An intelligent LED spectral regulation system based on multi-dimensional data fusion

Abstract

The present invention relates to the field of intelligent spectral regulation technology, and discloses an intelligent spectral regulation system based on multi-modal data fusion and dynamic optimization, including: a multi-modal data acquisition module, a real-time data processing module, a dynamic weight allocation algorithm, a deep reinforcement learning regulation algorithm, and a distributed collaborative optimization module. The system realizes the real-time fusion and dynamic priority adjustment of multi-dimensional data through the acquisition of multi-modal data (ambient spectrum, temperature and humidity, carbon dioxide concentration, user behavior) and dynamic weight allocation based on LSTM; combined with a deep reinforcement learning model, using a weighted reward function of spectral output error and energy consumption, it realizes the dynamic optimization of spectral regulation strategies, balancing the requirements of efficient light supplementation and low energy consumption; adopts a distributed optimization algorithm that combines edge computing and cloud collaboration to dynamically allocate computing resources, greatly improving the task processing efficiency and reducing the system power consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent spectrum regulation, and specifically relates to an intelligent LED spectrum regulation system based on multi-dimensional data fusion, which is used to achieve high-precision spectrum regulation in scenarios such as agricultural supplementary lighting, medical phototherapy, and intelligent lighting. Background Art

[0002] In current technologies, spectrum regulation systems usually rely on a single sensor to collect environmental data (such as temperature and humidity or spectrum intensity) for static supplementary lighting regulation. For example, after obtaining environmental spectrum data through a light sensor, preset regulation rules are used to control the spectrum output. However, this method has the following deficiencies:

[0003] 1. The dynamic coupling relationship between environmental spectrum and temperature and humidity, carbon dioxide concentration, and user behavior is not fully considered, resulting in the lack of accuracy and flexibility of spectrum output and being unable to adapt to the complex and changeable scenario requirements.

[0004] 2. The regulation mechanism with fixed weights is difficult to adjust the data priority according to real-time environmental changes, easily causing lag in regulation effects or waste of resources.

[0005] 3. The system usually adopts a single-node computing mode, resulting in response delays for tasks with high real-time requirements (such as agricultural supplementary lighting), and high energy consumption during long-term operation.

[0006] To overcome the above problems, dynamic weight adjustment methods or optimization schemes based on distributed computing have also been tried in existing technologies. However, these methods such as dynamic weight adjustment usually only target a single data source or a simple model, failing to achieve the fusion of multi-modal data and real-time priority adjustment, resulting in inaccuracies in weight calculation between multi-source data; distributed computing schemes usually migrate all tasks to the cloud for processing, increasing communication delays and being unable to meet the high real-time requirements of spectrum regulation; there is a lack of systematic algorithm design for multi-objective optimization (such as spectrum regulation accuracy and energy consumption control), making it difficult to balance resource allocation and energy consumption control.

[0007] Therefore, how to achieve the efficient fusion of multi-modal data, combine reinforcement learning and distributed collaborative optimization technologies, dynamically adjust the spectrum output weight, and balance system power consumption and regulation accuracy has become the technical problem to be solved by the present invention. Summary of the Invention

[0008] The present invention provides an intelligent LED spectrum regulation system, system, and computer-readable storage medium based on multi-dimensional data fusion, and its main purpose is to solve the problems of insufficient real-time performance, low resource allocation efficiency, and insufficient multi-modal data fusion.

[0009] To achieve the above object, the present invention provides an intelligent LED spectral regulation system based on multi-dimensional data fusion, which system includes:

[0010] A multi-dimensional data acquisition module, configured to acquire environmental spectral data, temperature and humidity data, carbon dioxide concentration data and user behavior data in real time, wherein: the environmental spectral data is acquired by a high-precision spectral sensor; the temperature and humidity data is acquired by a digital sensor; the carbon dioxide concentration data is measured by an infrared gas sensor; the user behavior data is analyzed by a camera in combination with an image recognition algorithm;

[0011] A data fusion processing module, connected to the multi-dimensional data acquisition module, configured to perform standardized preprocessing and dynamic fusion on the acquired multi-dimensional data, including: performing dimensionless processing on data of different physical quantities by using a standardization algorithm based on fuzzy logic; by means of a multi-modal adaptive weight allocation algorithm, calculating a weight coefficient of a spectral output parameter according to the time-series change characteristics of environmental spectral data, temperature and humidity data, carbon dioxide concentration data and user behavior data, and the weight coefficient is calculated by the following formula:

[0012]

[0013] Wherein: is the weight of the th type of data in spectral regulation; is the real-time value of the th type of acquired data; is a dynamically adjusted factor obtained by fitting based on historical data, and is obtained by training a time series prediction model LSTM in combination with historical data; is the dimension of the acquired data; represents the th spectral output coefficient at a time point; is the value of the th type of data;

[0014] A spectral regulation algorithm module, connected to the data fusion processing module, configured to perform real-time regulation on the LED spectral output according to the data fusion result, including: adopting a spectral optimization model based on deep reinforcement learning, with the model input being the weight coefficient and the spectral output target value after data fusion; during the model training process, the following objective function is adopted for dynamic optimization:

[0015]

[0016] Wherein: is the optimization target of the policy network; is the spectral regulation strategy based on the current network parameters ; is time Spectral output effect score at a moment; is the reward discount factor, which controls the influence weight of future rewards on the current strategy;

[0017] The LED spectral output module, connected to the spectral regulation algorithm module, is used to dynamically adjust the LED spectral output according to the optimization result, including: realizing spectral output through a multi-channel adjustable spectral LED light source; configuring a fast-response current driver to dynamically adjust the current value of each spectral channel according to the target parameters output by the spectral regulation algorithm module to achieve precise regulation of the spectrum.

[0018] The system collaborative control module, connected to the above-mentioned modules, is used to coordinate the operation of the multi-dimensional data acquisition module, data fusion processing module, spectral regulation algorithm module, and LED spectral output module, including: configuring an event-driven mechanism for dynamically triggering data acquisition and processing; dynamically allocating computing resources through a distributed collaborative optimization algorithm to reduce system power consumption and improve real-time performance.

[0019] Preferably, the fuzzy logic normalization algorithm in the data fusion processing module includes the following steps: normalizing the data of different physical quantities to the range between 0 and 1 respectively; based on the dynamic change range of the data, using a triangular membership function and a dynamically adjusted threshold interval to construct a fuzzy set to reduce the scale difference between physical quantities.

[0020] Preferably, in the multi-modal adaptive weight allocation algorithm, the dynamic adjustment factor is calculated as follows: collecting historical data in the target application scenario; performing time series fitting analysis based on the LSTM model to predict the long-term weight trend of each type of data on spectral output; using the gradient descent method to dynamically update the weight adjustment factor to optimize the real-time performance and robustness of weight calculation.

