LED Outdoor Light Pollution Suppression System and Method Combining Generative Adversarial Networks

Abstract

The present application provides an LED outdoor light pollution suppression system and method incorporating a generative adversarial network, which relates to the field of intelligent lighting. In the outdoor LED lighting scenario, the present application organically combines a cloud teacher model and a local student model through a cold start mechanism and real-time online guidance to achieve adaptive control of multi-objective constraints such as light uniformity, energy consumption, and glare index in a variable environment. The present application also performs cold start in different scenarios through a model snapshot library, and solidifies new snapshots when the performance is excellent, forming an accumulative evolution mechanism. In this way, when the system encounters a similar scenario again, it can immediately load existing excellent snapshots, which not only improves the algorithm efficiency, but also enhances the robustness of the solution and the long-term adaptive ability.

Description

Technical Field

[0001] This application relates to the field of intelligent lighting. Specifically, it relates to an LED outdoor light pollution suppression system and method combined with a generative adversarial network. Background Art

[0002] With the continuous improvement of urban lighting levels, outdoor LED lighting has been widely used in areas such as transportation, public places, and landscapes. However, outdoor light control often adopts fixed or relatively coarse-grained dimming strategies, making it difficult to respond promptly to weather changes, fluctuations in the number of people, and regional characteristic differences, resulting in unnecessary energy consumption and light pollution. Some existing technologies attempt to combine machine learning to intelligently adjust lighting schemes, but when facing a diverse and dynamically changing outdoor environment, there are still problems such as insufficient cold-start adaptability, inability to balance multiple objective constraints, and lack of an online learning mechanism. Especially when the environmental conditions are significantly different from the data distribution on which the model was pre-trained, the model often experiences dimming failure or requires a long re-training process, making it difficult to meet the requirements of real-time and stability for LED outdoor lighting. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, this application provides an LED outdoor light pollution suppression system and method combined with a generative adversarial network.

[0004] In a first aspect, this application provides an LED outdoor light pollution suppression method combined with a generative adversarial network, including:

[0005] Obtain the environmental data of the target area, perform matching analysis on the environmental data and the scene feature data in the scene library, and determine whether the environmental data belongs to a new scene condition based on similarity calculation;

[0006] In response to the environmental data belonging to a new scene condition, for the environmental data, select the initial parameter set with the highest similarity to the scene feature data from the model snapshot library; wherein, the initial parameter set is used as the initial state of the local student model; the student model is a lightweight generative adversarial network;

[0007] Use the student model to infer the environmental data and generate a dimming control instruction; obtain the environmental feedback data after the lighting device executes the dimming control instruction;

[0008] Use the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction; the student model adjusts its own parameters based on the guidance information.

[0009] As an alternative implementation, after the student model inference and teacher model guidance are completed for a preset number of times, the teacher model evaluates the performance of the current parameters of the student model, and when the preset conditions are met, solidifies the current parameters of the student model into a new model snapshot and stores it in the model snapshot library.

[0010] As an alternative implementation, the performance evaluation includes: the teacher model performs a multi-objective performance evaluation on the current parameters of the student model based on the environmental feedback data to generate a multi-objective evaluation result;

[0011] Among them, the multi-objective performance evaluation includes: illumination uniformity evaluation, device energy consumption evaluation, and glare index evaluation.

[0012] As an alternative implementation, the matching analysis of the environmental data with the scene feature data in the scene library includes:

[0013] Extracting an illuminance feature vector, a pedestrian flow feature vector, and a weather feature vector from the environmental data;

[0014] Calculating the similarity between each of the feature vectors and the corresponding feature vectors of each scene in the scene library respectively;

[0015] Selecting the scene with the highest similarity and exceeding the preset threshold as the target scene;

[0016] In response to all similarities being lower than the preset threshold, determining that the environmental data is a new scene condition.

[0017] As an alternative implementation, the selection of the initial parameter set with the highest similarity to the scene feature data from the model snapshot library includes:

[0018] Performing scene annotation on each model snapshot in the model snapshot library, where each model snapshot is associated with at least one scene feature vector or scene label;

[0019] Calculating the similarity between the feature vector of the current environmental data and the scene feature vectors of each model snapshot respectively;

[0020] Sorting by similarity and selecting the target model snapshot with the highest similarity and greater than the preset threshold;

[0021] Reading the model parameters of the target model snapshot and using them as the initial parameter set of the student model;

[0022] In response to the similarity of no model snapshot being greater than the preset threshold, triggering random initialization.

[0023] As an alternative implementation, the inference of the environmental data using the student model to generate a dimming control instruction includes:

[0024] Input the preprocessed environmental data into the generator network of the student model, perform forward inference, and obtain the lighting control parameters under the current scenario;

[0025] Based on the lighting control parameters and in combination with the preset adjustment range of the lighting equipment, map the control parameters to a dimming control instruction.

[0026] As an alternative implementation, the student model further includes a discriminator network; the discriminator network is used for:

[0027] Receive the dimming control parameters output by the generator network and the environmental feedback data;

[0028] Compare the environmental feedback data with the expected target feedback to determine the adversarial loss;

[0029] Based on the adversarial loss, perform a local gradient update on the generator network.

[0030] As an alternative implementation, using the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction includes:

[0031] Compare and analyze the environmental feedback data with the preset performance indicators to determine the differences between the lighting uniformity, equipment energy consumption, and glare index and the target requirements;

[0032] Run the discriminator network in the teacher model to evaluate the dimming strategy output by the student model and generate a simulation evaluation result;

[0033] Based on the multi-objective evaluation result and the simulation evaluation result, determine the difference information between the student model and the expected dimming effect;

[0034] Based on the difference information, generate guidance information for updating the student model.

[0035] As an alternative implementation, the guidance information includes: parameter gradient, target distribution, and soft label.

[0036] In a second aspect, the present application provides an LED outdoor light pollution suppression system combined with a generative adversarial network, including:

[0037] A matching unit, configured to obtain environmental data of a target area, perform matching analysis on the environmental data and the scene feature data in the scene library, and determine whether the environmental data belongs to a new scene condition based on similarity calculation;

[0038] a cold start unit, configured to select, in response to the environmental data belonging to a new scene condition, an initial parameter set having the highest similarity to the scene feature data from a model snapshot library for the environmental data; wherein the initial parameter set serves as an initial state of a local student model; and the student model is a lightweight generative adversarial network;

[0039] A control unit is configured to use the student model to infer the environmental data and generate a dimming control instruction; and obtain environmental feedback data after the lighting device executes the dimming control instruction;

[0040] The correction unit is used to use the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instructions; the student model adjusts its own parameters based on the guidance information.

