Energy-saving display method and system for LED display screen

Abstract

This invention discloses an energy-saving display method and system for LED displays, relating to the field of energy saving in LED displays. The method includes: acquiring intermediate data of the current display content through a video decoder interface to generate a predicted display content feature vector set for at least one future display frame; constructing a feature and response database based on a clustering algorithm; updating and optimizing the feature and response database based on the content feature vectors of the current display frame acquired in real time; calculating the current compensation coefficient for each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device; outputting the optimal driving parameter combination based on a deep reinforcement learning model; generating multi-level control instructions based on the optimal driving parameter combination, and dynamically optimizing and updating the model parameters based on actual results. The advantages of this invention are: effectively achieving significant reduction in LED display power consumption and extending device lifespan while ensuring display quality and temperature balance.

Description

Technical Field

[0001] This invention relates to the field of energy saving for LED displays, and more specifically to an energy-saving display method and system for LED displays. Background Technology

[0002] LED displays, with their advantages of high brightness, long lifespan, and high reliability, have been widely used in outdoor advertising, commercial displays, stage performances, traffic guidance, and monitoring and command, becoming a core carrier of information display. However, as display sizes and resolutions continue to increase, the high power consumption of LED displays is becoming increasingly prominent: large-size outdoor screens can consume tens of kilowatts per screen, leading not only to high operating costs but also exacerbating the load and heat dissipation pressure on the power supply system. Simultaneously, the photoelectric characteristics of LED devices are significantly affected by temperature (temperature drift effect), and localized overheating can cause brightness / color deviations and accelerated device aging, further reducing display consistency and lifespan. Current energy-saving control of LED displays largely relies on passive dimming or fixed power limits, making it difficult to simultaneously achieve the three core objectives of "energy saving, display quality, and temperature balance," and failing to adapt to dynamically changing display content and complex operating conditions. Therefore, a forward-looking and intelligent energy-saving control solution is urgently needed to achieve high efficiency and stable system operation while ensuring display quality.

[0003] Existing energy-saving methods mostly employ fixed rules or simple brightness adjustment, adjusting the overall brightness based on ambient light intensity. They lack the ability to respond in real time to changes in the displayed content. When the content being played switches from a dark field to a bright field, the system often responds with a lag, causing a sudden spike in power consumption or a jump in brightness, which affects the viewing experience. Secondly, existing energy-saving solutions are mostly statically configured and remain unchanged once deployed. However, as LED devices age and usage scenarios change, the fixed control parameters will gradually become mismatched, leading to a decrease in energy-saving effect or a deterioration in display quality. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides an energy-saving display method and system for LED displays. This technical solution solves the common defects of existing energy-saving control schemes for LED displays mentioned in the background, such as prediction lag, rigid models, single compensation, multi-objective imbalance, coarse control granularity, and insufficient adaptive capability. These defects prevent the achievement of efficient and stable energy-saving operation while ensuring display quality and temperature balance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An energy-saving display method for LED displays includes: The intermediate data of the currently displayed content is obtained through the video decoder interface, and a set of feature vectors for the predicted display content of at least one future display frame is generated. Tests were conducted to obtain the mapping relationship between the feature vector of the displayed content and the thermal response and light output based on the clustering algorithm, and a feature and response database was constructed. The feature and response database is updated and optimized based on the content feature vector of the current display frame acquired in real time. A hardware acquisition device is set up to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. With the optimization goals of reducing power consumption, improving display consistency, and balancing temperature, a multi-objective reward function is constructed, and the optimal combination of driving parameters is output based on a deep reinforcement learning model. Based on the optimal combination of driving parameters, multi-level control commands are generated, and the model parameters are dynamically optimized and updated based on actual results.

[0006] Preferably, the step of obtaining intermediate data of the currently displayed content through the video decoder interface and generating a predicted display content feature vector set for at least one future display frame specifically includes: The intermediate data during the decoding process of the current content display frame is obtained through the video decoder interface. The intermediate data includes: motion vector field, reference frame index and inter-frame prediction residual features. Based on motion vector field data, each motion vector contains a horizontal displacement component, a vertical displacement component, and the reference frame index referenced by the image patch; The motion velocity of each image block is calculated based on the horizontal and vertical displacement components, and the image blocks are divided into stationary blocks, slow-moving blocks, and fast-moving blocks according to the magnitude of the motion velocity. The angle of motion direction is calculated based on the ratio of the horizontal displacement component to the vertical displacement component. The corresponding reference frame information is found based on the reference frame index. It is determined whether the reference direction of each image block is forward reference or backward reference, and the frame interval between the current frame and the reference frame is calculated as the reference distance. Obtain the inter-frame prediction residual data for each image block, and calculate the mean of the absolute values ​​of the inter-frame prediction residuals within each image block as the average residual energy. Based on the magnitude of the average residual energy, image patches are divided into low-variation patches, medium-variation patches, and high-variation patches; The motion change type is identified based on the preset residual energy average threshold 'a' and motion velocity threshold 'b': If the average residual energy is greater than a and the motion speed is less than or equal to b, then it is determined to be a brightness change type. If the average residual energy is greater than a and the motion speed is greater than b, then it is judged to be a rapid motion change type; If the residual energy is less than or equal to a and the motion speed is greater than b, then it is judged to be a smooth motion type; If the average residual energy is less than or equal to a and the velocity is less than or equal to b, then it is judged to be a stationary or quasi-stationary type. Based on the motion change type identification results, prediction information for at least one future display frame is generated according to motion speed, motion direction and reference distance; The predicted multi-frame feature information is temporally aligned and normalized to generate a feature vector set of the predicted display content.

[0007] Preferably, the step of conducting the test experiment, based on a clustering algorithm, to obtain the mapping relationship between the display content feature vector and the thermal response and light output, and to construct a feature and response database specifically includes: The LED display screen was placed in a test dark chamber with a controlled ambient temperature. A gridded array of miniature temperature sensors was placed on the back of the display screen, and a spectroradiometer facing the display screen was placed in front of the screen. Play a sequence of calibration images covering various display content features. The parameter space of the calibration images includes: average image brightness, image complexity, proportion of bright areas, dynamic characteristics, and color distribution. For each calibration image, the display content feature vector, temperature feature vector, and photometric feature vector are acquired and displayed synchronously. Data was repeatedly collected under multiple ambient temperature conditions to construct a test sample set; The data in the test sample set is cleaned and normalized to remove outliers and map all features to the same numerical range. Based on the test sample set, the display content feature vector, the corresponding temperature response vector, and the photometric response vector of each sample are extracted; Principal component analysis algorithm is used to reduce the dimensionality of the display content feature vector, and principal component features with cumulative contribution rate reaching a preset cumulative contribution rate threshold are extracted to obtain the dimensionality-reduced display content feature vector. Using the dimensionality-reduced display content feature vector as input to the K-means clustering algorithm, the sample space is divided into K categories by iterative optimization to minimize the sum of squared Euclidean distances between samples within a class. For each category obtained from the division, perform the following calculation: Calculate the cluster center of this category, which is the mean vector of the dimensionality-reduced feature vectors of all samples in this category; Calculate the mean temperature response for this category, which is the mean of the temperature response vectors of all samples in this category; Calculate the mean photometric response of this category, which is the mean of the photometric response vectors of all samples in this category; Calculate the covariance vector of this category, which is the covariance vector of the dimensionality-reduced feature vectors of all samples in this category, and is used to characterize the correlation distribution of feature dimensions within this category; The calculation results are stored as a feature and response database, which includes a clustering index table and an original sample database. The clustering index table records the center coordinates, temperature response vector, and photometric response vector of each cluster; the original sample library retains all original samples and their corresponding complete feature vectors for fine matching.