[0021] Preferably, the training process of deep reinforcement learning in the spectral regulation algorithm module includes: generating an initial sample library of spectral regulation strategies based on offline collected multi-dimensional data; using a convolutional neural network to extract the features of multi-dimensional data; constructing a reinforcement learning environment, defining the reward function as the weighted comprehensive score of the accuracy score of spectral output and energy consumption; using distributed parallel computing to accelerate the strategy optimization of deep reinforcement learning.

[0022] Preferably, the multi-channel adjustable spectral LED light source in the LED spectral output module includes: LED light sources of four spectral bands, namely red, green, blue, and white; the light sources of each band are independently powered by a constant current driver; using PWM (pulse width modulation) technology to adjust the luminous intensity and duration of each band of light source in real time to achieve dynamic synthesis of the target spectrum.

[0023] Preferably, the event-driven mechanism in the system collaborative control module includes: detecting the real-time sampling frequency of the multi-dimensional data acquisition module; triggering the data fusion processing module to recalculate the weight coefficients when the spectral output deviation exceeds a preset threshold; managing the data transmission and control instructions between multiple modules through an asynchronous queue to avoid communication blocking between modules.

[0024] Preferably, the calculation formula of the weight coefficient further incorporates the confidence parameter of the data , and the confidence parameter is determined by the following formula:

[0025]

[0026] Where: is the sensitivity adjustment factor, determined by experimental fitting; is the value of the th type of data collected in real time; is the historical mean of this data category.

[0027] Preferably, the reward discount factor in the deep reinforcement learning model is dynamically adjusted according to the following formula:

[0028]

[0029] Where: is the adjustment factor, depending on the sensitivity requirements of spectral regulation; is time at which the spectral output score; is the target score threshold, used to measure whether the spectral regulation reaches the expected effect.

[0030] Compared with the problems described in the background art, the beneficial effects of the present invention are:

[0031] 1. Through the acquisition and preprocessing of multi-modal data (environmental spectrum, temperature and humidity, carbon dioxide concentration, user behavior), combined with the dynamic weight allocation algorithm based on LSTM, the real-time fusion of data input and dynamic priority adjustment are achieved. Compared with the traditional spectral regulation methods with single or fixed weights, this solution can quickly adapt to the different influence weights of various data on the spectral output in a dynamic and complex environment, and form a real-time regulation strategy. Through the adaptive adjustment of the spectral intensity, the system can maintain the spectral output error below 5% accuracy in different application scenarios (such as agricultural supplementary lighting, medical phototherapy), and at the same time meet the real-time response requirements under changing environmental conditions.

[0032] 2. By constructing a deep reinforcement learning model based on the Actor-Critic architecture and combining a weighted reward function of spectral output error and energy consumption, the dynamic optimization of the spectral regulation strategy is achieved. Through the combination of online inference and periodic training, the system can achieve dynamic trade-offs between efficient light supplementation and energy consumption control, avoiding the limitations of traditional static algorithms in different scenarios. In addition, the multi-objective optimization mechanism in the reward function ensures the performance stability and resource utilization efficiency of the system during long-term operation by balancing spectral accuracy and energy consumption. For example, in the medical light therapy scenario with high-precision requirements, the system can automatically adjust the spectral output intensity through the reinforcement learning model while reducing energy consumption by 15%.

[0033] 3. Through the resource allocation strategy of the cooperation between edge computing and cloud servers, this technical solution solves the computational conflict problem between real-time tasks and periodic tasks. A random forest regression model is used to predict the future task load distribution, and a greedy strategy is combined to achieve dynamic sorting of task priorities, greatly improving the computational resource allocation efficiency. The actual operation results show that when the computational load of the edge node approaches the upper limit, the automatic migration of low-priority tasks shortens the real-time task response time by 40%, and the overall power consumption of the system is reduced by 15% - 20%. In addition, by regularly analyzing the operation data to optimize the task allocation ratio, the long-term adaptability and resource utilization efficiency of the system are further enhanced.

[0034] 4. A complete closed-loop process from data collection, processing, regulation to feedback is realized, and a dynamic optimization logic combining real-time data and long-term data is constructed. Through the combination of distributed collaborative optimization and deep reinforcement learning algorithms, the system realizes the dynamic interaction of data input, model training, and resource allocation, solving the defects of isolated data processing and single optimization of regulation in the existing technology. The closed-loop logic in the technical path enhances the adaptability of multi-modal data in complex environments, enabling the system to meet the requirements of multi-scene switching, such as different fields like agriculture, medical, and industry, and achieving an overall improvement in system efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the architecture diagram of the intelligent spectral regulation system of the present invention.

[0036] Figure 2 It is the data processing flow chart of the intelligent spectral regulation of the present invention.

[0037] The implementation, functional features, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0039] An embodiment of this application provides an intelligent LED spectrum regulation system based on multi-dimensional data fusion. The system includes: a multi-dimensional data acquisition module, which is used to collect environmental spectrum data, temperature and humidity data, carbon dioxide concentration data, and user behavior data in real time, where: the environmental spectrum data is collected by a high-precision spectrum sensor; the temperature and humidity data is collected by a digital sensor; the carbon dioxide concentration data is measured by an infrared gas sensor; the user behavior data is analyzed through a camera combined with an image recognition algorithm;

[0040] A data fusion processing module, connected to the multi-dimensional data acquisition module, is used to perform standardized preprocessing and dynamic fusion on the collected multi-dimensional data, including: using a standardization algorithm based on fuzzy logic to perform dimensionless processing on data of different physical quantities; through a multi-modal adaptive weight allocation algorithm, according to the time-series change characteristics of environmental spectrum data, temperature and humidity data, carbon dioxide concentration data, and user behavior data, calculate the weight coefficient of the spectrum output parameter, and the weight coefficient is calculated by the following formula:

[0041]

[0042] Where: is the weight of the type of data in spectrum regulation; is the real-time value of the type of data collected; is a dynamic adjustment factor obtained by fitting based on historical data, and is trained by combining the time series prediction model LSTM with historical data; is the dimension of the collected data; represents the spectrum output coefficient at the th time point; is the value of the

[0043] A spectrum regulation algorithm module, connected to the data fusion processing module, is used to perform real-time regulation on the LED spectrum output according to the data fusion result, including: adopting a spectrum optimization model based on deep reinforcement learning, with the input of the model being the weight coefficient and the spectrum output target value after data fusion; during the model training process, the following objective function is used for dynamic optimization:

[0044]

[0045] Where: is the optimization target of the policy network; is the spectrum regulation strategy based on the current network parameters ; is the time spectrum output effect score at the is the reward discount factor, which controls the influence weight of future rewards on the current policy;

[0046] The LED spectral output module, connected to the spectral regulation algorithm module, is used to dynamically adjust the LED spectral output according to the optimization result, including: realizing spectral output through a multi-channel tunable spectral LED light source; configuring a fast-response current driver to dynamically adjust the current value of each spectral channel of each wavelength band according to the target parameters output by the spectral regulation algorithm module, so as to achieve precise regulation of the spectrum.