[0041] Compared with existing technologies, this application organically combines a cloud-based teacher model with a local student model in outdoor LED lighting scenarios through a "cold start mechanism" and "real-time online guidance," achieving adaptive control of multiple objective constraints such as illumination uniformity, energy consumption, and glare index in a changing environment. This technology offers the following outstanding technical advantages:

[0042] 1. Compared with traditional LED fixed dimming or rule-based solutions

[0043] This application uses a student model to quickly infer and generate dimming control instructions locally, and combines the multi-target evaluation and deep discrimination of the cloud-based teacher model to achieve instant adaptation to environmental changes. Traditional fixed time periods or simple light sensitivity adjustments often cannot take into account scenarios such as sudden weather events and drastic fluctuations in pedestrian flow, and there are problems such as dimming lag, light pollution, or energy waste. This application combines multi-dimensional indicators such as lighting uniformity, glare index, and energy consumption, and uses a large cloud-based model to comprehensively evaluate and generate guidance information, so that the dimming strategy achieves a better balance between safe lighting and energy conservation and emission reduction; conventional LED control mostly only considers a single goal (such as saving power or ensuring brightness), which is difficult to meet the complex lighting needs of cities.

[0044] 2. Compared with using only traditional machine learning or rule-based algorithms

[0045] This application utilizes the "model snapshot library" to select the most similar initial parameter set from existing model snapshots after identifying new scenario conditions, effectively avoiding dimming errors caused by training from scratch; when there are no selectable model snapshots, it can be initialized randomly or using a conservative strategy to ensure the most basic lighting function. If traditional machine learning solutions lack this cold start mechanism, they often exhibit significant dimming deviations in unfamiliar scenarios, increasing both safety and energy consumption risks. This application can also perform real-time distillation or fine-tuning on the output of the student model by the teacher model during the inference stage, ensuring that the system can adaptively and continuously optimize the dimming strategy in the face of situations such as peak and valley of the flow of people and sudden weather changes; while most general machine learning models require offline retraining and are slow to respond to environmental mutations once launched.

[0046] 3. Compared with ordinary generative adversarial networks (GANs)

[0047] This application deploys a lightweight GAN (student model) on the terminal side with limited computing power for quickly generating dimming control parameters; while discrimination or multi-objective comprehensive evaluation is performed by the cloud teacher model. This design not only retains the powerful modeling ability of GAN for highly nonlinear problems but also avoids the bottleneck of large-scale training of the entire GAN when local computing power is insufficient. Ordinary GANs often require a large amount of computing power and data and are difficult to directly use in outdoor real-time control scenarios. This application combines different objectives such as lighting uniformity, energy consumption, and glare index, and endows the student model with more accurate environmental adaptability through the "scenario library" and "model snapshot library" in the initial stage, avoiding problems such as mode collapse or overfitting due to training data deviation when GAN is actually deployed.

[0048] 4. Different from the offline distillation of general teacher-student models

[0049] Traditional knowledge distillation mostly performs soft label or feature alignment on the student model by the teacher model during offline training. Once the student model encounters a new environmental distribution after deployment, it requires offline fine-tuning or retraining, lacking flexibility. This application also continuously retains teacher-student interaction during the inference stage. Through real-time evaluation and adversarial discrimination of environmental feedback data, it provides "guidance information" online for the student model, enabling it to be quickly updated and adaptive, greatly enhancing its practicality in the dynamic outdoor lighting environment. In addition to offline distillation, this application also performs cold start in different scenarios through the "model snapshot library" and solidifies new model snapshots when performing well, forming an "accumulative" evolution mechanism. In this way, when the system encounters a similar scenario again, it can immediately load existing excellent model snapshots, which not only improves the algorithm efficiency but also enhances the robustness of the solution and the long-term adaptive ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of the LED outdoor light pollution suppression method combining a generative adversarial network provided by an embodiment of this application;

[0051] Figure 2 This is a flowchart of a method for matching and analyzing environmental data with scene feature data in a scene library provided by an embodiment of the present application;

[0052] Figure 3 This is a schematic diagram of an LED outdoor light pollution suppression system combined with a generative adversarial network provided by an embodiment of the present application. Detailed implementation manners

[0053] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments.

[0054] See Figure 1 As shown, this is a flowchart of an LED outdoor light pollution suppression method combined with a generative adversarial network provided by an embodiment of the present application. The method includes steps S101 to S104, where:

[0055] S101: Obtain environmental data of a target area, match and analyze the environmental data with scene feature data in a scene library, and determine whether the environmental data belongs to a new scene condition based on similarity calculation;

[0056] S102: In response to the environmental data belonging to a new scene condition, for the environmental data, select an initial parameter set with the highest similarity to the scene feature data from a model snapshot library; wherein, the initial parameter set is used as the initial state of a local student model; the student model is a lightweight generative adversarial network;

[0057] S103: Use the student model to infer the environmental data to generate a dimming control instruction; obtain environmental feedback data after the lighting device executes the dimming control instruction;

[0058] S104: Use a teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction; the student model adjusts its own parameters based on the guidance information.

[0059] Regarding the above S101:

[0060] In this embodiment, step S101 is mainly used to identify the scene condition where the current target area is located and determine whether it belongs to a new scene, so as to perform cold start initialization or directly reuse existing model snapshots in the subsequent process.

[0061] Specifically: First, environmental data is obtained through a variety of sensors (such as illuminance sensors, weather monitors, cameras, and pedestrian flow detectors, etc.) set within the target area. The environmental data includes illuminance values, weather types (such as sunny, rainy, hazy, etc.), time period information, and pedestrian flow values. Then, the environmental data is preprocessed (such as filtering, normalization, or feature extraction) to form a set of feature vectors describing the current environmental state.

[0062] Next, the above-mentioned feature vectors are subjected to matching analysis with the scene feature data in the pre-constructed scene library. The scene library can store several typical scenes and their corresponding feature vectors based on different environmental data collected historically or expert rules. For example, it is classified according to different climates (sunny, rainy, snowy), different time periods (daytime, night), different pedestrian flow intensities (peak, low peak), etc.

[0063] Furthermore, in order to quantitatively measure the similarity, this embodiment can adopt Euclidean distance, cosine similarity, or other common similarity measurement methods to compare the feature vectors of the current environmental data with each scene feature vector in the scene library one by one and calculate the similarity value. When the similarity value of a certain scene is greater than or equal to the preset threshold, it is determined that the current environmental data matches this scene; if the similarity values of all scenes are less than the preset threshold, it is determined that the current environmental data belongs to a new scene condition.