[0008] Preferably, the hardware acquisition device, based on the temperature drift characteristic curve of the LED device, calculates the current compensation coefficient for each pixel area of ​​the display screen, specifically including: By conducting photoelectric parameter calibration tests on LED beads of the same batch under varying temperature conditions, the temperature drift characteristic curve of LED devices was constructed. A miniature temperature sensor array is deployed on the PCB board on the back of each cabinet of the display screen to collect the back temperature data of each LED bead area in real time and obtain the actual temperature distribution vector. A miniature photometric sensor array is deployed in the bezel area of ​​the display screen to collect scattered light signals in a specific area in real time. After calibration and conversion, the actual output brightness data of the display screen is obtained, and the actual brightness distribution vector is acquired. The power factor, conversion efficiency, and output voltage and current parameters of the input power supply are collected by the power monitoring device and used as the power supply operation characteristic vector. The actual temperature distribution vector is compared pixel by pixel with the clustered temperature response vector to calculate the temperature deviation vector. The actual brightness distribution vector is compared pixel by pixel with the clustered brightness response vector to calculate the brightness deviation vector. Input the temperature deviation vector into the temperature drift characteristic curve of the LED device to obtain the current compensation amount required to offset the temperature deviation. The brightness deviation is used as a comprehensive compensation amount for environmental interference, and it is weighted and fused with the current compensation amount to generate the final current compensation coefficient for each pixel region.

[0009] Preferably, the step of constructing a multi-objective reward function with the optimization objectives of reducing power consumption, improving display consistency, and balancing temperature, and outputting the optimal combination of driving parameters based on a deep reinforcement learning model, specifically includes: The predicted display content feature vector set, clustered temperature and brightness vectors, actual temperature and brightness distribution vectors, current compensation coefficients, and power supply operation characteristics are jointly encoded to form the state space; Define the executable drive parameter adjustment action space, including the PWM duty cycle modulation amount of each pixel area, the drive voltage adjustment value of each cabinet, and the scan refresh rate configuration level of the whole screen; We construct a weighted reward function with optimization goals of reducing power consumption, improving display consistency, and balancing temperature, and incorporates compensation for excessive penalties and rewards for matching content changes. By using a deep reinforcement learning model with LSTM temporal memory units, temporal features are extracted from historical state sequences, and the optimal combination of driving parameters that maximizes long-term cumulative reward is output. The reward value is calculated by collecting actual operating data after execution, and the model parameters are updated by sampling through the experience playback pool to achieve continuous optimization of energy-saving control strategy and adaptive closed-loop energy control.

[0010] Preferably, the step of generating multi-level control commands based on the optimal combination of driving parameters and dynamically optimizing and updating the model parameters based on actual results specifically includes: Based on the PWM duty cycle modulation amount in the optimal driving parameter combination, the driving current of each pixel is independently adjusted by the constant current driving IC so that the actual output brightness of each pixel approaches the theoretical brightness value. Based on the drive voltage adjustment value in the optimal drive parameter combination, the drive voltage of each enclosure is dynamically adjusted to transfer some of the power consumption of the high-load enclosure to the low-load enclosure. Based on the scan refresh rate configuration level and real-time power consumption requirements in the optimal drive parameter combination, the PFC circuit parameters of the parallel power supply module are dynamically adjusted through the master-slave control protocol. The three-level control commands are integrated into a unified multi-parameter collaborative control signal, which is applied to each execution unit of the LED display screen. Collect the actual total power consumption of the LED display screen, the actual power consumption of each cabinet, and the real-time temperature of each pixel area after executing the optimal combination of driving parameters. The collected actual operating data is compared with the clustered theoretical temperature distribution matrix and theoretical brightness distribution matrix to calculate the data deviation value; The actual reward value for the current action is calculated based on the data deviation value, and this reward value is used as a feedback signal for reinforcement learning. The current state, the action performed, the reward value obtained, and the new state after the action are combined into an experience tuple and stored in the experience replay pool. Batch of experience tuples are randomly sampled from the experience replay pool to update the network parameters of the deep reinforcement learning model, so that the model can continuously optimize its decision-making strategy. The updated deep reinforcement learning model is used to regenerate the combination of driving parameters, thereby continuously optimizing the energy-saving control strategy and forming an adaptive closed-loop energy control.

[0011] Furthermore, this solution proposes an energy-saving display system for LED displays, used to achieve the energy-saving display method for LED displays as described above, including: The feature extraction module is used to acquire intermediate data of the currently displayed content through the video decoder interface, generate a predicted display content feature vector set for at least one future display frame; conduct test experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database; update and optimize the feature and response database according to the content feature vector of the current display frame acquired in real time; and set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. The dynamic optimization module is used to construct a multi-objective reward function with the optimization goals of reducing power consumption, improving display consistency, and balancing temperature. Based on a deep reinforcement learning model, it outputs the optimal combination of driving parameters. According to the optimal combination of driving parameters, it generates multi-level control instructions and dynamically optimizes and updates the model parameters based on the actual results.

[0012] Preferably, the feature extraction module includes: The content feature unit is used to obtain intermediate data of the currently displayed content through the video decoder interface and generate a set of predicted display content feature vectors for at least one future display frame. A database unit is used to conduct testing experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database. A database optimization unit is used to update and optimize the feature and response database based on the content feature vector of the current display frame acquired in real time. The compensation coefficient unit is used to set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device.