[0047] The system collaborative control module, connected to the above-mentioned modules, is used to coordinate the operation of the multi-dimensional data acquisition module, data fusion processing module, spectral regulation algorithm module and LED spectral output module, including: configuring an event-driven mechanism for dynamically triggering data acquisition and processing; dynamically allocating computing resources through a distributed collaborative optimization algorithm to reduce system power consumption and improve real-time performance.

[0048] Preferably, the fuzzy logic normalization algorithm in the data fusion processing module includes the following steps: normalizing the data of different physical quantities to the range between 0 and 1 respectively; based on the dynamic change range of the data, using a triangular membership function and a dynamically adjusted threshold interval to construct a fuzzy set to reduce the scale difference between physical quantities.

[0049] Preferably, in the multi-modal adaptive weight allocation algorithm, the dynamic adjustment factor is calculated as follows: collecting historical data in the target application scenario; performing time series fitting analysis based on the LSTM model to predict the long-term weight trend of each type of data on the spectral output; using the gradient descent method to dynamically update the weight adjustment factor to optimize the real-time performance and robustness of weight calculation.

[0050] Preferably, the training process of deep reinforcement learning in the spectral regulation algorithm module includes: generating an initial sample library of spectral regulation strategies based on offline collected multi-dimensional data; using a convolutional neural network to extract the features of multi-dimensional data; constructing a reinforcement learning environment, and defining the reward function as the weighted comprehensive score of the accuracy score of spectral output and energy consumption; adopting distributed parallel computing to accelerate the strategy optimization of deep reinforcement learning.

[0051] Preferably, the multi-channel tunable spectral LED light source in the LED spectral output module includes: LED light sources of four spectral bands of red, green, blue, and white; the light sources of each band are independently powered by a constant current driver; using PWM (pulse width modulation) technology to adjust the luminous intensity and duration of each band of light source in real time to realize the dynamic synthesis of the target spectrum.

[0052] Preferably, the event-driven mechanism in the system collaborative control module includes: detecting the real-time sampling frequency of the multi-dimensional data acquisition module; triggering the data fusion processing module to recalculate the weight coefficients when the spectral output deviation exceeds a preset threshold; managing the data transmission and control instructions between multiple modules through an asynchronous queue to avoid communication blocking between modules.

[0053] Preferably, the calculation formula of the weight coefficient further incorporates the confidence parameter of the data , and the confidence parameter is determined by the following formula:

[0054]

[0055] Where: is the sensitivity adjustment factor, determined by experimental fitting; is the value of the th type of data collected in real time; is the historical mean of this data category.

[0056] Preferably, the reward discount factor in the deep reinforcement learning model is dynamically adjusted according to the following formula:

[0057]

[0058] Where: is the adjustment factor, depending on the sensitivity requirements of spectral regulation; is time the spectral output score at the moment; is the target score threshold, used to measure whether the spectral regulation reaches the expected effect.

[0059] Preferably, the user behavior data analysis steps of the multi-dimensional data acquisition module include: obtaining the user image in real time through a camera; extracting features from the image using a pre-trained convolutional neural network; judging the current user behavior category based on a support vector machine model and matching it with the spectral output requirement library to generate spectral regulation target parameters.

[0060] Preferably, the distributed collaborative optimization algorithm in the system collaborative control module includes: dynamically allocating the task computing resources of multiple modules; during the operation of the spectral regulation algorithm module, preferentially allocating edge computing nodes to process real-time data, and uploading complex computing tasks to the cloud to complete; recording the operation status of each module through system logs and optimizing the resource allocation strategy in combination with long-term operation data.

[0061] Example 1:

[0062] The multi-modal data acquisition module of this embodiment includes the following devices and data sources: Ambient spectral data acquisition unit: Equipped with a high-precision spectral sensor (wavelength range 400 - 700nm, resolution 1nm), it acquires spectral curve data of the current light environment at a period of 1 second; Temperature and humidity data acquisition unit: Uses a digital temperature and humidity sensor (sampling accuracy ±0.1°C and ±1%RH), and acquires current environmental temperature and humidity data at a frequency of once per second; Carbon dioxide concentration data acquisition unit: Equipped with an infrared gas sensor, with a monitoring range of 400 - 2000ppm and a resolution of ±5ppm, it acquires the change in CO 2 concentration in real time; User behavior data acquisition unit: Obtains real-time user behavior videos through a camera, and combines with an image classification model based on the ResNet-50 convolutional neural network to classify user behaviors into "working", "resting", or "sleeping".

[0063] After data acquisition is completed, all data enters a preprocessing process to achieve standardization and normalization: Spectral data normalization: The light intensity data collected by the spectral sensor is processed by maximum normalization according to the following formula:

[0064]

[0065] where, is the original light intensity data collected, is the maximum light intensity value within the acquisition period.

[0066] Temperature, humidity, and CO 2 Data standardization: The z-score standardization algorithm is adopted:

[0067]

[0068] where: is the original data collected; is the historical average value of this data; is the historical standard deviation of this data.

[0069] The confidence level of the user behavior classification result is directly used as an input parameter by the classification probability output by the ResNet-50 network. For example, the confidence level for classification as "working" is 0.85.

[0070] After data preprocessing is completed, the dynamic weight allocation algorithm is used to calculate the influence weights of each type of data in spectral regulation. Its core steps are as follows: Input data: Standardized spectral data, temperature and humidity data, CO 2 concentration data, and user behavior confidence values; Dynamic weight factor Generation: Use LSTM (Long Short-Term Memory Network) to perform time series analysis on historical data and predict the importance weights of current data:

[0071]

[0072] Where: is the value of the type of data at the historical time point ; is the prediction window length (take historical data within 30 minutes as input).

[0073] The dynamically adjusted weight is calculated according to the following formula:

[0074]

[0075] Where: is the real-time weight of the type of data; is the type of data after real-time standardization.

[0076] Weight dynamic adjustment mechanism: If the change amplitude of a certain type of data (such as spectrum or CO 2 concentration) in the real-time data exceeds the 10% threshold, the weight adjustment factor will be updated in real time through the gradient descent method to ensure the sensitivity of the spectrum regulation response.

[0077] The calculated weight is used to fuse all the collected data to generate the input target value of the spectrum regulation algorithm. The data fusion process is as follows. The fusion formula: The spectrum target value is calculated according to the following formula:

[0078]

[0079] Where: is the target value output by the spectrum; is the input feature of the type of data.