[0064] Through the above process, this application can quickly identify or classify scenes in a changing outdoor environment and transmit the "new scene condition" information to subsequent steps to perform corresponding cold start initialization and student model update and other processes.

[0065] Regarding the above S102:

[0066] After identifying that the current environmental data belongs to a new scene condition, this embodiment selects the initial parameter set with the highest similarity to the scene feature data from the model snapshot library for initializing the local student model. The model snapshot library can be composed of several model weights obtained from offline training or solidified during previous operations, and is labeled according to typical weather types, pedestrian flow characteristics, and lighting requirements when established. By comparing with the scene feature data and calculating the similarity, the model parameters closest to the current new scene can be quickly located from these snapshots, reducing the deviation of the dimming strategy caused by the new scene.

[0067] Specifically, when the system detects that all scene features have no high similarity with the existing snapshots, it can also trigger random initialization or adopt a conservative initial strategy to ensure the most basic lighting function.

[0068] When the local student model is initialized, the parameters of the target model snapshot are loaded into its lightweight generative adversarial network structure. Among them, the generator network is mainly used to generate dimming strategies for environmental data, and the discriminator network can cooperate in subsequent steps to perform adversarial learning.

[0069] Due to the small scale of this model, it is more suitable for real-time inference and update on the terminal side with limited computing power, greatly shortening the adjustment time in new scenarios and improving the adaptability of outdoor lighting systems to environmental changes.

[0070] Regarding the above S103:

[0071] After the initialization of the student model is completed, the environmental data obtained in this embodiment (such as illuminance feature vectors, weather categories, pedestrian flow information, etc.) is input into the generator network of the student model for forward inference to obtain a set of lighting control parameters.

[0072] Specifically, the student model can output a lighting plan for this scenario based on the lightweight adversarial learning structure, combined with the current environmental characteristics, such as the brightness level, color temperature of the lights, or time-sharing dimming strategy.

[0073] In order to implement this generation result on actual lighting devices, this embodiment further quantifies or maps the lighting control parameters according to the preset adjustment range on the device side (such as several preset brightness levels, variable color temperature ranges, etc.) to form specific dimming control instructions.

[0074] Subsequently, the dimming control instruction is sent to the LED lighting devices in the target area through the control module, and the device is driven to perform the corresponding dimming operation.

[0075] For example, raise the brightness of the lamp post group to a certain level or switch to a low-energy consumption mode during a specified period. At this time, the system will monitor the environmental effects after the lighting devices are executed through various previously set sensors (the same sensors as in step S101), including illuminance, energy consumption changes, road visibility, and pedestrian flow induction. The obtained environmental feedback data will record the difference between the actual execution result and the expected plan, providing a basis for the evaluation and guidance of the teacher model in step S104.

[0076] With the above process, this embodiment can perform real-time inference and dimming control with a small model scale on the local side, and collect the actual feedback data to continuously optimize the lighting plan in a changing outdoor environment.

[0077] Regarding the above S104:

[0078] After the system obtains the environmental feedback data after the lighting device executes the dimming control instruction, it uploads the feedback data together with the control instruction previously sent to the lighting device to the cloud teacher model. The cloud teacher model can deeply evaluate the dimming effect of the student model in this scenario based on large-scale historical data or a more complex adversarial network structure.

[0079] For example, the teacher model can comprehensively score the execution results using multi-objective metrics (such as illuminance uniformity, glare index, device energy consumption, etc.), and simulate or compare the differences between the expected lighting state and the actual observed values in the built-in discriminant / simulation module.

[0080] Subsequently, based on the evaluation results and difference information, the teacher model generates "guidance information" for the student model. This information can be the gradient value, soft label, or directly recommended dimming strategy for correcting the parameters of the student model. After returning the guidance information to the local end through the network, the student model can apply it to the fine-tuning process of the generator or discriminator network, so as to output a better dimming instruction in subsequent scenarios.

[0081] For example, in this embodiment, the environmental feedback and student output can be uploaded uniformly after a certain period or number of executions (such as every several inference cycles). The teacher model generates a guidance information with a relatively large step update once after batch analysis in the cloud; or in the case where the network environment permits, real-time evaluation and immediate feedback can be performed more frequently. In this way, the student model can better integrate the large-scale learning ability and multi-objective optimization strategy in the cloud on the basis of rapid on-site adaptation, and continuously iterate to improve the outdoor lighting effect.

[0082] In the prior art, knowledge distillation (KD) usually refers to the offline training stage, where a larger and more accurate teacher model (Teacher) guides the student model (Student) to learn, so that the student model approaches the level of the teacher model in terms of accuracy or generation effect, while significantly reducing the model size and inference overhead. Such solutions generally use the output (soft label or intermediate features) of the teacher model as an additional supervision signal during training to help the student model better converge in scenarios with insufficient data annotation or complex distributions. However, most of these knowledge distillation techniques are limited to the offline training process. Once the model is deployed and the external environment changes, it often needs to be re-finely tuned or re-trained offline and cannot adapt quickly.

[0083] Based on this, the present application expands on the traditional teacher-student structure: during the actual operation phase (inference phase), the cloud teacher model is continuously retained and used to collect and evaluate real-time environmental feedback data, thereby dynamically generating "guidance information" for the student model, including adversarial loss gradients, policy correction suggestions, soft labels, etc. The student model makes small-step updates or local fine-tuning based on this guidance information locally (at the edge), and can quickly adapt to changing outdoor lighting scenarios. Compared with traditional offline distillation, the solution of the present application introduces a real-time teacher-student interaction mechanism, which not only retains the lightweight and high efficiency of the student model during inference, but also continuously obtains the high-precision support of the teacher model (the large cloud model), and continuously makes targeted optimizations and evolutions to the lighting strategy, greatly improving the system's ability to handle light pollution and multi-objective constraints (such as illuminance, energy consumption, glare index, etc.) in the actual outdoor environment.

[0084] Through the above improvements, the present invention can also perform teacher-student interaction during the inference phase. Compared with the traditional method of only distilling during offline training, it is more suitable for rapidly changing outdoor scenarios and greatly shortens the adaptation cycle to new scenarios and new conditions. At the same time, the cold start mechanism and model snapshot library introduced in the present application further enhance the system's initial adjustment ability for unknown scenarios, jointly constituting an online collaboration framework of "cloud teacher model - local student model" to achieve efficient and adaptive control of LED outdoor light pollution suppression.