[0013] Preferably, the dynamic optimization module includes: The parameter combination unit is used to construct a multi-objective reward function with the optimization goals of power consumption reduction, display consistency improvement and temperature balance, and output the optimal combination of driving parameters based on a deep reinforcement learning model. The model optimization unit is used to generate multi-level control instructions based on the optimal combination of driving parameters, and to dynamically optimize and update the model parameters based on the actual results.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an energy-saving display solution for LED displays. It acquires intermediate data such as motion vectors and residual energy during the decoding process of the displayed content in real time through a video decoder interface, predicting the content change trend and high-energy-consuming area distribution of future frames. Combined with an offline feature and response database, it accurately predicts the thermal response and light output state of each area of ​​the display screen, obtaining the photoelectric and thermal properties of the LED display screen under dynamic operating conditions. Furthermore, it integrates real-time collected temperature drift characteristics and current compensation data, aiming to reduce power consumption, improve display consistency, and balance temperature. A deep reinforcement learning model is used to perform multi-level coordinated control of pixel-level PWM duty cycle, cabinet-level drive voltage, and power supply-level refresh rate. This approach achieves a shift from "post-event response" to "proactive prediction," effectively improving the accuracy of energy-saving control and display consistency of LED displays under dynamic content, significantly reducing overall power consumption, balancing screen temperature, and thus effectively extending the lifespan of LED devices and improving the overall system operating efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of an energy-saving display method for an LED display screen according to the present invention; Figure 2 The flowchart of the present invention is as follows: intermediate data of the currently displayed content is obtained through the video decoder interface to generate a predictive display content feature vector set for at least one future display frame. Figure 3 The flowchart of the present invention is as follows: Based on the clustering algorithm, the feature vector of the displayed content is obtained to obtain the mapping relationship between thermal response and light output, and a feature and response database is constructed. Figure 4 This is a flowchart of the present invention for calculating the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of LED devices; Figure 5 The flowchart for constructing a multi-objective reward function based on a deep reinforcement learning model and outputting the optimal combination of driving parameters is provided for this invention. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 As shown, an energy-saving display method for an LED display screen includes: The intermediate data of the currently displayed content is obtained through the video decoder interface, and a set of feature vectors for the predicted display content of at least one future display frame is generated. Tests were conducted to obtain the mapping relationship between the feature vector of the displayed content and the thermal response and light output based on the clustering algorithm, and a feature and response database was constructed. The feature and response database is updated and optimized based on the content feature vector of the current display frame acquired in real time. A hardware acquisition device is set up to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. With the optimization goals of reducing power consumption, improving display consistency, and balancing temperature, a multi-objective reward function is constructed, and the optimal combination of driving parameters is output based on a deep reinforcement learning model. Based on the optimal combination of driving parameters, multi-level control commands are generated, and the model parameters are dynamically optimized and updated based on actual results.

[0018] It can be explained that when LED displays are playing dynamic content, their power consumption, heat generation, and display effects exhibit strong time-varying and spatial non-uniformity. Traditional fixed-parameter or simple feedback control is difficult to adapt to complex and ever-changing display scenarios. This solution achieves adaptive energy-saving control of LED displays through the following six steps: Step 1: Obtain intermediate data such as motion vector and residual energy of the current frame from the video decoder interface. Divide the image patch into four types according to the motion speed and residual size: brightness change, fast motion, smooth motion, and stillness. Predict the motion region, brightness change trend and high energy consumption region probability of the future frame respectively, and generate a prediction feature vector set. Step 2: Place the display screen in the test darkroom and play calibration images covering various types of content. Simultaneously collect display content features, temperature response, and photometric response data. Through PCA dimensionality reduction and K-means clustering, divide the samples into several typical categories and establish a feature and response database containing cluster centers and mean temperature / photometric responses for online matching. Step 3: Extract the content features of the current frame in real time, calculate the Mahalanobis distance between it and each cluster center, and obtain the theoretical temperature distribution and brightness distribution by weighting the inverse distance. For samples with low matching degree, the system caches them and periodically starts incremental clustering updates so that the database can continuously adapt to new content. Step 4: By deploying temperature sensors and photometers on the screen, the temperature distribution and brightness output of each area are collected in real time. The measured data are compared with the theoretical distribution to calculate the temperature deviation and brightness deviation. The temperature deviation is input into the LED temperature drift characteristic curve to obtain the current compensation amount. The current compensation amount is then weighted and fused with the brightness deviation to generate the final current compensation coefficient for each pixel area. Step 5: Encode the predicted features, theoretical distribution, measured data, compensation coefficients, etc. into a state space. Construct a reward function with the goals of power consumption reduction, consistency improvement and temperature balance. Output the optimal combination of driving parameters, including pixel-level PWM duty cycle, cabinet-level driving voltage and power supply-level refresh rate, through a deep reinforcement learning model with LSTM. Step Six: Transform the optimal parameters into three levels of control commands: pixel-level current adjustment, cabinet-level voltage adjustment, and power supply-level power matching. These commands are applied to the display screen, and the actual power consumption and temperature data after execution are collected. The reward value is calculated and fed back to update the model parameters, forming a continuously optimized adaptive closed loop. Through the coordinated efforts of these six steps, this solution achieves a leap from "post-event response" to "proactive prediction", from "global unification" to "pixel-level precision", and from "static configuration" to "dynamic learning". While ensuring display consistency, it significantly reduces power consumption, balances screen temperature, effectively extends the lifespan of LED devices, and improves the overall operating efficiency of the system.

[0019] Reference Figure 2 As shown, the step of obtaining intermediate data of the currently displayed content through the video decoder interface and generating a predicted display content feature vector set for at least one future display frame specifically includes: The intermediate data during the decoding process of the current content display frame is obtained through the video decoder interface. The intermediate data includes: motion vector field, reference frame index and inter-frame prediction residual features. Based on motion vector field data, each motion vector contains a horizontal displacement component, a vertical displacement component, and the reference frame index referenced by the image patch; The motion velocity of each image block is calculated based on the horizontal and vertical displacement components, and the image blocks are divided into stationary blocks, slow-moving blocks, and fast-moving blocks according to the magnitude of the motion velocity. The angle of motion direction is calculated based on the ratio of the horizontal displacement component to the vertical displacement component. The corresponding reference frame information is found based on the reference frame index. It is determined whether the reference direction of each image block is forward reference or backward reference, and the frame interval between the current frame and the reference frame is calculated as the reference distance. Obtain the inter-frame prediction residual data for each image block, and calculate the mean of the absolute values ​​of the inter-frame prediction residuals within each image block as the average residual energy. Based on the magnitude of the average residual energy, image patches are divided into low-variation patches, medium-variation patches, and high-variation patches; The motion change type is identified based on the preset residual energy average threshold 'a' and motion velocity threshold 'b': If the average residual energy is greater than a and the motion speed is less than or equal to b, then it is determined to be a brightness change type. If the average residual energy is greater than a and the motion speed is greater than b, then it is judged to be a rapid motion change type; If the residual energy is less than or equal to a and the motion speed is greater than b, then it is judged to be a smooth motion type; If the average residual energy is less than or equal to a and the velocity is less than or equal to b, then it is judged to be a stationary or quasi-stationary type. Based on the motion change type identification results, prediction information for at least one future display frame is generated according to motion speed, motion direction and reference distance; The predicted multi-frame feature information is temporally aligned and normalized to generate a feature vector set of the predicted display content.