[0080] When the weight of the spectrum data is relatively high (such as in the agricultural scenario where spectrum accuracy is prioritized), the system preferentially meets the spectrum regulation requirements; when the user behavior confidence is relatively high (such as in the medical phototherapy scenario where the user state is prioritized), the system dynamically increases the influence weight of the user behavior data.

[0081] Suppose the system is applied to the agricultural supplementary lighting scenario: The spectrum sensor detects that the current environmental light is insufficient (the light intensity in the 400 - 500nm band is lower than 30%); the temperature and humidity sensor detects that the humidity is too high (85%RH); CO 2The concentration is 1000 ppm; the user behavior is classified as "rest".

[0082] Through dynamic weight assignment, the weight of spectral data is assigned a value of 0.6, and the weight of temperature and humidity data is 0.2, and the weight of CO 2 data is 0.15, and the weight of user behavior data is 0.05. The generated spectral target value after fusion is:

[0083]

[0084] The system adjusts the LED spectral output according to this target value to achieve dynamic light compensation and environmental adaptation.

[0085] Example 2:

[0086] This example is based on the deep reinforcement learning algorithm, and realizes intelligent spectral regulation by dynamically regulating the multi-channel LED spectral output. The hardware system includes the following components. Data acquisition module: Obtain multi-modal data such as environmental spectrum, temperature and humidity, and carbon dioxide concentration; Computing and processing unit: Equipped with an embedded chip (such as NVIDIA Jetson AGX Orin), supporting real-time inference and optimization calculation of deep learning models; Spectral output module: Equipped with a multi-channel adjustable spectral LED light source, and each channel is independently configured with a constant current driver, supporting PWM modulation; Communication module: Supports data transmission between the edge node and the cloud for model training and task assignment optimization.

[0087] The definitions of state, action, and reward function are as follows. State space ( ): The state space includes multi-modal data of the environment, such as spectral intensity, temperature and humidity, carbon dioxide concentration, and the user behavior classification result. The specific definitions are as follows:

[0088]

[0089] Among them, each variable is real-time data after normalization.

[0090] Action space ( ): The action space is defined as the PWM duty cycle adjustment value of each spectral channel LED:

[0091]

[0092] Each PWM duty cycle has a value range of [0, 1], representing the luminous intensity of the corresponding band LED.

[0093] Reward function ( ) The reward function is determined by the comprehensive score of the spectral output error and energy consumption, as follows:

[0094]

[0095] Where: is the actual spectral output value; is the target spectral output value (obtained by the weight allocation algorithm in Example 1); is the system power consumption (measured in real time by the current sensor of the spectral output module); is the power consumption impact factor, which adjusts the trade-off between spectral accuracy and energy consumption.

[0096] Also, during the training process of the reinforcement learning model, such as dataset construction: Under laboratory conditions, paired samples of different environmental variables (such as spectral intensity, temperature and humidity, CO 2 concentration, etc.) and the corresponding spectral regulation target values are collected to generate an offline training dataset. The data volume is not less than 10,000 groups, covering typical scenarios such as agricultural supplementary lighting, medical phototherapy, and intelligent lighting.

[0097] Network structure design, the reinforcement learning model is based on the Actor-Critic architecture and consists of the following two parts: Actor network: The input state space outputs the probability distribution of the action space : Where are the parameters of the Actor network. Critic network: The input state and action output the value estimate of the current policy: Where are the parameters of the Critic network.

[0098] Reward function optimization: Use the Proximal Policy Optimization (PPO) algorithm to train the model and update the parameters of the Actor and Critic networks. The objective function is: Where: is the reward discount factor (the value range is [0, 1], used to balance short-term and long-term rewards); is the time instant reward at time t. The optimization process of PPO uses the gradient descent method to limit the step size of policy update to ensure convergence stability.

[0099] Training configuration: Learning rate: 0.0003; Batch size: 64; Number of training iterations: 2000 rounds; Use a distributed training framework (such as PyTorch Distributed Data Parallel) to improve training efficiency.

[0100] In terms of the online inference and deployment of the model, its real-time inference process is as follows: Multimodal data is collected and standardized per second and then input into the reinforcement learning model; The Actor network generates the probability distribution of spectral output actions according to the input state based on the input state ; Select the action with the highest probability as the execution result, and adjust the PWM duty cycle of the spectral output module in real time.

[0101] In addition, in terms of edge-cloud collaborative deployment: Model inference is deployed on edge nodes to achieve real-time response; The cloud regularly updates the model parameters and distributes them to edge devices via OTA (Over-the-Air) to ensure the long-term adaptability of the model.

[0102] Assume that the system is applied to a medical phototherapy scenario. The following are the actual application steps: The environmental spectrum detects that the intensity of blue light (450nm) is insufficient, the temperature and humidity are maintained at 25°C and 40%RH, and the user behavior is classified as "sleep"; According to the dynamic weight allocation algorithm in Embodiment 1, the target spectral output is generated, and it is necessary to enhance the output of the blue light band; The reinforcement learning model outputs the regulation strategy in real time, increases the PWM duty cycle of the blue light channel to 0.85, and reduces the output of other bands at the same time; After the spectral regulation is completed, the error between the actual spectral output and the target value is less than 5%, meeting the accuracy requirements of medical phototherapy.

[0103] Embodiment 3:

[0104] The distributed collaborative optimization algorithm is used to efficiently allocate computing resources between edge computing nodes and cloud servers, reducing the overall power consumption of the system and improving real-time performance. This system consists of the following hardware: Edge computing node: Equipped with an embedded computing module (such as NVIDIA Jetson Xavier NX), supporting real-time data processing and spectral regulation, providing real-time computing support for the spectral output module and multimodal data acquisition module; Cloud server: Equipped with a high-performance GPU computing unit (such as NVIDIA A100) and a distributed database for long-term data storage, model training, and policy optimization; Communication module: The edge node and the cloud server are connected through a 5G communication module or Wi-Fi 6, supporting high-bandwidth and low-latency data transmission.

[0105] Dynamic decision-making rules for task allocation: The system classifies real-time tasks (such as multi-modal data fusion, spectral regulation reasoning) and periodic tasks (such as model training, long-term policy optimization). For real-time tasks: They are preferentially executed on edge nodes to ensure low-latency response. For periodic tasks: They are migrated to the cloud server for execution to reduce the computing load on edge nodes; the length of the task queue on the edge node For dynamically making decisions on task migration: where is the threshold of the task queue length, which is dynamically adjusted according to the operation historical data.

[0106] Long-term resource allocation optimization strategy: The system records the operation data of edge nodes and cloud servers (such as task completion time, resource occupancy rate, power consumption, etc.) every 10 minutes, and optimizes the resource allocation strategy through the following steps:

[0107] Data collection: Record the task running time and computing load and power consumption .

