient of the mapping model. The sensitivity coefficient represents the color temperature change caused by a unit change in the mapping parameters, obtained by locally linearizing the mapping model.

[0095] To ensure the stability of the calibration process, the calculated parameter corrections are limited to restrict the maximum magnitude of a single correction. The limited correction is then added to the current mapping parameters to obtain the new calibrated parameter values. The calibrated parameters take effect immediately and are used for subsequent color temperature-drive parameter mapping calculations. The system also records the calibration history; when the cumulative calibration amount exceeds a preset threshold, it prompts for system maintenance or LED light source replacement.

[0096] Step 6.6, Calibration parameter persistence and recovery; Based on the mapping parameter correction values ​​obtained from the outer loop calibration, a parameter persistence mechanism is used to save the calibration results to non-volatile memory, resulting in a calibration parameter backup that can be recovered after a system restart.

[0097] Specifically, the parameter correction values ​​accumulated during incremental calibration reflect the actual changes in the characteristics of the LED light source. This information needs to be retained after the system is powered off so that the calibrated parameters can be used directly upon the next startup, avoiding the need for recalibration.

[0098] Parameter persistence employs a periodic saving strategy. A parameter saving operation is triggered when the cumulative change in calibration parameters exceeds a preset threshold or when the time since the last save exceeds a preset interval. Saved information includes: key point parameters of the color temperature-duty cycle mapping curve for both warm and cool LEDs, calibration correction values ​​for each key point, calibration timestamps, and cumulative LED operating time. Parameter saving utilizes a redundant storage mechanism, simultaneously saving both current and previous version parameters to prevent parameter loss due to power outages during the saving process. Upon system startup, the saved calibration parameters are read from non-volatile memory, verified for integrity, and then loaded into working memory. If verification fails, the previous version parameters are loaded or factory default parameters are restored, and an exception log is recorded. This parameter persistence mechanism ensures the long-term effectiveness of the calibration results, avoiding the need for recalibration on every startup.

[0099] Step 6.7, Color temperature output stability monitoring and anomaly handling; Based on historical data collected through color temperature feedback, a stability monitoring method is used to evaluate the long-term stability of color temperature output, and abnormal situations detected are handled to obtain an LED color temperature adjustment strategy.

[0100] Specifically, stability monitoring assesses the system's operating status by analyzing the statistical characteristics of color temperature feedback data. Monitoring indicators include: the mean color temperature deviation, reflecting whether there is a systematic bias in the system; the standard deviation of the color temperature deviation, reflecting the degree of fluctuation in the color temperature output; the maximum color temperature deviation, reflecting whether there is an abnormally large deviation; and the rate of change of calibration parameters, reflecting the drift speed of LED characteristics. When the monitoring indicators exceed the preset normal range, the system determines that an anomaly exists and triggers the corresponding processing procedure. For cases where the mean deviation exceeds the standard, the outer loop calibration is strengthened, the calibration cycle is shortened, and the deviation elimination speed is accelerated. For cases where the standard deviation exceeds the standard, the sensor and drive circuit are checked for faults, and if necessary, the system is switched to a backup channel or degraded for operation. For cases where the rate of change of calibration parameters exceeds the standard, it is determined that the LED light source may be nearing the end of its lifespan, and a maintenance reminder notification is generated. After the anomaly handling is completed, the system returns to normal operation mode and continues to output a stable color temperature control effect.

[0101] Furthermore, since traditional proportional-integral (PI) control may experience performance degradation when system parameters change, Model Predictive Control (MPC) can be used to replace PLC for fast-response inner-loop control. The aim is to achieve multi-step anticipatory control by establishing a predictive model of the system, thereby improving adaptability and control accuracy to changes in system parameters. Specifically, the core idea of ​​MPC is to calculate the system output for several future steps based on the predictive model of the system in each control cycle. Through optimization, a control sequence that best approximates the predicted output to the target trajectory is found, and the first control variable of the control sequence is executed. This process is repeated in the next cycle. The predictive model uses a discrete state-space model of the LED drive system, and the model parameters are estimated from historical input-output data using system identification methods. The optimization objective function includes a term for the deviation between the predicted output and the target trajectory, and a penalty term for changes in the control variable. By adjusting the weights of these two terms, tracking performance and control smoothness can be balanced. MPC can naturally handle system constraints, such as upper and lower limits of the drive duty cycle and the duty cycle rate of change constraints, ensuring that the control variable remains within the allowable range. When system parameters change, model predictive control can adaptively adjust the control strategy and maintain good control performance by updating the predictive model parameters online.