[0085] Exemplarily, an LED lighting system is deployed at an intersection of a certain urban road. The system is equipped with a variety of sensors, including an illuminometer, a weather monitor, a pedestrian flow detector, and a camera. During the initial stage, the lighting system has established a "model snapshot library" through offline training and prior historical data analysis, which stores several GAN model snapshots trained or solidified under different typical scenarios (such as sunny days during the day, rainy nights, peak pedestrian flow, low pedestrian flow, etc.); and a "teacher model" with a larger scale or higher precision is retained on the cloud server.

[0086] During the winter evening period, the weather at the intersection suddenly changes from light rain to heavy snow, and the pedestrian flow also rapidly decreases due to the bad weather. The system detects environmental information such as a significant decrease in the current brightness through the illuminometer, obtains the snow day signal through the weather monitor, and observes the decrease in pedestrians through the pedestrian flow detector, and converts it into a set of feature vectors.

[0087] The system first preprocesses the feature vector locally (such as normalization, filtering), and compares its similarity with the existing scene feature vectors in the scene library; since the scene of snowing in the evening with low pedestrian flow may not be fully covered or the similarity is insufficient, the system determines that it "belongs to new scene conditions", and transmits this information to the next step S102 for cold start initialization.

[0088] Since it is determined as a new scenario, the system extracts the model snapshot parameters that are most similar to "rain and snow weather" or "low light at night" from the "model snapshot library". For example, there may be a snapshot corresponding to the scenario of "rainy and cloudy night, low pedestrian flow" in the library. Although it is not a perfect match for the current situation, the similarity is the highest (e.g., 0.78 ≥ the threshold of 0.75), so it is selected by the system.

[0089] It should be noted that if the similarity of any snapshot does not exceed 0.75, the system can "trigger random initialization or adopt a conservative initial strategy", but in this example, a "relatively close" scenario snapshot is successfully found;

[0090] The system loads the snapshot parameters into the generator and discriminator structures of the local student model (a lightweight GAN) (if the discriminator also needs to be initialized locally), forming the "initial state" of the student model.

[0091] After loading the initial parameters, the system enters the online inference mode:

[0092] For example, the real-time environmental features collected at the current moment (such as heavy snow at night, intersection brightness about 20 lux, very low pedestrian flow, wind force level 2, etc.) are input into the generator network of the student model;

[0093] The generator network outputs a set of dimming parameters for controlling the street lights, such as:

[0094] Brightness level: mapped from 0 - 10 levels to 8 levels; color temperature: 2900K; dimming time period: 20:00 - 23:00;

[0095] The system further quantifies these dimming parameters into dimming control instructions and sends them to the LED street light group through a wireless or wired control module; after the street lights execute, the system continuously monitors in the next 10 minutes:

[0096] Whether the actual illuminance reaches the expectation (such as 70 lux); whether the energy consumption is too high; whether the visibility of pedestrians is good (which can be assisted by camera image recognition for evaluation); glare index (glare sensor or calculation model);

[0097] The data collected form environmental feedback data, including the actual illuminance value of 60 lux, energy consumption of 85% of the rated power, pedestrian flow distribution, etc., and are packed for use in the next step S104 after a period of time.

[0098] In specific implementation, at a certain interval (for example, once every 15 minutes) or when network resources are sufficient, the system will upload the dimming control instructions output by the student model and the environmental feedback data of that period to the cloud teacher model for evaluation:

[0099] The cloud-based teacher model combines large-scale historical data (such as the optimal brightness level at night with multiple rain and snow weather conditions and multi-objective performance criteria) to deeply discriminate or simulate the dimming result of the student model this time.

[0100] Exemplarily, the evaluation results show that the actual illuminance is slightly lower than the target (60 lux < 70 lux), the energy consumption is relatively high (85% quota), and the glare index is within a reasonable range.

[0101] The teacher model converts the adversarial loss or multi-objective score of this difference information (for example, the need to increase brightness or optimize energy consumption) into "guidance information", which can be an instruction such as "it is recommended to increase the brightness to meet the pedestrian safety requirements, but optimizing energy consumption can further reduce the lamp post power during non-pedestrian periods". It can also specifically generate a set of gradient update values or soft labels to guide the local student model to converge towards the "higher brightness - lower energy consumption optimal solution" during subsequent dimming.

[0102] The guidance information is sent back to the student model through the network, and the student model performs a small-step update locally (fine-tuning some weights of the generator network) in order to output a better dimming strategy during the next inference.

[0103] In this way, after experiencing multiple cycle iterations, the student model can gradually adapt to the specific scenario of low pedestrian flow during a snowstorm at night, and find a lighting configuration that can not only ensure the pedestrian's field of vision but also be relatively energy-efficient and avoid glare. Compared with traditional GANs or pure manual strategies without a cloud-based teacher model, the response speed and adaptability to environmental changes have been significantly improved, truly realizing the adaptive suppression of LED outdoor light pollution.

[0104] As an optional implementation manner, it further includes: after completing the preset number of inferences of the student model and the guidance of the teacher model, the teacher model evaluates the current parameters of the student model for performance, and when the preset conditions are met, solidifies the current parameters of the student model into a new model snapshot and stores it in the model snapshot library.

[0105] In a specific implementation, in order to enable the student model to continuously accumulate adaptability during long-term operation and provide a more accurate initial strategy for subsequent cold starts, this embodiment performs a performance evaluation and snapshot solidification process after several inferences and teacher guidance. Specifically, when the system detects that the student model has completed the iteration of the preset number of times (for example, every 10 rounds of local inferences + cloud teacher guidance), the cloud-based teacher model will evaluate the overall performance of the current student model based on multi-objective metrics (such as illuminance uniformity, glare index, energy consumption level, etc.) and calculate the corresponding adversarial loss or multi-objective score value.

[0106] If the multi-objective score value is higher than a pre-set threshold (e.g., ≥80 points), or the adversarial loss is lower than a certain reference standard, it indicates that the student model has reached a relatively optimal level in this scenario. At this time, the teacher model can solidify the current parameters of the student model (i.e., the network weights of the generator and discriminator) into a new model snapshot, and store it in the model snapshot library by annotating "applicable scenario features" and other means.

[0107] In this way, when the system encounters a similar scenario next time (for example, the weather and the crowd flow pattern are similar), it can directly use this newly "verified excellent" model snapshot for cold start, thereby reducing the time for re-training or significant adjustment, and further improving the adaptation speed and dimming quality.

[0108] Conversely, if the evaluation result does not meet the expectation, the teacher model only gives temporary guidance or fine-tuning to the student model, without solidifying the model snapshot, so as to prevent too many sub-optimal versions from accumulating in the library. Through this mechanism of "selectively updating the model snapshot", this embodiment avoids the storage redundancy caused by frequent writing, and also ensures that most of the "verified" relatively optimal network parameters are saved in the model snapshot library, providing a more reliable initialization for subsequent new scenario conditions.