[0108] Data modeling: Based on the random forest regression model, predict the task load distribution in the next 1 hour and generate optimization suggestions.

[0109] Strategy update: Use the genetic algorithm to optimize the edge-cloud task allocation ratio, and the objective function is the weighted minimization of the system's task completion time and power consumption: where and are the weight factors of time and power consumption, which are dynamically set based on the application scenario.

[0110] In terms of the implementation of real-time task scheduling, the scheduling trigger mechanism: The system sets the trigger conditions for task scheduling. For example, when the CPU / GPU utilization rate of the edge node exceeds 80% or the task queue length exceeds 5 items, task migration is triggered; the task migration process: Task classification: Check the task priority, and preferentially retain tasks with high real-time requirements (such as spectral regulation calculation) on the edge node; Migration mechanism: Send low-priority tasks (such as data storage and mid-term analysis) to the cloud server for processing through the communication module; Feedback mechanism: After the cloud processing is completed, the results are transmitted to the edge node in real time through the 5G communication module for subsequent execution of control instructions; Task scheduling algorithm: The scheduling algorithm adopts a task allocation logic based on the greedy strategy: Sort according to the priority from high to low, and allocate computing resources item by item.

[0111] In terms of system energy consumption optimization, such as power consumption control of edge nodes: when the edge node is in an idle task state, it automatically switches to the low-power mode (such as reducing the processor main frequency to 50%), and uses dynamic voltage and frequency scaling (DVFS) technology to dynamically adjust the voltage and frequency of the processor according to the task load; improvement of resource utilization rate of cloud servers: the cloud server reduces the task allocation frequency and reduces the energy consumption of ineffective calculations through batch task processing technology (such as data batch merging processing).

[0112] Suppose the system is applied to the agricultural supplementary lighting scenario. The following is the actual application process of the system's distributed collaborative optimization. For example, scenario input: the edge node collects multi-modal data (spectral intensity, temperature and humidity, CO 2 concentration, etc.) per second and completes real-time spectral regulation inference; the data storage task and the model retraining task are migrated to the cloud server.

[0113] Real-time task allocation: when the CPU utilization rate of the edge node reaches 85%, the task scheduling module triggers the migration of low-priority tasks; the data storage task is transmitted to the cloud server for processing through the 5G communication module.

[0114] And long-term optimization can be carried out. For example, the system analyzes the operation data (such as task completion time and power consumption) every 10 minutes and predicts the task load distribution in the next 1 hour; the optimized strategy shows that in the current scenario, the cloud should undertake 70% of the computing load and the edge node undertakes 30%.

[0115] After distributed collaborative optimization, the task delay is reduced from 100 ms to 60 ms, and the overall power consumption of the system is reduced by 15%, meeting the dual requirements of real-time spectral regulation and energy consumption for agricultural supplementary lighting.

[0116] Example 4:

[0117] In modern agricultural greenhouse cultivation, the growth of plants highly depends on the light intensity and spectral quality. Traditional supplementary lighting systems mostly use a fixed ratio of red and blue light for supplementary lighting, ignoring the different spectral requirements of different crops, different growth stages and environmental conditions, resulting in unsatisfactory supplementary lighting effects, high energy consumption and serious resource waste.

[0118] For example, in a smart agricultural greenhouse cultivation base, multiple crops are planted, including tomatoes, cucumbers and strawberries. In order to improve the yield and quality of the crops, users need to supplement specific spectral bands during the day, suppress the excessive respiration of plants at night, and at the same time reduce the energy consumption of supplementary lighting. The system needs to adapt the spectral output in real time and dynamically adjust the spectral composition according to the environmental data (light intensity, temperature and humidity, CO 2 concentration) and the current growth stage of the crops to achieve the dual goals of precise supplementary lighting and energy conservation.

[0119] The intelligent spectral regulation system can be deployed inside the greenhouse, including the following modules. For example, the multimodal data acquisition module: Environmental sensors installed inside the greenhouse collect data on light intensity (350 - 750 nm), temperature (range: 15 - 35 °C), humidity (range: 30 - 80%), and carbon dioxide concentration (range: 400 - 2000 ppm) in real time; the camera module performs real-time image acquisition of the color and morphology of crop leaves for analyzing the health status of the crops; the spectral regulation module: Equipped with a 12-channel tunable LED light source, including multiple bands such as red light (660 nm), blue light (450 nm), green light (520 nm), far-red light (730 nm), etc., supporting fine adjustment of the PWM duty cycle; the computing and control module: The system uses an edge computing node (such as NVIDIA Jetson Nano) to achieve data fusion and spectral regulation inference, and the cloud server (supporting GPU acceleration) is used for long-term strategy optimization and model update.

[0120] The multimodal data acquisition module collects environmental data at a frequency of 1 second, and after normalization, inputs it into the dynamic weight allocation algorithm based on LSTM. The algorithm calculates the real-time priorities of different spectral bands according to the dynamic changes in environmental light intensity, carbon dioxide concentration, and temperature and humidity. Combining the spectral requirements of the crop growth stage (for example, the demand for blue light is higher than that for red light in the tomato seedling stage) and the supplementary lighting strategy set by the user, it generates the target spectral output value \(S_{\text{target}}\). The spectral regulation module dynamically adjusts the LED duty cycle of each band according to the calculation results. For example, when the environmental light intensity is insufficient and the carbon dioxide concentration is low, the system automatically increases the output of the red and blue light bands and reduces the energy consumption of the green light band.

[0121] At the same time, to improve real-time performance and reduce energy consumption, the system adopts a distributed optimization mechanism of edge-cloud collaboration. For example, the edge node: realizes real-time data fusion and spectral regulation inference, and preferentially processes real-time tasks sensitive to latency; the cloud server: analyzes the environmental data and crop growth status of the past 24 hours every day to optimize the future spectral regulation strategy; uses a random forest regression model to predict the future load and generate the task allocation ratio to ensure the load balance of the edge node; task migration and feedback: when the task queue of the edge node exceeds the load threshold, the system migrates low-priority tasks (such as historical data analysis) to the cloud for execution; after the task is completed, the cloud feedbacks the optimization result and updates the inference model of the edge device.

[0122] Taking tomato cultivation as an example, the light supplementation performance of the system in the seedling stage (the red light / blue light demand ratio is about 1:2) and the flowering stage (the red light / blue light demand ratio is about 2:1) is as follows. Spectral accuracy: Through multi-modal data fusion and dynamic regulation, the system ensures that the output error of the target spectrum is less than 3%, significantly improving the accuracy of light supplementation. Energy consumption control: Through distributed collaborative optimization, edge nodes process latency-sensitive tasks in real time, and the cloud completes low-priority computing tasks, reducing the overall power consumption by 15%-20%. Crop yield and quality improvement: After the system runs for 3 months, the yield per tomato plant increases by 15%, the fruit color becomes more vivid, and the quality is significantly improved.