[0102] An adaptive color temperature adjustment system for LED lamps, used to execute the aforementioned adaptive color temperature adjustment method for LED lamps, such as... Figure 3 As shown, it includes: The environmental perception module is used to acquire raw spectral signal data and ambient light color temperature values ​​and preprocess them to obtain an environmental perception dataset. The user status recognition module extracts user physiological status and behavioral features based on the environmental perception dataset, and uses multimodal fusion to identify user status categories and obtain user status category labels. The feature fusion module is used to standardize the environmental perception dataset and user state category labels, and uses an adaptive weight allocation strategy and attention mechanism to fuse features to obtain a comprehensive feature vector. The decision optimization module makes color temperature decisions based on comprehensive feature vectors, continuously optimizes the optimal color temperature decision, and obtains the optimal target color temperature value and decision confidence. The trajectory generation module is used to obtain the current color temperature value, determine the transition level based on the optimal target color temperature value and the current color temperature value, adjust the transition time based on the decision confidence level, and generate a color temperature gradient trajectory sequence using S-curve interpolation; The drive control module calculates the target drive duty cycle of warm and cool LEDs based on the color temperature gradient trajectory sequence, generates a drive control signal through pulse width modulation, and modulates and calibrates the drive control signal to obtain the LED color temperature adjustment strategy.

[0103] In one embodiment of the present invention, a specific example is provided: A field test was conducted on a rehabilitation ward of an ophthalmology hospital from March to May 2025, lasting 60 days. The test area was a single-patient rehabilitation ward, approximately 20 square meters in size, with a ceiling height of 2.8 meters, facing southeast, and equipped with standard medical lighting. During the test, an LED color temperature adaptive adjustment system based on the method of this invention was deployed. A 6-channel multispectral sensor array (3 nodes) was installed on the ceiling of the ward, an infrared thermal imaging sensor (resolution 320×240 pixels, frame rate 15fps) was installed 1.5 meters above the head of the bed, and a 77GHz millimeter-wave radar sensor was installed at the foot of the bed. The ward housed 3 patients recovering from cataract surgery and 3 patients undergoing glaucoma treatment, with each patient observed for 10 days. The system sampling period was set to 5 seconds, continuously collecting ambient light information and patient status information.

[0104] Table 1 shows an example of environmental perception data collected by the system on a typical workday (the vernal equinox, a sunny day): Table 1: Example of environmental perception data collected on a typical workday (Spring Equinox, sunny);

[0105] Table 2 shows examples of user status data collected by the system: Table 2: Examples of collected user status data;

[0106] Based on the collected data, the system automatically generates the optimal target color temperature value adapted to the patient's condition and environmental conditions through multi-source data fusion and reinforcement learning decision-making. It then achieves smooth color temperature adjustment through graded gradual transition control. Test results show that the system's output LED color temperature value remains within the range of 3800K to 5200K during the day and decreases to the range of 2400K to 3000K at night. The average time for the color temperature transition is 120 to 180 seconds, and the patient's perception score for color temperature changes is less than 1.5 out of 10, indicating that the color temperature adjustment process is essentially imperceptible. Throughout the test, the patient did not experience any eye discomfort symptoms caused by light stimulation, and the lighting environment remained comfortable, effectively supporting the patient's recovery process.