[0109] With the help of the above-mentioned dynamically refreshed model snapshot library, not only can the present invention gradually evolve a better dimming strategy in the existing scenarios, but also lay a foundation for the rapid cold start of potential new scenarios, effectively enhancing the long-term adaptability of the LED outdoor light pollution suppression system.

[0110] As an optional implementation manner, the performance evaluation includes: the teacher model conducts a multi-objective performance evaluation on the current parameters of the student model based on the environmental feedback data, and generates a multi-objective evaluation result;

[0111] Among them, the multi-objective performance evaluation includes: illuminance uniformity evaluation, device energy consumption evaluation, and glare index evaluation.

[0112] After completing several inferences and guidance from the cloud teacher model, the teacher model will conduct a performance evaluation on the current parameters of the student model. For this purpose, this embodiment measures the lighting strategy of the student model in the current scenario from three dimensions of illuminance uniformity, device energy consumption, and glare index by analyzing the environmental feedback data collected by the sensor.

[0113] Regarding the illuminance uniformity evaluation:

[0114] The teacher model can calculate the brightness distribution of the lighting area based on the data obtained by the road surface illuminance sensor or camera, and use common uniformity formulas or other industry standards to determine whether the light and dark distribution of the road surface is uniform.

[0115] Regarding the device energy consumption evaluation:

[0116] Based on the power meter or energy consumption record of the lighting device, the actual power consumption of the device within a certain period is statistically calculated and compared with the theoretical power or expected energy consumption of the corresponding gear to determine whether the dimming strategy output by the local student model achieves the energy-saving goal in terms of energy consumption;

[0117] Regarding the glare index evaluation:

[0118] The system can be equipped with glare sensors or adopt calculation methods such as the UGR (Unified Glare Rating) model. The teacher model can quantify the visual comfort of pedestrians under the lighting state. If the glare index is too high, it indicates that the strategy generates unnecessary light pollution or safety hazards.

[0119] In actual implementation, the teacher model can combine these three evaluation results into a multi-objective evaluation result. For example, the comprehensive score (such as 0-100 points) is generated by means of weighted summation or interval scoring. If the comprehensive score exceeds a certain preset threshold (for example, ≥80), or the three indicators of illuminance uniformity, device energy consumption, and glare index all reach certain standards, it is considered that the current parameters of the student model perform excellently in multi-objective performance; otherwise, it is determined that there is still room for improvement. This multi-objective evaluation result not only provides a basis for subsequent judgment in step S104 on whether to solidify the snapshot, but also can guide the student model to focus on the optimization direction of specific indicators in the adversarial loss.

[0120] In this way, through such a multi-objective performance evaluation mechanism, this embodiment can not only avoid excessive pursuit under a single indicator (such as only pursuing energy conservation but ignoring the illuminance demand), but also achieve a balance between lighting comfort and energy consumption, so that the LED outdoor light pollution suppression system can always maintain a reasonable lighting effect and resource utilization rate in different scenarios.

[0121] Please refer to Figure 2 , Figure 2 which is a flowchart of a method for matching and analyzing environmental data with scene feature data in the scene library provided by an embodiment of this application, including steps S201 to S204, where:

[0122] S201: Extract the illuminance feature vector, pedestrian flow feature vector, and weather feature vector from the environmental data;

[0123] S202: Calculate the similarity between each of the feature vectors and the corresponding feature vectors of each scene in the scene library;

[0124] S203: Select the scene with the highest similarity and exceeding the preset threshold as the target scene;

[0125] S204: In response to all similarities being lower than the preset threshold, determine that the environmental data is a new scene condition.

[0126] In this embodiment, in order to more precisely identify the scenario of the outdoor environment, the present application splits and extracts feature vectors for the environmental data collected by the sensors.

[0127] In specific implementation, for the illuminance feature vector: the luminance values (unit: lux) of several key positions are collected by an illuminance sensor, or formed by a combination of the average value and the maximum and minimum values ;

[0128] For the pedestrian flow feature vector, pedestrian counting is performed according to a pedestrian flow detector or a camera, and the pedestrian flow intensity, change rate, etc. are statistically analyzed within a short-time window to obtain ;

[0129] For the weather feature vector, the current meteorological information, such as temperature, rainfall / snowfall, wind level, humidity, etc., is obtained by a weather monitor or an online meteorological API, and integrated into .

[0130] After splicing or combining the above three types of features, a complete environmental feature vector D can be formed:

[0131]

[0132] In the subsequent matching and analysis process, this embodiment will traverse the feature vectors of all registered scenarios in the scenario library , and perform similarity calculations on D and respectively. Among them, represents the feature vector of the i-th scenario in the scenario library, which is used to characterize the key environmental parameters under this scenario.

[0133] In terms of similarity measurement, the Euclidean distance (such as ) or cosine similarity and other common methods can be preferably adopted. Specifically:

[0134]

[0135] When the similarity result is greater than or equal to the preset threshold (for example ), it indicates that the current environmental data matches the scenario . This embodiment can select the scenario with the highest similarity and exceeding as the "target scenario" after calculating all scenarios, and transmit the scenario information to the subsequent steps (such as cold start initialization).

[0136] If all or calculation results in the scenario library fail to exceed the preset threshold , then this embodiment is regarded as a new scenario condition. The system can utilize the cold start mechanism or random initialization subsequently to avoid forcibly applying an inconsistent scenario snapshot, which may trigger an incorrect dimming strategy.

[0137] Through the above method, the three types of key feature vectors are extracted and matched step by step, which can not only finely distinguish different scenarios such as different weather (rain / snow, sunny), different pedestrian flows (peak, off-peak), and different illuminance levels (daytime, night), but also flexibly expand more environmental data dimensions in actual deployment. Compared with the traditional method of simply distinguishing scenarios only by illuminance or time period, this embodiment can more accurately determine whether a new scenario condition has been entered, providing higher-precision support for the subsequent adaptation of the LED dimming strategy.

[0138] As an alternative implementation, the selection of the initial parameter set with the highest similarity to the scenario feature data from the model snapshot library includes:

[0139] Perform scenario annotation on each model snapshot in the model snapshot library, where each model snapshot is associated with at least one scenario feature vector or scenario label;

[0140] Calculate the similarity between the feature vector of the current environmental data and the scenario feature vectors of each model snapshot respectively;

[0141] Sort by similarity and select the target model snapshot with the highest similarity and greater than the preset threshold;

[0142] Read the model parameters of the target model snapshot and use them as the initial parameter set of the student model;

[0143] In response to the similarity of no model snapshot being greater than the preset threshold, trigger random initialization.