[0123] Through the combination of the dynamic weight allocation algorithm and the reinforcement learning model, this system achieves high-precision and high-responsiveness spectral regulation. The distributed optimization mechanism balances real-time performance and resource efficiency, overcoming the energy consumption and latency problems of the single-node computing mode. The practical application in the agricultural light supplementation scenario proves that the system has significant technical advantages in energy conservation, yield increase, and quality improvement.

[0124] Example 5:

[0125] In the treatment of skin diseases, for example, phototherapy is widely used to treat diseases such as psoriasis, eczema, and melasma. Its curative effect depends on the accuracy of spectral output and the stability of intensity. However, traditional phototherapy devices mostly adopt fixed spectral output or simple manual adjustment modes, making it difficult to dynamically adjust the spectral output according to the patient's skin type, lesion location, and treatment response, which easily leads to over-irradiation or insufficient curative effect, and even increases the risk of side effects.

[0126] The system implementation method is as follows. In terms of hardware and data configuration, a multi-modal data acquisition module is adopted: An ambient light sensor installed on the phototherapy device collects indoor light intensity (range: 100 - 1000 lux), humidity (range: 30 - 60%), and the temperature on the patient's skin surface (range: 32 - 37°C) in real time. The camera module obtains real-time image data of the patient's skin lesions (resolution 1080p) for evaluating the size, color, and treatment response of the skin lesions. Spectral regulation module: The phototherapy device is equipped with a narrowband UVB light source (311nm), a red light source (630nm), and a blue light source (450nm), with a light intensity output range of 0 - 500mW / cm², supporting real-time dynamic adjustment. Computing and control module: The system uses an embedded processor (such as NXP i.MX8) to process data fusion and real-time spectral regulation, and connects to a cloud server (supporting AI model updates) to optimize long-term treatment plans.

[0127] In the dynamic regulation of spectral output, patient data fusion can be carried out: The system collects real-time data of the patient's skin at a frequency of 5 seconds, including temperature, lesion area change (unit: cm²), and skin color RGB values. Combining ambient light intensity and humidity information, the output priority of each spectral band is calculated through a dynamic weight allocation algorithm based on LSTM; Spectral regulation strategy: Through the dynamic calculation of the target output value and combining the photosensitivity of the patient's lesion and the current treatment response, a spectral regulation strategy is generated. For example, when the lesion area decreases to 50% of the initial value and the skin temperature rises by more than 0.5°C, the system automatically reduces the output intensity of UVB light and increases the proportion of red light to promote healing; Real-time feedback and adjustment: The system collects lesion images through a camera to analyze the treatment effect. When it is found that there is an increase in pigmentation or the inflammatory reaction subsides, the spectral output plan is adjusted in real time.

[0128] To verify the superiority of the technical solution of the present invention, the following comparative experiment was designed:

[0129] Experimental subjects: Randomly select 100 psoriasis patients and divide them into two groups, with 50 people in each group. The experimental group is treated with the intelligent spectral regulation system of the present invention, and the control group uses a traditional phototherapy device (fixed spectral output, 80% UVB ratio, 20% red light ratio).

[0130] Treatment parameters: The experimental group dynamically adjusts the spectral ratio, the UVB range is 50%-80%, and the red light range is 20%-50%. The specific parameters are automatically adjusted according to the patient's real-time data. The control group has a fixed spectral ratio, and the output of UVB and red light does not change with the treatment response; Treatment cycle: Both groups are treated for 8 weeks, 3 times a week, and each treatment lasts for 10 minutes.

[0131] In the experimental group, the average area of the lesions decreased by 72%, while in the control group, it was 58%. In the experimental group, obvious improvement of the lesions occurred at the 4th week, while the improvement trend in the control group appeared at the 6th week. The incidence of skin photosensitivity reactions (such as erythema, burning sensation) in the experimental group was 8%, significantly lower than 20% in the control group. This indicates that the dynamic regulation strategy of the present invention effectively reduces the side effects caused by over-irradiation. The recovery time of the experimental group was shortened by an average of 15% after each treatment, and the operation interruptions caused by equipment parameter adjustment during the treatment were significantly reduced; and the experimental group achieved precise regulation of real-time spectrum output through multi-modal data fusion and dynamic weight allocation algorithm. The spectrum error was always kept within 3%, significantly improving the phototherapy effect compared with the error of the fixed output in the control group (about 15%); and the distributed optimization mechanism realized the collaborative optimization of edge computing and the cloud, avoiding unnecessary excessive spectrum output, and reducing the overall energy consumption of the equipment by about 18%. And in terms of optimizing the patient experience, the dynamic regulation strategy combined with the real-time feedback of the patient not only improved the treatment effect, but also effectively controlled the side effects, which is conducive to providing a safer and more comfortable treatment experience for the patient.

[0132] Example 6:

[0133] In this embodiment, a dynamic adjustment mechanism is added to the data fusion processing module, which specifically includes the following steps, that is, the improvement of data preprocessing: the multi-modal data collected is subjected to multi-stage normalization and standardization processing, and the specific method is as follows:

[0134] Normalization of spectral data: The original spectral data collected is normalized according to the following formula:

[0135] where, is the maximum light intensity value in the current acquisition period. The function of this formula is to eliminate the scale difference of spectral data under different intensity conditions and improve the stability of subsequent processing.

[0136] Standardization of temperature, humidity and carbon dioxide concentration data: Z-score standardization is adopted, and the specific formula is:

[0137]

[0138] where, is the real-time data collected, and are the historical mean and standard deviation of this type of data respectively. This method can dynamically adapt to the data distribution characteristics under different environmental conditions.

[0139] In addition, for the calculation of the weight coefficient, the parameter definition and implementation path in the formula are improved. For example, the dynamic adjustment factor is generated by the following steps:

[0140] Historical data analysis: Based on the LSTM (Long Short-Term Memory) model, perform time series fitting on the historical change trends of each type of data to generate preliminary predicted values of the weight factors; Real-time adjustment mechanism: When real-time data is updated, use the gradient descent method to perform dynamic optimization to ensure that it can quickly respond to data changes. The optimization formula is as follows:

[0141]

[0142] Among them, is the learning rate, is the loss function for weight optimization, defined as the weighted sum of squares of the spectral output error.

[0143] Weight coefficient The final calculation formula of is as follows:

[0144]

[0145] Among them, is the confidence parameter of the data, and its calculation formula is:

[0146]

[0147] Parameter is determined by experimental fitting and represents the sensitivity adjustment factor of the confidence level; is the historical mean. The design of this formula can balance the importance of real-time data and the weight influence of historical data; and, in the optimization of the spectral regulation algorithm module, such as the improvement of the objective function, in the deep reinforcement learning model, the optimization objective function for the spectral regulation strategy is adjusted, and an energy consumption balance factor is added to make the balance between the accuracy of the spectral output and the energy consumption more optimized. The optimized objective function is:

[0148]

[0149] Among them, is the energy consumption weight adjustment factor, is the system power consumption at time , which is measured by a real-time current sensor. By adjusting , the requirements for energy consumption and accuracy in different scenarios can be adapted.