[0020] This solution obtains intermediate decoding data through the video decoder interface, effectively utilizing the inter-frame prediction information naturally present during video compression encoding, thus avoiding the enormous computational overhead of performing complete decoding and analysis on each frame. When decoding the current frame, the video decoder already contains key information such as motion vectors and reference relationships for predicting subsequent frames. By obtaining this intermediate data through an appropriate interface, the content change trend of future frames can be predicted without adding extra decoding burden. The core of this prediction mechanism based on intermediate decoding data lies in using motion vector fields to reflect the direction and speed of image block movement, using reference frame indices to clarify the reference relationship between each image block and subsequent frames, and using inter-frame prediction residual features to assess the magnitude of image content change. Combining these three elements constructs a statistical feature prediction of future frame content, providing data support for subsequent forward-looking energy-saving control. Each element in the predicted display content feature vector represents a predicted feature of a future frame, and each future frame typically contains the following structured information: Frame sequence number identifier: used to identify the temporal position of the predicted frame; Brightness spatial distribution trend: predicted direction and magnitude of brightness changes in each region; Characteristics of motion area distribution: predicted location, direction, and speed of motion; High-energy-consuming area prediction information: Predicts the location and duration of areas with high brightness and high saturation; Among them, for the type of brightness change, the spatial distribution trend of brightness is predicted; For types of rapid motion changes, predict the distribution characteristics of the motion region; For smooth motion types, predict the motion trajectory and location distribution characteristics; For stationary or quasi-stationary types, high-energy-consuming areas can be predicted by combining brightness information. It should be noted that thresholds a and b are dynamically determined based on statistical data from the current frame or historical frames. Specifically, the mean plus standard deviation is calculated, and the mean plus standard deviation is used as the high threshold, while the mean minus standard deviation is used as the low threshold.

[0021] Reference Figure 3 As shown, the step of obtaining the mapping relationship between the display content feature vector and thermal response and light output based on the clustering algorithm, and constructing a feature and response database specifically includes: The LED display screen was placed in a test dark chamber with a controlled ambient temperature. A gridded array of miniature temperature sensors was placed on the back of the display screen, and a spectroradiometer facing the display screen was placed in front of the screen. Play a sequence of calibration images covering various display content features. The parameter space of the calibration images includes: average image brightness, image complexity, proportion of bright areas, dynamic characteristics, and color distribution. For each calibration image, the display content feature vector, temperature feature vector, and photometric feature vector are acquired and displayed synchronously. Data was repeatedly collected under multiple ambient temperature conditions to construct a test sample set; The data in the test sample set is cleaned and normalized to remove outliers and map all features to the same numerical range. Based on the test sample set, the display content feature vector, the corresponding temperature response vector, and the photometric response vector of each sample are extracted; Principal component analysis algorithm is used to reduce the dimensionality of the display content feature vector, and principal component features with cumulative contribution rate reaching a preset cumulative contribution rate threshold are extracted to obtain the dimensionality-reduced display content feature vector. Using the dimensionality-reduced display content feature vector as input to the K-means clustering algorithm, the sample space is divided into K categories by iterative optimization to minimize the sum of squared Euclidean distances between samples within a class. For each category obtained from the division, perform the following calculation: Calculate the cluster center of this category, which is the mean vector of the dimensionality-reduced feature vectors of all samples in this category; Calculate the mean temperature response for this category, which is the mean of the temperature response vectors of all samples in this category; Calculate the mean photometric response of this category, which is the mean of the photometric response vectors of all samples in this category; Calculate the covariance vector of this category, which is the covariance vector of the dimensionality-reduced feature vectors of all samples in this category, and is used to characterize the correlation distribution of feature dimensions within this category; The calculation results are stored as a feature and response database, which includes a clustering index table and an original sample database. The clustering index table records the center coordinates, temperature response vector, and photometric response vector of each cluster; the original sample library retains all original samples and their corresponding complete feature vectors for fine matching.

[0022] This can be explained by the fact that, building upon the feature vector generated in the previous stage for predicting the displayed content, this solution further addresses the problem of establishing a mapping relationship between the displayed content and the thermal / optical response. Due to the dynamic display characteristics of LED displays, the displayed content in different areas and at different times will lead to differentiated heat generation and brightness output, and this mapping relationship is highly nonlinear and time-varying, making it difficult to describe directly with a simple mathematical model. Therefore, this solution constructs a feature and response database by combining offline testing with cluster analysis, providing data support for rapid matching and accurate prediction in the online stage. Specifically, this solution places the LED display screen in a test darkroom with a controllable ambient temperature. By playing a series of calibration images covering parameters such as average image brightness, image complexity, proportion of bright areas, dynamic characteristics, and color distribution, it simulates various display scenarios that may occur in real applications. For each calibration image, display content characteristics, temperature response, and photometric response are collected synchronously to establish the correspondence between input and output. By repeatedly collecting data at different ambient temperatures, the influence of environmental factors on thermal / optical response is further considered. The calibration image sequence is essentially a series of representative static frames that cover various possible display content modes. In the data processing stage, the collected samples are first cleaned and normalized to ensure data quality. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional display content features, reducing computational complexity while retaining the main information. In this embodiment, the preset cumulative contribution rate threshold is 95%, meaning that the extracted principal components can retain more than 95% of the information in the original data. Finally, using the dimensionality-reduced features as input, the K-means clustering algorithm is used to divide the sample space into K categories. The K value is determined by the elbow method, that is, the K value at the inflection point of the curve of the sum of squared clustering errors as a function of K value is selected as the optimal number of clusters. For each category obtained by the division, not only are its cluster center, mean temperature response and mean photometric response calculated, but also the covariance vector of the category is calculated to characterize the correlation distribution of the feature dimensions within the category. The covariance vector can be used in the online stage to calculate the Mahalanobis distance between the sample to be matched and the cluster center. The final feature and response database adopts a hierarchical structure: the cluster index table records the center coordinates, temperature response vector, and photometric response vector of each cluster, which is used for fast coarse matching in the online stage; the original sample library retains all original samples and their complete feature vectors, which are used for fine matching when the coarse matching accuracy is insufficient, thus balancing response speed and matching accuracy.

[0023] The step of updating and optimizing the feature and response database based on the content feature vector of the current display frame acquired in real time specifically includes: The content feature vector of the current display frame is acquired in real time, and after principal component transformation, its Mahalanobis distance with each cluster center is calculated. The top M clusters with the smallest distance are selected as candidate nearest neighbors, and the theoretical temperature distribution and theoretical brightness distribution of the current display frame are calculated based on the inverse distance weighting. Calculate the distance between the current feature and the nearest neighbor cluster center. If the distance is greater than a preset threshold, it means that the current displayed content is significantly different from all historical patterns in the database, and the frame is marked as a low-confidence sample. Cache low-confidence samples and their corresponding measured data, and start incremental clustering update after the number of samples reaches the set value; After the update is complete, the updated database will be deployed to the online features and response database, enabling the database to continuously adapt to new content features.

