AI algorithm-based LED backlight brightness uniformity optimization method

Abstract

This invention relates to the field of semiconductor display control technology and discloses an AI-based method for optimizing the brightness uniformity of LED backlights. The method includes: acquiring the physical state parameters and target brightness of the LED backlight array and constructing a physical topology map of the backlight; inputting this data into a graph neural network embedded with physical laws for network inference; combining preset optical superposition equations and thermodynamic diffusion equations to perform physical constraint calculations on the prediction results; outputting the target driving duty cycle and transient thermal risk index of each semiconductor LED; performing descending sorting and allocating phase offsets for local regions where the transient thermal risk index exceeds a threshold, determining the target starting phase, and generating a timing control signal; and driving the LED array according to the timing control signal. This invention compensates for optical diffusion crosstalk and thermal attenuation by embedding photothermal physical equations into network inference and performing time-domain phase misalignment driving.

Description

Technical Field

[0001] This invention relates to the field of semiconductor display control technology, specifically to an AI algorithm-based method for optimizing the brightness uniformity of LED backlights. Background Technology

[0002] As modern home display devices evolve towards higher dynamic range and more refined local dimming, high-density LED backlight arrays are widely used due to their superior contrast performance. However, with the dramatic increase in the density of light-emitting components, individual diodes are no longer isolated entities in physical space. Traditional display backlight control mechanisms generally employ a static open-loop lookup table strategy, relying solely on the pixel grayscale input from the image front-end to directly map the driving duty cycle. This purely data instruction-based process completely ignores the multidimensional physical coupling effects in complex hardware environments. Under actual operating conditions, densely lit light sources will experience disordered diffusion and crosstalk within the optical film, causing the actual light output to deviate from the theoretical value. Simultaneously, the heat released by the high-frequency conduction of components will rapidly spread laterally through the underlying substrate, causing severe heat accumulation in localized areas. This dynamic drift of the microscopic thermal field directly interferes with the band structure of the semiconductor, causing a nonlinear decay in electro-optical conversion efficiency, ultimately leading to frequent uneven brightness, localized overexposure, and color shift phenomena in large areas of high brightness or edges with strong contrast.

[0003] To compensate for display errors caused by physical distortions, existing technologies tend to introduce external temperature feedback mechanisms. However, conventional thermal monitoring heavily relies on external sensors such as thermistors mounted around the backplate or at key nodes. This crude hardware stacking not only significantly increases wiring complexity and material costs, but also, due to the hysteresis of the dielectric heat conduction path, the obtained data often has deep spatial blind spots and time lags, making it difficult to accurately reflect the transient junction temperature of the chips behind tens of millions of independent pixels. If the underlying driver circuit is used to directly measure the electrical parameters of the components, the sampling action is prone to conflict with the high-frequency light emission cycle due to the lack of a scientific time-domain isolation mechanism. This causes the detection signal to crosstalk the normal PWM waveform, resulting in visible flickering and visual artifacts on the screen. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based method for optimizing the brightness uniformity of LED backlights, which solves the problems of uneven brightness and thermal attenuation caused by neglecting optical spatial crosstalk and local heat accumulation in existing backlight control mechanisms.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an LED backlight brightness uniformity optimization method based on AI algorithms, comprising the following steps:

[0006] Obtain the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array;

[0007] Based on the spatial distribution of each semiconductor light-emitting diode, the physical state parameters, and the target brightness, a physical topology diagram of the backlight is constructed.

[0008] The physical topology map of the backlight is input into a preset physical law embedded graph neural network for network inference, and the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode are output. In the network inference process, the physical law embedded graph neural network combines the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculation on the prediction results output by the network, and iteratively adjusts the target driving duty cycle based on the calculated brightness prediction error, and updates the transient thermal risk index based on the junction temperature prediction value.

[0009] Based on the target driving duty cycle and transient thermal risk index corresponding to each semiconductor light-emitting diode, a descending sorting and phase offset are performed on the local area where the transient thermal risk index exceeds a preset threshold to determine the target starting phase of each semiconductor light-emitting diode, and a timing control signal containing the corresponding target starting phase and the corresponding target driving duty cycle is generated for each semiconductor light-emitting diode.

[0010] When constructing the physical topology of the backlight, each semiconductor light-emitting diode in the LED backlight array is used as a node in the topology graph. The initial feature vector of the node includes the target brightness and physical state parameters: the junction temperature value and aging decay factor at the current moment. Edges are constructed based on the physical distance between the semiconductor light-emitting diodes, and optical diffusion weights and thermal conduction weights are assigned.