[0150] And online training is combined with offline training. For example, the reinforcement learning model adopts a combination of online inference and periodic offline training to improve the adaptability of the model. Offline training is carried out through a pre-sampled multi-dimensional data set, and the data volume is not less than 20,000 groups, covering agricultural supplementary lighting, medical phototherapy, and intelligent lighting scenarios. Online inference is based on real-time input states and dynamically outputs spectral regulation actions.

[0151] The experimental verification process is as follows. Data source: Collect multi-modal data in an agricultural greenhouse, including light intensity (350 - 750 nm), temperature and humidity (range: 10 - 50 °C, 30% - 90% RH), and carbon dioxide concentration (range: 400 - 2000 ppm). Verification index: Evaluate the technical effect of this embodiment by comparing the accuracy of the spectral output and the system energy consumption before and after dynamic regulation. Experimental result: The spectral output error is reduced from 5% to 2.5%, and the accuracy is improved by 50%. The overall power consumption of the system is reduced by 18%, and the real-time response time is shortened by 30%.

[0152] In terms of hardware configuration, such as the data acquisition module: high-precision spectral sensor, digital temperature and humidity sensor, and infrared gas sensor; computing unit: embedded AI chip (NVIDIA Jetson Xavier NX); spectral output module: 12-channel adjustable spectral LED light source, supporting PWM modulation; and in terms of software, the reinforcement learning model can be implemented using PyTorch, deployed through edge computing nodes, and the model parameters are regularly optimized in combination with the cloud.

[0153] Example 7:

[0154] Based on the original technology, this embodiment designs a dynamic feedback regulation mechanism based on real-time fusion of multi-dimensional data to improve the accuracy and response efficiency of the spectral regulation system. This solution is applicable to intelligent agricultural light supplementation, medical phototherapy, and intelligent lighting scenarios.

[0155] Step 1: Multi-modal data acquisition and standardization processing. For example, in terms of the composition of the data acquisition module, it includes:

[0156] Environmental spectral data: The spectral intensity curve is collected through a high-precision spectral sensor with a wavelength range of 400 - 700 nm and a resolution of 1 nm at a period of 1 second; temperature and humidity data: A digital temperature and humidity sensor with accuracies of ±0.1 °C and ±1% RH respectively is used, and the collection period is 1 second; CO 2 Concentration data: An infrared gas sensor is used, with a monitoring range of 400 - 2000 ppm and a resolution of ±5 ppm; user behavior data: Real-time video is collected using a camera and classified (working, resting, sleeping) in combination with the ResNet-50 convolutional neural network, and the classification confidence is directly represented by the network output probability.

[0157] In terms of data standardization processing, the spectral data normalization formula: Where, is the original collected light intensity data, is the maximum light intensity value within the collection period; for temperature, humidity, and CO 2 Data standardization: The z-score standardization algorithm is adopted: Among them, is the collected real-time value, is the historical average value, is the historical standard deviation.

[0158] Step 2: Optimization of dynamic weight allocation. To achieve dynamic priority adjustment of multimodal data, a dynamic feedback mechanism is designed to clarify the logic and formula source of weight allocation. For example:

[0159] Generation of dynamic adjustment factor: The dynamic adjustment factor is generated based on the LSTM model. Its core process is as follows: Input data: The collected real-time data sequence ; Time series prediction: Fit the historical data through LSTM to predict the weight adjustment trend; Optimization formula: Perform real-time optimization on and update it using the gradient descent method: Among them, is the learning rate, is the optimization objective function, defined as the weighted sum of squares of the spectral output error.

[0160] Weight coefficient calculation formula:

[0161]

[0162] Among them: : The real-time weight of the th class of data; : The normalized real-time data value; : The dynamic adjustment factor; : The data confidence parameter, determined by the following formula: The parameter is the sensitivity adjustment factor, is the historical average value.

[0163] Step 3: Optimization of spectral regulation strategy. To improve the accuracy of spectral regulation and energy consumption balance, the reward function and target formula of the deep reinforcement learning model can be redesigned. Such as the optimization of the target function:

[0164]

[0165] Among them: : The spectral output accuracy score; : The real-time energy consumption, measured by the current sensor; : The energy consumption weight adjustment factor, used to balance spectral accuracy and power consumption.

[0166] And for the dynamic adjustment of the reward function: According to the requirements of the actual application scenario (agriculture or medical), by adjusting values to achieve the spectral regulation goal. For example, in the agricultural scenario: improve , and prioritize reducing energy consumption; in the medical scenario: reduce , and prioritize ensuring spectral accuracy;

[0167] Step 4: Dynamically allocate computing resources. To solve the resource allocation problem of the system under high-load conditions and further reduce power consumption, the system designs a distributed computing resource dynamic allocation mechanism based on real-time triggering and long-term optimization. This mechanism includes four core modules: task migration trigger conditions, synchronization mechanism, load prediction, and long-term optimization.

[0168] In the trigger conditions for task migration, such as the metrics monitored by edge nodes in real time, CPU utilization rate (%) is used to monitor the real-time computing pressure of edge computing nodes; GPU utilization rate (%) reflects the resource usage of computing-intensive tasks such as spectral regulation inference; task queue length (number of tasks) indicates the backlog of tasks to be processed by edge nodes currently.

[0169] Its trigger rule is: the system sets dynamic thresholds. When the computing resources of edge nodes are close to saturation, the migration of low-priority tasks is triggered:

[0170]

[0171] Among them: : the task queue length of the edge node; : the dynamically adjusted task queue threshold, which is updated in real time according to historical load data and current computing pressure. Classification of high-priority and low-priority tasks: High-priority tasks, such as spectral regulation inference and multimodal data fusion, need to be completed on edge nodes first to ensure real-time performance; low-priority tasks, such as historical data storage and long-term model training, can be processed later and migrated to the cloud server.

[0172] In terms of the synchronization mechanism and real-time performance guarantee, to ensure the real-time performance and synchronization of the system during task migration, the following synchronization mechanism is designed, such as asynchronous queue management: The system adopts a distributed asynchronous queue to manage the task transmission between edge nodes and cloud servers respectively. By following the "first in, first out" rule, communication blocking between modules is avoided to ensure the correct order of data processing.