[0024] This can be explained by the fact that during the online prediction phase, the model searches for historical patterns in the feature space that are most similar to the currently displayed content using distance metrics. These historical patterns are then weighted and averaged based on their temperature and photometric responses to obtain the current theoretical prediction value. The inverse distance weighting ensures that more similar samples contribute more significantly. The accumulation of low-confidence sample data and online updates enable the model to continuously learn. As the system's runtime increases, the database covers more and more display content features and responses, leading to continuously improving prediction accuracy and reliability. This allows the model to adapt to scenarios not covered in the offline phase, such as personalized user habits and new display content formats. The incremental clustering update process involves storing samples marked as low-confidence and their corresponding measured data (including content feature vectors, measured temperature vectors, and measured brightness vectors) into a cache queue. A maximum capacity for the cache queue is set. When the number of cached samples reaches a preset trigger threshold, the incremental clustering update process is initiated. A fixed update interval is also set. If the time since the last update is less than this interval, samples continue to accumulate until the time condition is met before initiating the update, avoiding frequent updates that could cause system oscillations. Based on the statistical distribution of the distances from historical samples to their respective cluster centers, merging and splitting thresholds are set. For each new sample in the cache queue, its Mahalanobis distance to each existing cluster center is calculated, and the minimum distance and its corresponding cluster are identified. If the minimum distance is less than or equal to the merging threshold, the new sample is assigned to the corresponding cluster, and the cluster's center coordinates, mean temperature response, etc., are updated incrementally. The brightness response mean and covariance vector are estimated. If the minimum distance is greater than the merging threshold but less than or equal to the splitting threshold, the sample is temporarily stored in the boundary buffer. After accumulating a sufficient number of similar samples, a new cluster is considered to be split from the original cluster. If the minimum distance is greater than the splitting threshold, a new cluster is created, initialized with the sample as the center, with an initial sample count of 1, an initial temperature response mean of the sample's measured temperature, and an initial brightness response mean of the sample's measured brightness. After the incremental update is completed, a new version of the feature and response database is generated. The total offset of the cluster centers of the new database compared to the original database is calculated. If the total offset exceeds a preset threshold, a smooth switching mechanism is activated. The smooth switching adopts a weighted linear transition method, gradually increasing the weight of the new database and decreasing the weight of the original database over multiple consecutive control cycles to ensure a smooth transition of control parameters and avoid system oscillations caused by database mutations. The calculation of the theoretical temperature distribution and theoretical brightness distribution of the current display frame includes: The weight of the i-th candidate nearest neighbor is obtained by dividing the reciprocal of the distance of the i-th candidate nearest neighbor by the sum of the reciprocals of the distances of the M candidate nearest neighbors. The theoretical temperature distribution is the sum of the products of the mean temperature of the M candidate nearest neighbors and their corresponding weights; The theoretical brightness distribution is the sum of the products of the brightness mean of the M candidate nearest neighbors and their corresponding weights.

[0025] Reference Figure 4 As shown, the calculation of the current compensation coefficient for each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device specifically includes: By conducting photoelectric parameter calibration tests on LED beads of the same batch under varying temperature conditions, the temperature drift characteristic curve of LED devices was constructed. A miniature temperature sensor array is deployed on the PCB board on the back of each cabinet of the display screen to collect the back temperature data of each LED bead area in real time and obtain the actual temperature distribution vector. A miniature photometric sensor array is deployed in the bezel area of ​​the display screen to collect scattered light signals in a specific area in real time. After calibration and conversion, the actual output brightness data of the display screen is obtained, and the actual brightness distribution vector is acquired. The power factor, conversion efficiency, and output voltage and current parameters of the input power supply are collected by the power monitoring device and used as the power supply operation characteristic vector. The actual temperature distribution vector is compared pixel by pixel with the clustered temperature response vector to calculate the temperature deviation vector. The actual brightness distribution vector is compared pixel by pixel with the clustered brightness response vector to calculate the brightness deviation vector. Input the temperature deviation vector into the temperature drift characteristic curve of the LED device to obtain the current compensation amount required to offset the temperature deviation. The brightness deviation is used as a comprehensive compensation amount for environmental interference, and it is weighted and fused with the current compensation amount to generate the final current compensation coefficient for each pixel region.

[0026] This solution works by collecting the actual temperature and brightness of the display screen in real time and comparing them with the theoretical response obtained through clustering, calculating the temperature deviation and brightness deviation for each pixel region. The temperature deviation is input into the temperature drift characteristic curve of the LED device to obtain a precise current compensation amount based on the physical mechanism, which is used to offset the brightness change caused by temperature drift. The brightness deviation is used as a comprehensive compensation amount to compensate for non-temperature drift factors such as device aging and ambient light interference. The two are weighted and fused to generate the final current compensation coefficient for each pixel region, achieving closed-loop precise control of the display screen brightness and ensuring that the display effect is always consistent with the theoretical prediction under various operating conditions.

[0027] Reference Figure 5 As shown, the construction of the multi-objective reward function, based on a deep reinforcement learning model, and the output of the optimal combination of driving parameters specifically includes: The predicted display content feature vector set, clustered temperature and brightness vectors, actual temperature and brightness distribution vectors, current compensation coefficients, and power supply operation characteristics are jointly encoded to form the state space; Define the executable drive parameter adjustment action space, including the PWM duty cycle modulation amount of each pixel area, the drive voltage adjustment value of each cabinet, and the scan refresh rate configuration level of the whole screen; We construct a weighted reward function with optimization goals of reducing power consumption, improving display consistency, and balancing temperature, and incorporates compensation for excessive penalties and rewards for matching content changes. By using a deep reinforcement learning model with LSTM temporal memory units, temporal features are extracted from historical state sequences, and the optimal combination of driving parameters that maximizes long-term cumulative reward is output. The reward value is calculated by collecting actual operating data after execution, and the model parameters are updated by sampling through the experience playback pool to achieve continuous optimization of energy-saving control strategy and adaptive closed-loop energy control.