[0144] In this embodiment, each snapshot in the model snapshot library is associated with at least one "scenario feature vector" or "scenario label", which may include main climate information, pedestrian flow range, illuminance interval, etc. Through offline training or historical operation processes, the system has performed "scenario annotation" on these snapshots, such as "rainy night", "sunny day", "off-peak evening with low pedestrian flow", etc.

[0145] When it is detected that the current environmental data (such as the feature vector generated by S101) matches a certain scenario or is determined to be a new scenario condition, this embodiment will traverse all the annotated scenario features in the model snapshot library to calculate the similarity;

[0146] The feature vector of the current environmental data Perform cosine similarity or Euclidean distance calculation with the scenario feature vectors attached to each model snapshot;

[0147] Sort all model snapshots in descending order of similarity and eliminate those model snapshots with similarity lower than a preset threshold;

[0148] If there is at least one model snapshot with similarity greater than or equal to the preset threshold, select the one with the highest similarity as the "target model snapshot"; if no snapshot similarity can reach the preset threshold, it is regarded as lacking a matching model snapshot, triggering "random initialization".

[0149] After determining the "target model snapshot", this embodiment reads the generative adversarial network parameters (i.e., generator and discriminator weights, or only generator weights) in this model snapshot and loads them as the initial parameter set of the student model to the terminal side. In this way, the student model can obtain a policy basis relatively close to the current scenario without starting from scratch, quickly reducing the possible lighting errors or dimming deviations in the new scenario.

[0150] When all model snapshots do not meet the threshold requirements, it means that the current environmental data is extremely different from the existing scenarios. In this case, the system will adopt the random initialization method to ensure that old model snapshots that do not match the environment seriously will not be loaded by mistake. This strategy also lays the foundation for dynamically generating new high-similarity model snapshots through online learning or cold start in the future.

[0151] Through the above process, fast retrieval and matching of model snapshots can be achieved in diverse outdoor environments, thus greatly reducing the mismatch risk during the initialization of the student model and saving a large amount of online retraining time.

[0152] As an alternative implementation, using the student model to infer the environmental data to generate a dimming control instruction includes:

[0153] Input the preprocessed environmental data into the generator network of the student model, perform forward inference, and obtain the lighting control parameters in the current scenario;

[0154] Based on the lighting control parameters and combined with the preset adjustment range of the lighting device, map the control parameters to a dimming control instruction.

[0155] In this embodiment, after the preprocessing of the current environmental data (such as normalization, feature vector splicing) is completed, the system inputs this data into the generator network of the student model for a forward pass. This generator network has been initialized in the early stage (such as through cold start or adversarial learning) and can output a set of lighting control parameters (such as brightness coefficient, color temperature setting K value, dimming time period, etc.) according to features such as weather, light, and people flow.

[0156] To enable these control parameters to be executed by actual lighting devices, in this embodiment, after the output of the generator network, with reference to the preset adjustment range of the target LED luminaires or pole groups (such as brightness levels from 0 to 10 and color temperature range from 2500K to 4000K), the above parameters are quantified or mapped.

[0157] Specifically, if the brightness coefficient output by the generator is 0.8, it can be correspondingly mapped to the 8th level of the device; if the color temperature parameter is 0.3, it can be converted to the actual command value of 3000K.

[0158] Finally, the system will organize these discretized or specific parameters into "dimming control commands" and send them to the LED luminaires through wired or wireless control modules (such as bus protocols, network protocols, etc.). Once the luminaires receive the corresponding commands, they will complete the adjustment of brightness, color temperature, or time-sharing operations according to the commands.

[0159] In this embodiment, depending on the compatibility of specific devices, multiple fields (brightness, color temperature, on-time period, etc.) can be encapsulated in the command format to achieve more flexible dimming.

[0160] Through this process, the continuous output of the generator network can be seamlessly mapped to the dimming commands executable by the devices, effectively reducing the intermediate conversion links and connecting with subsequent steps (such as sensor feedback collection) to achieve real-time or near-real-time light pollution suppression and lighting optimization in a changing outdoor environment.

[0161] As an alternative implementation, the student model further includes a discriminator network; the discriminator network is used for:

[0162] Receiving the dimming control parameters output by the generator network and the environmental feedback data;

[0163] Comparing the environmental feedback data with the expected target feedback to determine the adversarial loss;

[0164] Based on the adversarial loss, performing local gradient update on the generator network.

[0165] In this embodiment, in order to perform fast and fine-grained adversarial correction on the dimming strategy at the local end, the student model not only includes a generator network but also a discriminator network. Each time, the discriminator network will receive both the lighting control parameters (such as brightness level, color temperature setting, time-sharing strategy, etc.) output by the generator network and the environmental feedback data (such as the actual illuminance, energy consumption information, glare index, etc.) detected after execution, and compare the environmental feedback data with the "target feedback" preset by the system or obtained through online learning.

[0166] When the discriminator detects a large deviation between the output of the generator and the target feedback, it calculates the adversarial loss, which can reflect the degree of insufficient matching of the dimming strategy to the environmental requirements. If the adversarial loss is too high, it indicates that the strategy generated by the current generator cannot well meet the requirements such as lighting uniformity or energy consumption constraints in this environment.

[0167] In this embodiment, the adversarial loss output by the discriminator is mapped into a local gradient, and the corresponding weights of the generator network are updated in small steps (such as adjusting the parameters of certain layers of the generator).

[0168] This local adversarial update helps the student model optimize its response to environmental changes in real time. Even when not communicating frequently with the cloud teacher model, it can rely on the evaluation of the local discriminator to obtain a certain degree of adaptive correction. Complementary to the global and multi-objective evaluation done by the cloud teacher model in the previous step, the discriminator network mainly focuses on fast adversarial and immediate fine-tuning, which can significantly reduce the error accumulation of light pollution or excessive energy consumption in low-latency scenarios.

[0169] Through this mechanism, this embodiment can maintain a more flexible dimming strategy adaptation ability on edge devices with limited computing power, providing a technical solution with multi-level cooperation between local and cloud for the continuous optimization of outdoor LED lighting.

[0170] As an alternative implementation, the generating of the guidance information by using the teacher model in the cloud based on the environmental feedback data and the dimming control instruction includes:

[0171] Comparing and analyzing the environmental feedback data with preset performance indicators to determine the differences between lighting uniformity, device energy consumption, and glare index and the target requirements;

[0172] Running a discriminator network in the teacher model to evaluate the dimming strategy output by the student model and generating a simulation evaluation result;

[0173] Based on the multi-objective evaluation result and the simulation evaluation result, determining the difference information between the student model and the expected dimming effect;

[0174] Based on the difference information, generating guidance information for updating the student model.