[0173] In terms of data consistency maintenance, during the migration process, for example, task data is encrypted and verified by timestamp to prevent data duplication or loss during transmission. After receiving the results processed by the cloud, the edge node re-verifies the task ID to ensure that the results correspond one-to-one with the tasks. At the same time, its real-time requirement is as follows: task migration latency: the system requires the task migration latency to be controlled within 100 ms to ensure the real-time requirements of high-priority tasks; dynamic adjustment of network bandwidth: when the network latency exceeds the set threshold (such as 200 ms), the system automatically reduces the migration frequency and preferentially retains local computing. And in terms of the load prediction mechanism, to optimize the long-term resource allocation efficiency, the system predicts future task loads based on a random forest model and dynamically adjusts the task allocation strategy. The input features of the random forest model, such as the model input features cover the following three types of data:

[0174] Real-time task load metrics: CPU / GPU utilization of the edge node; current task queue length; average value of current task completion time (unit: ms); Historical operation data metrics: task migration frequency of the edge node; average length of the cloud task queue; system energy consumption (unit: watt); External environment metrics: current network bandwidth (unit: Mbps); network latency (unit: ms); Its weight allocation rule is: the system sets different weights according to the importance of the features, that is, the weight of real-time task load metrics: 50%, which preferentially reflects the current load; the weight of historical operation data metrics: 30%, which is used for long-term trend analysis; the weight of external environment metrics: 20%, which evaluates the network conditions for migration.

[0175] And in terms of model output: predict the task load distribution within the next 1 hour; dynamically adjust the task division ratio between the edge node and the cloud.

[0176] And during long-term operation, the system optimizes the task allocation strategy through a genetic algorithm, with the goal of minimizing the weighted sum of task completion time and power consumption:

[0177]

[0178] Among them: : Completion time of the th task; : Power consumption of the th task; 、 : Weight factors for time and power consumption, dynamically set.

[0179] The system can also achieve dynamic allocation of computing resources, data collection and analysis through the following closed-loop logic: real-time collection of the operation data of edge nodes and cloud servers; task triggering and allocation: task migration or local processing according to the triggering rules; load prediction and optimization: combining the random forest prediction model and the genetic algorithm to optimize the long-term task allocation strategy; feedback and adjustment: the system regularly adjusts the triggering conditions and allocation ratios according to the operation data to ensure long-term stability.

[0180] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent LED spectrum control system based on multi-dimensional data fusion, characterized in that: The system includes: The multi-dimensional data acquisition module is used to collect environmental spectrum data, temperature and humidity data, carbon dioxide concentration data and user behavior data in real time, among which: environmental spectrum data is collected by high-precision spectrum sensors; temperature and humidity data are collected by digital sensors; carbon dioxide concentration data is measured by infrared gas sensors; user behavior data is analyzed by cameras combined with image recognition algorithms; The data fusion processing module is connected to the multidimensional data acquisition module and is used to perform standardized preprocessing and dynamic fusion on the collected multidimensional data, including: using a fuzzy logic-based standardization algorithm to perform dimensionless processing on the data of different physical quantities; using a multimodal adaptive weight allocation algorithm, according to the time series variation characteristics of environmental spectrum data, temperature and humidity data, carbon dioxide concentration data and user behavior data, the weight coefficient of the spectral output parameter is calculated, and the weight coefficient is calculated by the following formula: in: For spectrum control The weight of the class data; For the collection of Real-time value of class data; It is a dynamic adjustment factor obtained based on historical data fitting, which is obtained by combining the time series prediction model LSTM with historical data training; is the dimension of the collected data; Indicates Spectral output coefficient at each time point; For the The value of the class data; The spectrum control algorithm module is connected to the data fusion processing module and is used to control the LED spectrum output in real time according to the data fusion result, including: using a spectrum optimization model based on deep reinforcement learning, the model input is the weight coefficient and the spectrum output target value after data fusion; in the model training process, the following objective function is used for dynamic optimization: in: is the optimization goal of the policy network; Based on the current network parameters Spectral control strategy; For time The spectrum output effect score at the moment; is the reward discount factor, which controls the weight of the impact of future rewards on the current strategy; An LED spectrum output module is connected to the spectrum control algorithm module and is used to dynamically adjust the LED spectrum output according to the optimization result, including: realizing spectrum output through a multi-channel adjustable spectrum LED light source; configuring a fast-response current driver to dynamically adjust the current value of each band spectrum channel according to the target parameter output by the spectrum control algorithm module; The system collaborative control module is connected to the above modules and is used to coordinate the operation of the multi-dimensional data acquisition module, the data fusion processing module, the spectrum control algorithm module and the LED spectrum output module, including: configuring an event-driven mechanism for dynamically triggering data acquisition and processing; and dynamically allocating computing resources through a distributed collaborative optimization algorithm.

2. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The fuzzy logic normalization algorithm in the data fusion processing module includes the following steps: normalizing the data of different physical quantities to a range between 0 and 1 respectively; based on the dynamic range of data changes, using triangular membership functions and dynamically adjusted threshold intervals to construct fuzzy sets to reduce the scale differences between physical quantities.

3. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: In the multi-modal adaptive weight allocation algorithm, the dynamic adjustment factor The calculation includes: collecting historical data in the target application scenario; performing time series fitting analysis based on the LSTM model to predict the long-term weight trend of each type of data on the spectral output; and using the gradient descent method to dynamically update the weight adjustment factor to optimize the real-time and robustness of the weight calculation.

4. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The training process of deep reinforcement learning in the spectral control algorithm module includes: generating an initial sample library of spectral control strategies based on multidimensional data collected offline; using a convolutional neural network to extract features of multidimensional data; constructing a reinforcement learning environment and defining a reward function as a weighted comprehensive score of the accuracy score of the spectral output and energy consumption; and using distributed parallel computing to accelerate the strategy optimization of deep reinforcement learning.

5. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The multi-channel adjustable spectrum LED light source in the LED spectrum output module includes: LED light sources of four spectral bands: red, green, blue and white; the light source of each band is independently powered by a constant current driver; and the PWM pulse width modulation technology is used to adjust the luminous intensity and duration of the light source of each band in real time.

6. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The event-driven mechanism in the system collaborative control module includes: detecting the real-time sampling frequency of the multi-dimensional data acquisition module; triggering the data fusion processing module to recalculate the weight coefficient when the spectral output deviation exceeds the preset threshold; and managing the data transmission and control instructions between multiple modules through asynchronous queues.

7. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The calculation formula of the weight coefficient further combines the confidence parameter of the data , the confidence parameter is determined by the following formula: in: is the sensitivity adjustment factor, determined by experimental fitting; For real-time collection The value of the class data; is the historical mean of this data category.

8. The intelligent LED spectrum control system based on multi-dimensional data fusion according to claim 1, characterized in that: The reward discount factor The value of is adjusted dynamically according to the following formula: in: is the adjustment factor, which depends on the sensitivity requirement of spectral regulation; For time Spectral output score at the moment; Score threshold for the target.

Images