[0028] This can be explained by the fact that, based on the clustered temperature / brightness and the actual temperature / brightness, this solution further introduces a deep reinforcement learning model to solve the multi-objective optimization decision problem. Reinforcement learning is needed because energy-saving control faces three conflicting optimization objectives: reducing power consumption requires decreasing drive current, but this leads to a decrease in brightness; improving display consistency requires compensating for temperature deviations, but this increases power consumption; and balancing screen temperature requires distributing the load, but this may sacrifice local display effects. Traditional fixed-rule or single-objective optimization methods struggle to balance these conflicting objectives in dynamically changing display content. The weighted reward function designed in this solution is precisely to quantify these optimization objectives and guide the reinforcement learning model to learn the optimal decision strategy. This reward function comprises five components: Power consumption reduction reward: It is proportional to the amount of reduction in the total measured power consumption after the action is performed. The greater the power consumption reduction, the greater the reward value, which encourages the model to save energy as much as possible while ensuring display quality. Display consistency bonus: It is inversely proportional to the sum of the squares of the differences between the actual brightness and the theoretical brightness in each area. The smaller the deviation, the greater the bonus. It encourages the model to make the actual output approach the theoretical value through current compensation, so as to ensure the uniformity of the screen. Temperature balance reward: It is inversely proportional to the sum of the squares of the differences between the actual temperature and the average temperature in each region. The more balanced the temperature distribution, the greater the reward. This encourages the model to distribute heat throughout the screen and avoid local overheating that could accelerate device aging. Overcompensation penalty: When the current compensation coefficient exceeds the preset safety threshold, a negative reward is given to prevent the model from excessively increasing the current in pursuit of consistency, which would lead to a surge in power consumption or device overload. Content change matching reward: The reward is proportional to the degree of matching between the current driving parameters and the best parameters predicted by future frame features. The higher the degree of matching, the greater the reward, which encourages the model to adjust in advance according to future content. These five rewards are weighted and summed using coefficients to form a comprehensive reward signal, ensuring the model receives clear feedback after each decision, indicating whether the current action is good or bad. Secondly, the deep reinforcement learning model with LSTM temporal memory units used in this solution has the core advantage of extracting temporal patterns from historical state sequences. Since the content displayed on LED screens has strong temporal correlation, the content of the previous frame often determines the trend of the next frame. LSTM (Long Short-Term Memory), as a special type of recurrent neural network, can remember long-term dependencies, learning patterns such as "what precursors usually precede the appearance of a certain type of high-dynamic content" or "what brightness changes often follow a specific content pattern" from the state sequences of the past N moments. For example, the model can learn that "in the first three frames of a full white screen, there are usually transition frames with gradually increasing brightness," thus adjusting the driving parameters in advance to avoid brightness jumps caused by response lag during screen transitions. In the specific decision-making process, the model first fuses the current state (including theoretical temperature / brightness, actual temperature / brightness, current compensation coefficient, future frame features, current power consumption, etc.) with the temporal features extracted by LSTM to form a complete understanding of the current situation. Then, through the decision layer of the deep Q-network, it evaluates the expected cumulative reward that can be obtained by taking each possible action in the current state, and finally selects the action with the largest expected reward as the optimal combination of driving parameters for output. This action includes the PWM duty cycle modulation amount of each pixel area, the driving voltage adjustment value of each cabinet, and the scanning refresh rate configuration level of the entire screen, which is a set of multi-dimensional fine control instructions. Through this mechanism, the reinforcement learning model is no longer simply responding to the current deviation, but can predict future trends based on historical experience, find a dynamic balance point among multiple objectives such as power consumption, consistency, and thermal balance, and achieve truly intelligent energy-saving control. It should be noted that deep reinforcement learning models with LSTM temporal memory units specifically include: The input to LSTM is a sequence of historical states over multiple consecutive time points. The state at each time point includes the predicted display content features, theoretical temperature and brightness distribution, actual temperature and brightness distribution, current compensation coefficient, and power supply operation features. The historical state sequence is constructed using a sliding window method, with the window length N set according to the system response time and the frequency of content changes. The output of LSTM is a hidden state vector that incorporates temporal information. This vector not only contains the state information at the current moment, but more importantly, it encodes the change patterns contained in the historical state sequence. The connection between the LSTM layer and the reinforcement learning network is as follows: the hidden state vector output by the LSTM is concatenated with the state vector at the current time step and input together into the subsequent fully connected layer. Finally, the output layer generates the Q value of each action. This structure enables the model to fully refer to the temporal patterns learned from historical experience when evaluating the value of each action in the current state. The training method uses the backpropagation algorithm over time. During experience replay, it not only samples single-step experience, but also samples an experience sequence of length N, so that the LSTM layer can learn the temporal dependencies between states.

[0029] The process of generating multi-level control commands based on the optimal combination of driving parameters, and dynamically optimizing and updating model parameters based on actual results, specifically includes: Based on the PWM duty cycle modulation amount in the optimal driving parameter combination, the driving current of each pixel is independently adjusted by the constant current driving IC so that the actual output brightness of each pixel approaches the theoretical brightness value. Based on the drive voltage adjustment value in the optimal drive parameter combination, the drive voltage of each enclosure is dynamically adjusted to transfer some of the power consumption of the high-load enclosure to the low-load enclosure. Based on the scan refresh rate configuration level and real-time power consumption requirements in the optimal drive parameter combination, the PFC circuit parameters of the parallel power supply module are dynamically adjusted through the master-slave control protocol. The three-level control commands are integrated into a unified multi-parameter collaborative control signal, which is applied to each execution unit of the LED display screen. Collect the actual total power consumption of the LED display screen, the actual power consumption of each cabinet, and the real-time temperature of each pixel area after executing the optimal combination of driving parameters. The collected actual operating data is compared with the clustered theoretical temperature distribution matrix and theoretical brightness distribution matrix to calculate the data deviation value; The actual reward value for the current action is calculated based on the data deviation value, and this reward value is used as a feedback signal for reinforcement learning. The current state, the action performed, the reward value obtained, and the new state after the action are combined into an experience tuple and stored in the experience replay pool. Batch of experience tuples are randomly sampled from the experience replay pool to update the network parameters of the deep reinforcement learning model, so that the model can continuously optimize its decision-making strategy. The updated deep reinforcement learning model is used to regenerate the combination of driving parameters, thereby continuously optimizing the energy-saving control strategy and forming an adaptive closed-loop energy control.

[0030] This can be explained by the fact that the hierarchical control execution of this scheme includes three levels, each corresponding to a different control mechanism: The principle of pixel-level current regulation: Modern LED constant current driver ICs generally support PWM dimming. By inputting a square wave signal with a specific duty cycle into the DIM pin, the switching action of the internal power transistor is controlled. When the PWM signal is high, the power transistor is turned on, and the LED beads pass through the set constant current value; when the PWM signal is low, the power transistor is turned off, and the LED bead current is zero. The average current of the LED bead is equal to the product of the constant current value and the PWM duty cycle. Since the PWM frequency is higher than 100Hz, the human eye cannot perceive the switching process, only the change in average brightness. Therefore, this solution converts the PWM duty cycle modulation output by the reinforcement learning model into a square wave signal with the corresponding duty cycle, and inputs it to the DIM pin of the driver IC to achieve precise adjustment of the current of each pixel. The principle of cabinet-level power consumption balancing: Different display content in each cabinet of an LED display leads to varying power loads, resulting in uneven voltage distribution. This solution employs two control methods: First, it dynamically adjusts the equivalent resistance value using a digital potentiometer to maintain a consistent total voltage across the branches of different load cabinets. Second, it utilizes the IFB terminal of the LED driver chip connected to the FB port of the power management system. When a cabinet's port voltage deviates from the standard value, the driver chip generates a reference current of corresponding magnitude and direction, feeding it back to the FB port, thereby dynamically adjusting the output voltage of the power management system. For example, when a cabinet's voltage is low due to high brightness display, the feedback current flows out of the FB port, increasing the output voltage; conversely, it decreases the output voltage, achieving dynamic power consumption balancing between cabinets. The principle of power matching regulation at the power supply level: The power factor correction (PFC) circuit of the switching power supply achieves input current following input voltage by controlling the conduction time of the switching transistor. This solution coordinates the working state of multiple parallel power supplies based on the refresh rate configuration level and real-time power consumption requirements of the reinforcement learning model output, through a master-slave control protocol: when the total power consumption is low, some slave power supplies are put into sleep mode, and the master power supply operates at its optimal efficiency point; when the total power consumption increases, the slave power supplies are gradually woken up to share the load. At the same time, by adjusting the reference voltage and switching frequency of the PFC control circuit, the overall power factor is kept above 0.98, reducing reactive power loss.