[0175] In this embodiment, in order to more comprehensively evaluate the dimming effect of the student model in the current scenario, on the one hand, the teacher model makes multi-objective comparisons between the preset performance indicators (such as lighting uniformity, energy consumption standard, glare threshold, etc.) and the environmental feedback data; on the other hand, it starts a discriminator network or a simulation module in the teacher model to deeply evaluate the dimming strategy actually output by the student model, thereby generating a "simulation evaluation result".

[0176] For multi-objective evaluation, the system first reads the actual illuminance data, energy consumption records, and glare detection values, compares them with the predefined target requirements (such as illuminance uniformity ≥ 0.7, device energy consumption below a certain threshold, and glare index below a certain maximum tolerance), and calculates the difference degrees of illuminance uniformity, energy consumption, and glare from the ideal state.

[0177] For discriminator evaluation, in the teacher model, a more complex discriminator network or simulation engine can also be run to perform adversarial or simulation-based determination on the dimming strategies (such as brightness levels, color temperature settings, etc.) given by the student model. For example, the discriminator can compare the "ideal dimming distribution" with the "student model output" and give an adversarial loss or simulation score.

[0178] For comprehensive difference information, when the multi-objective evaluation results (such as the three scores of illuminance uniformity / energy consumption / glare) and the simulation results of the discriminator are obtained, the teacher model will perform weighted or parallel analysis on these results to determine the comprehensive difference information between the "current output of the student model" and the "desired dimming effect". If the difference is large, it means that the parameters of the student model need to be further corrected; if the difference is small, it means that the student model is relatively excellent.

[0179] Finally, based on the above difference information, the teacher model outputs a "guidance information" for updating the student model.

[0180] Exemplarily, the guidance information can be presented as:

[0181] Gradient update: directly give the gradients of some generator or discriminator layers to facilitate the local student model to perform backpropagation;

[0182] Soft label: provide the local student model with an additional target distribution in adversarial learning;

[0183] Suggested strategies: such as "significantly reduce the brightness during low-traffic periods to save energy" and "increase the brightness to maintain a safe illuminance".

[0184] The teacher model will send the guidance information to the local student model through the network. After receiving it, the latter can perform small-step parameter fine-tuning (such as local gradient update), so as to output dimming instructions that better meet the requirements of illuminance uniformity, low energy consumption, and low glare in subsequent scenarios. Through this process of "multi-objective evaluation, discriminator simulation, and difference information generation", the system can utilize the advantages of large-scale computing and adversarial evaluation in the cloud to provide real-time and accurate guidance for the online learning of the student model.

[0185] Exemplarily, a batch of LED lamps are deployed in a central square of a certain city for night landscape lighting and pedestrian passage. The system deploys a student model (lightweight GAN) locally, while a teacher model is deployed in the cloud, which contains a high-precision discrimination / simulation module and an evaluation logic integrating multiple objective metrics (such as illuminance uniformity, energy consumption, glare index, etc.).

[0186] On ordinary nights, the flow of people in the square is relatively even. The student model generator gives a set of dimming control instructions based on the current environment (stable flow of people, low illuminance requirement, energy consumption budget), mainly including:

[0187] Brightness level: gear 6; color temperature: 3200K; time period: slightly reduce the brightness from 22:00 to 00:00.

[0188] About 10 minutes after execution, the system obtains information such as the actual energy consumption being 78% of the budget value, the average road surface illuminance being about 55 lux, and the glare index G value being within the safe range during this period through illuminometers, energy consumption measurement modules, and glare detection sensors installed throughout the square. This environmental feedback data, together with the dimming control instructions such as "brightness gear = 6, color temperature = 3200K" issued by the student model at that time, is packaged and uploaded to the cloud teacher model.

[0189] The teacher model is pre-configured with a set of performance metrics: illuminance uniformity target ≥ 0.7; energy consumption recommendation ≤ 75% of the budget value; glare index G ≤ a certain safety threshold.

[0190] The teacher model compares the environmental feedback data with these preset performance metrics item by item and obtains the following information:

[0191] The illuminance uniformity is 0.68 (slightly lower than 0.7); the actual energy consumption is about 78% (higher than the upper limit of 75%), indicating a slight overrun; the glare index G is safe and measured within a reasonable threshold by the sensor.

[0192] Therefore, the teacher model initially determines that there are differences from the target in terms of illuminance uniformity and energy consumption.

[0193] The teacher model further runs its discriminator or simulation module to deeply evaluate the dimming strategy output by the student model:

[0194] The discriminator can make an adversarial comparison between "brightness gear = 6, color temperature = 3200K" and the ideal gear / color temperature distribution; or simulate through a simulation tool "if the brightness is only turned to gear 5 and the color temperature is 3000K during the period from 22:00 to 00:00", whether the energy consumption is lower and the illuminance uniformity is better?

[0195] Finally, the simulation evaluation results show that "moderately reducing the brightness and adjusting the color temperature" can, on the premise of ensuring the safety of pedestrian passage, balance the energy consumption target and the need for more uniform illuminance.

[0196] After synthesizing the multi-objective evaluation results and the discriminant / simulation evaluation results, the teacher model forms "difference information", indicating that the current dimming strategy of the student model is in:

[0197] Illuminance uniformity: about 0.02 different from the ideal value; Energy consumption: about 3% exceeding the recommended budget; Glare index: meets the requirements;

[0198] Moreover, based on the discriminator network or simulation analysis, the brightness can be moderately reduced (to level 5.5) and the color temperature can be lowered to 3100K during the period from 22:00 to 23:00 to balance the illuminance uniformity and energy consumption indicators.

[0199] After obtaining the above difference information, the teacher model generates a "guidance information". This can include:

[0200] Gradient update: For example, giving gradient correction to some weights of the generator network part to make it tend to lower brightness settings during subsequent inferences;

[0201] Strategy suggestion: For example, directly prompting that "the brightness should be set at level 5.5 and the color temperature is 3100K" and giving the corresponding time period range;

[0202] Soft label: If the adversarial training idea is adopted, an improved target distribution can be sent to the student model to make it more likely to converge to an excellent solution in the next round of inference.

[0203] The system then sends this guidance information back to the local student model through the network. After the student model performs the next inference or automatically conducts local fine-tuning (such as applying gradient update), it can output a better dimming instruction during the subsequent night time period, achieving a balance among energy consumption, illuminance uniformity, and glare.