[0107] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An adaptive adjustment method for the color temperature of an LED lamp, characterized in that, Includes the following steps: The raw spectral signal data and ambient light color temperature values ​​are acquired and preprocessed to obtain the environmental perception dataset. Based on the environmental perception dataset, user physiological state and behavioral features are extracted, and multimodal fusion is used to identify user state categories to obtain user state category labels; The environmental perception dataset and user state category labels are standardized, and feature fusion is performed using an adaptive weight allocation strategy and an attention mechanism to obtain a comprehensive feature vector. Color temperature decision-making is based on comprehensive feature vectors, and the optimal color temperature decision is continuously optimized to obtain the optimal target color temperature value and decision confidence. Specifically, this includes: determining whether the number of samples in the experience replay buffer, the number of pre-training rounds of the reinforcement learning model, and the decision performance on the validation set have reached preset thresholds; when all conditions are met, switching to the reinforcement learning model decision mode; otherwise, maintaining the rule engine decision mode. When the system is in rule engine decision mode, color temperature decision-making is performed. The rule engine includes basic color temperature rules, environmental compensation rules, user state adjustment rules, and sensitivity protection rules. The outputs of each rule are weighted and summed to obtain the optimal target color temperature value. When the system is in reinforcement learning model decision mode, a system is defined that includes comprehensive feature vectors, when... The state space of the current LED output color temperature value, recent color temperature adjustment history, and reliability scores of each data source is used to define a continuous action space representing the target color temperature adjustment amount. A reinforcement learning model is designed using an actor-critic architecture. The actor network outputs action values ​​based on the current state, and the critic network evaluates the value of the state-action pair. A priority experience replay mechanism is used for sample storage and sampling, and the parameters of the actor network and the critic network are optimized using gradient descent. The current state vector is input into the actor network to obtain the action value. The action value is added to the current color temperature value and the amplitude is limited to obtain the optimal target color temperature value. The decision confidence is obtained by comprehensively considering the reliability score of the input data, the Q-value evaluation of the critic network, and the matching degree between the current state and the training sample distribution. The process involves obtaining the current color temperature value, determining the transition level based on the optimal target color temperature value and the current color temperature value, adjusting the transition time based on the decision confidence level, and generating a color temperature gradient trajectory sequence using S-curve interpolation. Specifically, this includes: calculating the difference between the optimal target color temperature value and the current color temperature value; determining a fine adjustment level when the absolute value of the difference is less than a first threshold; a standard transition level when the absolute value of the difference is between the first and second thresholds; and a large adjustment level when the absolute value of the difference is greater than the second threshold, thus obtaining the transition level and the corresponding transition time parameters; and mapping the decision confidence level to the transition time. The adjustment factor for time is multiplied by the base transition time corresponding to the transition level to obtain the actual transition time. The number of trajectory points is calculated based on the actual transition time and the system control cycle. The normalized time parameter for each trajectory point is calculated, and the normalized color temperature value is calculated. This value is then converted to the actual color temperature value through linear mapping to obtain the color temperature gradient trajectory sequence. The smoothness of the color temperature gradient trajectory sequence is verified by checking the color temperature change rate of adjacent trajectory points, the abrupt change point of the second derivative of the trajectory, and the monotonicity of the overall trajectory. Trajectory segments that do not meet the requirements are corrected to obtain the final trajectory sequence. The target driving duty cycle of warm and cool LEDs is calculated based on the color temperature gradient trajectory sequence. A driving control signal is generated by pulse width modulation. The driving control signal is then modulated and calibrated to obtain the LED color temperature adjustment strategy.

2. The adaptive adjustment method for LED lamp color temperature according to claim 1, characterized in that, The steps for obtaining the environmental awareness dataset include: Deploy multispectral sensor nodes and adopt a distributed acquisition architecture to synchronously measure ambient light and obtain raw spectral signal datasets; Dark current correction and sliding window mid-range filtering are performed on the data of each channel to remove sampling points whose rate of change exceeds the preset anomaly judgment threshold; The spectral power distribution curve is estimated by spectral reconstruction algorithm, the tristimulus values ​​and chromaticity coordinates are calculated, and the blackbody radiation trajectory point closest to the chromaticity coordinates of the light source is found on the CIE chromaticity diagram to obtain the ambient light color temperature value. Calculate the solar altitude angle, and obtain the time rhythm reference color temperature value by looking up the mapping table between the solar altitude angle and the natural light color temperature; The ambient light color temperature value, spectral signal intensity value, and time rhythm reference color temperature value are encapsulated to obtain the environmental perception dataset.

3. The adaptive adjustment method for LED lamp color temperature according to claim 1, characterized in that, The steps for obtaining the user status category label include: Infrared thermal imaging sensors are used to acquire thermal images of the user's facial region. Non-uniformity correction, temperature calibration, and spatial filtering are performed on the thermal image sequence. Background modeling and foreground segmentation methods are used to extract the user region. A cascaded classifier is used to extract facial thermal distribution features and eye state features to obtain the infrared modal user state feature vector. Frequency-modulated continuous wave millimeter-wave radar is used to detect the target area. The radar echo signal is mixed and processed by fast Fourier transform. The constant false alarm rate detection algorithm is used to extract motion features and micro-motion features to obtain the user state feature vector of the radar mode. The data quality of infrared and radar modes is evaluated, confidence scores for each mode are generated, contribution weights for each mode are determined, and feature vectors of each mode are concatenated to obtain the fused user state feature vector and mode fusion confidence. The fused user state feature vector is input into the pre-trained classification model, and the category with the highest probability is selected to obtain the user state category label and classification confidence.