[0031] Furthermore, based on the same inventive concept as the energy-saving display method for LED displays described above, this solution proposes an energy-saving display system for LED displays, comprising: The feature extraction module is used to acquire intermediate data of the currently displayed content through the video decoder interface, generate a predicted display content feature vector set for at least one future display frame; conduct test experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database; update and optimize the feature and response database according to the content feature vector of the current display frame acquired in real time; and set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. The dynamic optimization module is used to construct a multi-objective reward function with the optimization goals of reducing power consumption, improving display consistency, and balancing temperature. Based on a deep reinforcement learning model, it outputs the optimal combination of driving parameters; based on the optimal combination of driving parameters, it generates multi-level control instructions, and dynamically optimizes and updates the model parameters according to the actual results. The feature extraction module includes: The content feature unit is used to obtain intermediate data of the currently displayed content through the video decoder interface and generate a set of predicted display content feature vectors for at least one future display frame. A database unit is used to conduct testing experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database. A database optimization unit is used to update and optimize the feature and response database based on the content feature vector of the current display frame acquired in real time. The compensation coefficient unit is used to set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. The dynamic optimization module includes: The parameter combination unit is used to construct a multi-objective reward function with the optimization goals of power consumption reduction, display consistency improvement and temperature balance, and output the optimal combination of driving parameters based on a deep reinforcement learning model. The model optimization unit is used to generate multi-level control instructions based on the optimal combination of driving parameters, and to dynamically optimize and update the model parameters based on the actual results.

[0032] In summary, the advantages of this invention are: it uses intermediate data from video decoding to predict the displayed content, combines offline clustering modeling with online adaptive updates, and achieves multi-objective optimization through deep reinforcement learning with temporal memory. While ensuring display quality and temperature balance, it significantly reduces the power consumption of LED displays and extends device lifespan.

[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An energy-saving display method for LED displays, characterized in that, include: The intermediate data of the currently displayed content is obtained through the video decoder interface, and a set of feature vectors for the predicted display content of at least one future display frame is generated. Tests were conducted to obtain the mapping relationship between the feature vector of the displayed content and the thermal response and light output based on the clustering algorithm, and a feature and response database was constructed. The feature and response database is updated and optimized based on the content feature vector of the current display frame obtained in real time. A hardware acquisition device is set up to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. With the optimization goals of reducing power consumption, improving display consistency, and balancing temperature, a multi-objective reward function is constructed, and the optimal combination of driving parameters is output based on a deep reinforcement learning model. Based on the optimal combination of driving parameters, multi-level control commands are generated, and the model parameters are dynamically optimized and updated based on actual results.

2. The energy-saving display method for LED displays according to claim 1, characterized in that, The step of obtaining intermediate data of the currently displayed content through the video decoder interface and generating a predicted display content feature vector set for at least one future display frame specifically includes: The intermediate data during the decoding process of the current content display frame is obtained through the video decoder interface. The intermediate data includes: motion vector field, reference frame index and inter-frame prediction residual features. Based on motion vector field data, each motion vector contains a horizontal displacement component, a vertical displacement component, and the reference frame index referenced by the image patch; The motion velocity of each image block is calculated based on the horizontal and vertical displacement components, and the image blocks are divided into stationary blocks, slow-moving blocks, and fast-moving blocks according to the magnitude of the motion velocity. The angle of motion direction is calculated based on the ratio of the horizontal displacement component to the vertical displacement component. The corresponding reference frame information is found based on the reference frame index. It is determined whether the reference direction of each image block is forward reference or backward reference, and the frame interval between the current frame and the reference frame is calculated as the reference distance. Obtain the inter-frame prediction residual data for each image block, and calculate the mean of the absolute values ​​of the inter-frame prediction residuals within each image block as the average residual energy. Based on the magnitude of the average residual energy, image patches are divided into low-variation patches, medium-variation patches, and high-variation patches; The motion change type is identified based on the preset residual energy average threshold 'a' and motion velocity threshold 'b': If the average residual energy is greater than a and the motion speed is less than or equal to b, then it is determined to be a brightness change type. If the average residual energy is greater than a and the motion speed is greater than b, then it is judged to be a rapid motion change type; If the residual energy is less than or equal to a and the motion speed is greater than b, then it is judged to be a smooth motion type; If the average residual energy is less than or equal to a and the velocity is less than or equal to b, then it is judged to be a stationary or quasi-stationary type. Based on the motion change type identification results, prediction information for at least one future display frame is generated according to motion speed, motion direction and reference distance; The predicted multi-frame feature information is temporally aligned and normalized to generate a feature vector set of the predicted display content.

3. The energy-saving display method for LED displays according to claim 2, characterized in that, The aforementioned testing and experimentation, based on clustering algorithms, involves obtaining the mapping relationship between the display content feature vector and thermal response and light output, and constructing a feature and response database, specifically including: The LED display screen was placed in a test dark chamber with a controlled ambient temperature. A gridded array of miniature temperature sensors was placed on the back of the display screen, and a spectroradiometer facing the display screen was placed in front of the screen. Play a sequence of calibration images covering various display content features. The parameter space of the calibration images includes: average image brightness, image complexity, proportion of bright areas, dynamic characteristics, and color distribution. For each calibration image, the display content feature vector, temperature feature vector, and photometric feature vector are acquired and displayed synchronously. Data was repeatedly collected under multiple ambient temperature conditions to construct a test sample set; The data in the test sample set is cleaned and normalized to remove outliers and map all features to the same numerical range. Based on the test sample set, the display content feature vector, the corresponding temperature response vector, and the photometric response vector of each sample are extracted; Principal component analysis algorithm is used to reduce the dimensionality of the display content feature vector, and principal component features with cumulative contribution rate reaching a preset cumulative contribution rate threshold are extracted to obtain the dimensionality-reduced display content feature vector. Using the dimensionality-reduced display content feature vector as input to the K-means clustering algorithm, the sample space is divided into K categories by iterative optimization to minimize the sum of squared Euclidean distances between samples within a class. For each category obtained from the division, perform the following calculation: Calculate the cluster center of this category, which is the mean vector of the dimensionality-reduced feature vectors of all samples in this category; Calculate the mean temperature response for this category, which is the mean of the temperature response vectors of all samples in this category; Calculate the mean photometric response of this category, which is the mean of the photometric response vectors of all samples in this category; Calculate the covariance vector of this category, which is the covariance vector of the dimensionality-reduced feature vectors of all samples in this category, and is used to characterize the correlation distribution of feature dimensions within this category; The calculation results are stored as a feature and response database, which includes a clustering index table and an original sample database. The clustering index table records the center coordinates, temperature response vector, and photometric response vector of each cluster; the original sample library retains all original samples and their corresponding complete feature vectors for fine matching.