[0204] With multiple rounds of cycling, the student model gradually converges to a strategy that better meets the night-time needs of the square, achieving lighting control that takes into account pedestrian safety, energy conservation, and landscape beauty. Through the multi-objective comparison of the cloud teacher model and the discriminant / simulation evaluation, and then outputting difference information and generating guidance information, the local student model can continuously perform online updates according to real-time feedback, significantly reducing light pollution and meeting the urban night lighting indicators.

[0205] Based on the same inventive concept, an LED outdoor light pollution suppression system combined with a generative adversarial network corresponding to the LED outdoor light pollution suppression method combined with a generative adversarial network is also provided in the embodiments of the present application. Since the principle of solving problems by the system in the embodiments of the present application is similar to that of the LED outdoor light pollution suppression method combined with a generative adversarial network in the above embodiments of the present application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be elaborated.

[0206] Referring to Figure 3 As shown, it is a schematic diagram of an LED outdoor light pollution suppression system combined with a generative adversarial network provided by an embodiment of the present application. The system includes:

[0207] A matching unit 10, configured to obtain environmental data of a target area, perform matching analysis on the environmental data and scene feature data in a scene library, and determine whether the environmental data belongs to a new scene condition based on a similarity calculation;

[0208] A cold start unit 20, configured to, in response to the environmental data belonging to a new scene condition, select, for the environmental data, an initial parameter set with the highest similarity to the scene feature data from a model snapshot library; wherein, the initial parameter set is used as the initial state of a local student model; the student model is a lightweight generative adversarial network;

[0209] A control unit 30, configured to use the student model to infer the environmental data and generate a dimming control instruction; obtain environmental feedback data after the lighting device executes the dimming control instruction;

[0210] A correction unit 40, configured to use a teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction; the student model adjusts its own parameters based on the guidance information.

[0211] Those of ordinary skill in the art can realize that, combining the units and algorithm steps of each example described in the embodiments disclosed in the present application, they can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

Claims

1. An LED outdoor light pollution suppression method combined with a generative adversarial network, characterized in that Including: Obtain the environmental data of the target area, perform matching analysis on the environmental data and the scene feature data in the scene library, and determine whether the environmental data belongs to a new scene condition based on the similarity calculation; In response to the environmental data belonging to a new scene condition, for the environmental data, select the initial parameter set with the highest similarity to the scene feature data from the model snapshot library; wherein, the initial parameter set is used as the initial state of the local student model; the student model is a lightweight generative adversarial network; Use the student model to infer the environmental data and generate a dimming control instruction; obtain the environmental feedback data after the lighting device executes the dimming control instruction; Use the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction; the student model adjusts its own parameters based on the guidance information; The step of using the student model to infer the environmental data and generate a dimming control instruction includes: Input the preprocessed environmental data into the generator network of the student model, perform forward inference, and obtain the lighting control parameters in the current scene; Based on the lighting control parameters and in combination with the preset adjustment range of the lighting device, map the control parameters to a dimming control instruction; It further includes: after completing the preset number of inferences of the student model and guidance of the teacher model, the teacher model evaluates the current parameters of the student model, and when the preset conditions are met, solidify the current parameters of the student model into a new model snapshot and store it in the model snapshot library; The performance evaluation includes: the teacher model performs multi-objective performance evaluation on the current parameters of the student model based on the environmental feedback data and generates a multi-objective evaluation result; wherein, the multi-objective performance evaluation includes: illumination uniformity evaluation, device energy consumption evaluation, and glare index evaluation; The step of using the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction includes: Compare and analyze the environmental feedback data with the preset performance indicators to determine the differences between the illumination uniformity, device energy consumption, and glare index and the target requirements; Run the discriminator network in the teacher model to evaluate the dimming strategy output by the student model and generate a simulation evaluation result; Based on the multi-objective evaluation result and the simulation evaluation result, determine the difference information between the student model and the expected dimming effect; Generate guidance information for updating the student model based on the difference information.

2. The method for suppressing LED outdoor light pollution by combining a generative adversarial network according to claim 1, characterized in that, The step of performing matching analysis on the environmental data and the scene feature data in the scene library includes: Extract the illuminance feature vector, pedestrian flow feature vector, and weather feature vector from the environmental data; Calculate the similarity between each of the feature vectors and the corresponding feature vectors of each scene in the scene library respectively; Select the scene with the highest similarity and exceeding the preset threshold as the target scene; In response to all similarities being lower than the preset threshold, determine that the environmental data is a new scene condition.

3. The method for suppressing LED outdoor light pollution by combining a generative adversarial network according to claim 1, characterized in that, The step of selecting the initial parameter set with the highest similarity to the scene feature data from the model snapshot library includes: Perform scene annotation on each model snapshot in the model snapshot library, where each model snapshot is associated with at least one scene feature vector or scene label; Calculate the similarity between the feature vector of the current environmental data and the scene feature vectors of each model snapshot respectively; Sort by similarity and select the target model snapshot with the highest similarity and greater than the preset threshold; Read the model parameters of the target model snapshot and use them as the initial parameter set of the student model; In response to the similarity of no model snapshot being greater than the preset threshold, trigger random initialization.

4. The method for suppressing LED outdoor light pollution by combining a generative adversarial network according to claim 1, characterized in that, The student model further includes a discriminator network; the discriminator network is used to: Receive the dimming control parameters output by the generator network and the environmental feedback data; Compare the environmental feedback data with the expected target feedback to determine the adversarial loss; Based on the adversarial loss, perform local gradient update on the generator network.

5. The LED outdoor light pollution suppression method combining a generative adversarial network according to claim 1, characterized in that The guidance information includes: parameter gradients, target distributions, and soft labels.

6. The LED outdoor light pollution suppression system combined with a generative adversarial network is used to implement the LED outdoor light pollution suppression method combined with a generative adversarial network according to any one of claims 1-5, and is characterized in that Includes: A matching unit, configured to obtain environmental data of a target area, perform matching analysis on the environmental data and the scene feature data in the scene library, and determine whether the environmental data belongs to a new scene condition based on similarity calculation; A cold start unit, configured to, in response to the environmental data belonging to a new scene condition, select, for the environmental data, an initial parameter set with the highest similarity to the scene feature data from the model snapshot library; wherein, the initial parameter set is used as the initial state of the local student model; the student model is a lightweight generative adversarial network; A control unit, configured to use the student model to infer the environmental data and generate a dimming control instruction; obtain the environmental feedback data after the lighting device executes the dimming control instruction; A correction unit, configured to use the teacher model in the cloud to generate guidance information based on the environmental feedback data and the dimming control instruction; the student model adjusts its own parameters based on the guidance information.

Images