4. The adaptive adjustment method for LED lamp color temperature according to claim 1, characterized in that, The steps for obtaining the comprehensive feature vector include: The environmental perception dataset and user state category labels are resampled onto a unified time grid, and each feature dimension is standardized. Reliability assessments are conducted based on the quality metrics of each data source and the characteristics of the current scenario to obtain a reliability score for each data source. Based on the reliability score and the judgment results of the dominant factors in the scenario, the fusion weight values ​​of environmental characteristics, time characteristics and user characteristics are determined; The various standardized features are scaled and concatenated to form the input feature matrix. The attention weight matrix is ​​calculated, and the attention weight matrix is ​​multiplied by the value vector to obtain the fused high-dimensional feature representation. Principal component analysis is used to reduce the dimensionality of high-dimensional features, resulting in a comprehensive feature vector.

5. The adaptive adjustment method for LED lamp color temperature according to claim 1, characterized in that, The method of generating a color temperature gradient trajectory sequence using S-curve interpolation also includes: A generalized S-curve with adjustable parameters is used. The generalized S-curve introduces shape control parameters on the basis of the standard logistic function. The value of the shape parameter is determined according to the user's visual sensitivity level.

6. The adaptive adjustment method for LED lamp color temperature according to claim 1, characterized in that, The steps for obtaining the LED color temperature adjustment strategy include: Based on the color temperature gradient trajectory sequence, the theoretical luminous flux ratio of warm and cool LEDs is calculated using a color temperature driving parameter mapping model, and the driving current ratio is calculated by combining the luminous flux driving current characteristic curve; the LED driving control signal is generated using a pulse width modulation method. A composite control structure of feedforward and feedback is adopted to achieve fast response control in the inner loop. The feedforward channel predicts the changing trend of LED output based on the change of the target duty cycle, and the feedback channel uses a proportional-integral control algorithm to adjust the duty cycle according to the deviation. A color temperature feedback sensor is used to periodically collect the actual output color temperature value and compare it with the color temperature value of the trajectory target to calculate the color temperature deviation value. The color temperature driving parameter mapping model is corrected online. The parameter correction amount is calculated, and the parameter correction amount is superimposed on the current mapping parameter after being limited to obtain the calibrated mapping parameter.

7. The adaptive adjustment method for LED lamp color temperature according to claim 6, characterized in that, The method of using a feedforward and feedback composite control structure to achieve fast response control of the inner loop also includes: Model predictive control is adopted, and the system output of the LED driving system is calculated based on the discrete state-space prediction model of the LED driving system for the future preset number of steps. The control sequence is obtained by solving the objective function. The objective function includes the deviation term between the predicted output and the target trajectory and the penalty term for the change of control quantity.

8. An adaptive color temperature adjustment system for LED lamps, characterized in that, An adaptive adjustment method for LED lamp color temperature according to any one of claims 1-7 includes: The environmental perception module is used to acquire raw spectral signal data and ambient light color temperature values ​​and preprocess them to obtain an environmental perception dataset. The user status recognition module extracts user physiological status and behavioral features based on the environmental perception dataset, and uses multimodal fusion to identify user status categories and obtain user status category labels. The feature fusion module is used to standardize the environmental perception dataset and user state category labels, and uses an adaptive weight allocation strategy and attention mechanism to fuse features to obtain a comprehensive feature vector. The decision optimization module makes color temperature decisions based on comprehensive feature vectors, continuously optimizes the optimal color temperature decision, and obtains the optimal target color temperature value and decision confidence. The trajectory generation module is used to obtain the current color temperature value, determine the transition level based on the optimal target color temperature value and the current color temperature value, adjust the transition time based on the decision confidence level, and generate a color temperature gradient trajectory sequence using S-curve interpolation; The drive control module calculates the target drive duty cycle of warm and cool LEDs based on the color temperature gradient trajectory sequence, generates a drive control signal through pulse width modulation, and modulates and calibrates the drive control signal to obtain the LED color temperature adjustment strategy.

Images