4. The energy-saving display method for an LED display screen according to claim 3, characterized in that, The step of updating and optimizing the feature and response database based on the content feature vector of the current display frame acquired in real time specifically includes: The content feature vector of the current display frame is obtained in real time, and after principal component transformation, its Mahalanobis distance with each cluster center is calculated. The top M clusters with the smallest distance are selected as candidate nearest neighbors, and the theoretical temperature distribution and theoretical brightness distribution of the current display frame are calculated based on the inverse distance weighting. Calculate the distance between the current feature and the nearest neighbor cluster center. If the distance is greater than a preset threshold, it means that the current displayed content is significantly different from all historical patterns in the database, and the frame is marked as a low-confidence sample. Cache low-confidence samples and their corresponding measured data, and start incremental clustering update after the number of samples reaches the set value; After the update is complete, the updated database will be deployed to the online features and response database, enabling the database to continuously adapt to new content features.

5. The energy-saving display method for an LED display screen according to claim 4, characterized in that, The hardware acquisition device, based on the temperature drift characteristic curve of the LED device, calculates the current compensation coefficient for each pixel area of ​​the display screen, specifically including: By conducting photoelectric parameter calibration tests on LED beads of the same batch under varying temperature conditions, the temperature drift characteristic curve of LED devices was constructed. A miniature temperature sensor array is deployed on the PCB board on the back of each cabinet of the display screen to collect the back temperature data of each LED bead area in real time and obtain the actual temperature distribution vector. A miniature photometric sensor array is deployed in the bezel area of ​​the display screen to collect scattered light signals in a specific area in real time. After calibration and conversion, the actual output brightness data of the display screen is obtained, and the actual brightness distribution vector is acquired. The power factor, conversion efficiency, and output voltage and current parameters of the input power supply are collected by the power monitoring device and used as the power supply operation characteristic vector. The actual temperature distribution vector is compared pixel by pixel with the clustered temperature response vector to calculate the temperature deviation vector. The actual brightness distribution vector is compared pixel by pixel with the clustered brightness response vector to calculate the brightness deviation vector. Input the temperature deviation vector into the temperature drift characteristic curve of the LED device to obtain the current compensation amount required to offset the temperature deviation. The brightness deviation is used as a comprehensive compensation amount for environmental interference, and it is weighted and fused with the current compensation amount to generate the final current compensation coefficient for each pixel region.

6. The energy-saving display method for an LED display screen according to claim 5, characterized in that, The optimization objectives, namely reducing power consumption, improving display consistency, and balancing temperature, involve constructing a multi-objective reward function and, based on a deep reinforcement learning model, outputting the optimal combination of driving parameters. Specifically, this includes: The predicted display content feature vector set, clustered temperature and brightness vectors, actual temperature and brightness distribution vectors, current compensation coefficients, and power supply operation characteristics are jointly encoded to form the state space; Define the executable drive parameter adjustment action space, including the PWM duty cycle modulation amount of each pixel area, the drive voltage adjustment value of each cabinet, and the scan refresh rate configuration level of the whole screen; We construct a weighted reward function with optimization goals of reducing power consumption, improving display consistency, and balancing temperature, and incorporates compensation for excessive penalties and rewards for matching content changes. By using a deep reinforcement learning model with LSTM temporal memory units, temporal features are extracted from historical state sequences, and the optimal combination of driving parameters that maximizes long-term cumulative reward is output. The reward value is calculated by collecting actual operating data after execution, and the model parameters are updated by sampling through the experience playback pool to achieve continuous optimization of energy-saving control strategy and adaptive closed-loop energy control.

7. The energy-saving display method for an LED display screen according to claim 6, characterized in that, The process of generating multi-level control commands based on the optimal combination of driving parameters, and dynamically optimizing and updating model parameters based on actual results, specifically includes: Based on the PWM duty cycle modulation amount in the optimal driving parameter combination, the driving current of each pixel is independently adjusted by the constant current driving IC so that the actual output brightness of each pixel approaches the theoretical brightness value. Based on the drive voltage adjustment value in the optimal drive parameter combination, the drive voltage of each enclosure is dynamically adjusted to transfer some of the power consumption of the high-load enclosure to the low-load enclosure. Based on the scan refresh rate configuration level and real-time power consumption requirements in the optimal drive parameter combination, the PFC circuit parameters of the parallel power supply module are dynamically adjusted through the master-slave control protocol. The three-level control commands are integrated into a unified multi-parameter collaborative control signal, which is applied to each execution unit of the LED display screen. Collect the actual total power consumption of the LED display screen, the actual power consumption of each cabinet, and the real-time temperature of each pixel area after executing the optimal combination of driving parameters. The collected actual operating data is compared with the clustered theoretical temperature distribution matrix and theoretical brightness distribution matrix to calculate the data deviation value; The actual reward value for the current action is calculated based on the data deviation value, and this reward value is used as a feedback signal for reinforcement learning. The current state, the action performed, the reward value obtained, and the new state after the action are combined into an experience tuple and stored in the experience replay pool. Batch of experience tuples are randomly sampled from the experience replay pool to update the network parameters of the deep reinforcement learning model, so that the model can continuously optimize its decision-making strategy. The updated deep reinforcement learning model is used to regenerate the combination of driving parameters, thereby continuously optimizing the energy-saving control strategy and forming an adaptive closed-loop energy control.

8. An energy-saving display system for LED displays, characterized in that, An energy-saving display method for an LED display screen as described in any one of claims 1-7 includes: The feature extraction module is used to acquire intermediate data of the currently displayed content through the video decoder interface, generate a predicted display content feature vector set for at least one future display frame; conduct test experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database; update and optimize the feature and response database according to the content feature vector of the current display frame acquired in real time; and set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device. The dynamic optimization module is used to construct a multi-objective reward function with the optimization goals of reducing power consumption, improving display consistency, and balancing temperature. Based on a deep reinforcement learning model, it outputs the optimal combination of driving parameters. According to the optimal combination of driving parameters, it generates multi-level control instructions and dynamically optimizes and updates the model parameters based on the actual results.

9. An energy-saving display system for LED displays according to claim 8, characterized in that, The feature extraction module includes: The content feature unit is used to obtain intermediate data of the currently displayed content through the video decoder interface and generate a set of predicted display content feature vectors for at least one future display frame. A database unit is used to conduct testing experiments, obtain the mapping relationship between the display content feature vector and thermal response and light output based on clustering algorithms, and construct a feature and response database. A database optimization unit is used to update and optimize the feature and response database based on the content feature vector of the current display frame acquired in real time. The compensation coefficient unit is used to set up a hardware acquisition device to calculate the current compensation coefficient of each pixel area of ​​the display screen based on the temperature drift characteristic curve of the LED device.

10. An energy-saving display system for LED displays according to claim 9, characterized in that, The dynamic optimization module includes: The parameter combination unit is used to construct a multi-objective reward function with the optimization goals of power consumption reduction, display consistency improvement and temperature balance, and output the optimal combination of driving parameters based on a deep reinforcement learning model. The model optimization unit is used to generate multi-level control instructions based on the optimal combination of driving parameters, and to dynamically optimize and update the model parameters based on the actual results.

Images