[0011] Among them, the optical diffusion weight is used to characterize the spatial crosstalk rate of light between adjacent semiconductor light-emitting diodes, and is obtained by fitting the optical diffusion distribution function preset by the factory and historical experience measurement data; the thermal conduction weight is used to characterize the reciprocal of the thermal flow impedance between adjacent semiconductor light-emitting diodes, and is calculated in advance based on the thermal conductivity of the backlight substrate and the physical geometric distance.

[0012] During the network inference process of the graph neural network embedded in the physical laws, after the network performs single-layer feature updates, it calculates the predicted combined brightness value of each semiconductor light-emitting diode using a preset optical superposition equation. The optical superposition equation is:

[0013] ;

[0014] in: Indicates the first Predicted combined brightness of individual semiconductor light-emitting diodes; Indicates the relationship between the physical topology graph and the first A node has a set of adjacent nodes connected by edges, containing nodes. itself; Represents a node For nodes The optical diffusion weights are obtained by fitting historical empirical data; Represents a node The output is the target drive duty cycle; This represents the electro-optical conversion efficiency function, whose variables include nodes. junction temperature value and aging degradation factor The mapping table of the function is pre-calibrated based on laboratory temperature cycling and aging test data of different batches of light-emitting diodes; This represents the reference maximum brightness of a single semiconductor light-emitting diode at its rated power, and is a factory-calibrated constant.

[0015] When the calculated composite brightness prediction value When the absolute value of the difference between the brightness and the target brightness is greater than a preset brightness error threshold, the difference is taken as the brightness prediction error, and the network iteratively adjusts the output based on this difference using a preset numerical optimization algorithm. This process continues until the brightness error threshold condition is met, thereby achieving precise compensation for optical spatial crosstalk.

[0016] Simultaneously, during network inference, the predicted junction temperature of each semiconductor light-emitting diode at the next moment is calculated using a pre-defined thermodynamic diffusion equation. The thermodynamic diffusion equation is:

[0017] ;

[0018] in: Indicates the first Predicted junction temperature of a semiconductor light-emitting diode at the next moment; and They represent the first The node and the first The junction temperature (i.e., physical state parameter) of each node at the current moment. This represents the transient thermal resistance coefficient, obtained based on test data of the heat dissipation characteristics of the backlight module material. This indicates the rated input power of a single semiconductor light-emitting diode; Used to characterize the proportion of non-luminous dissipation in the conversion of electrical power into thermal power; This represents the set of adjacent nodes whose spatial distance is within a preset thermal influence range; Represents a node With nodes The heat conduction weight between them.

[0019] Predicted junction temperature at the next moment When the preset safe junction temperature threshold is exceeded, the network increases the transient thermal risk index output by the corresponding node by a preset step size or a preset ratio to complete the update.

[0020] Furthermore, to avoid the generation of transient thermal superposition peaks caused by adjacent LEDs in high-thermal-risk areas turning on simultaneously at the same time, this invention uses a transient thermal risk index for phase misalignment control. Specifically, the logic is as follows: It is determined whether each transient thermal risk index exceeds a trigger threshold, which is the preset threshold; if it does, the corresponding LED and its adjacent LEDs form a target set as the local region, and these are sorted in descending order to obtain the ranking. Initial phase offset The calculation formula is:

[0021] ;

[0022] in: The preset unit clock phase offset is determined by the minimum addressing clock cycle of the drive control chip. Then... The target starting phase is obtained by superimposing it with the initial phase and then packaged into a timing control signal to drive the corresponding semiconductor light-emitting diode to control its conduction duration and turn-on time.

[0023] In another aspect of the invention, the process of acquiring the physical state parameters of the semiconductor light-emitting diode is configured to be performed within the vertical blanking interval of the display system. Within this interval, the forward drive current is paused, and a reverse bias voltage is applied to collect the reverse junction capacitance and reverse leakage current. The reverse junction capacitance and reverse leakage current have a strict physical correspondence with the current junction temperature and aging degradation factor, and state conversion can be performed using a preset mapping function.

[0024] This invention also provides an LED backlight brightness uniformity optimization system based on AI algorithms, comprising:

[0025] The data acquisition module is used to acquire the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array;

[0026] The topology construction module is used to construct a backlight physical topology based on the spatial distribution of each of the semiconductor light-emitting diodes, as well as the physical state parameters and target brightness.

[0027] The inference module is used to input the physical topology map of the backlight into a preset physical law embedded graph neural network for network inference, and output the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode; wherein, during the network inference process, the physical law embedded graph neural network combines the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculation on the prediction results output by the network, and iteratively adjusts the target driving duty cycle based on the calculated brightness prediction error, and updates the transient thermal risk index based on the junction temperature prediction value;

[0028] The signal generation module is used to perform descending sorting and allocate phase offset for local areas where the transient thermal risk index exceeds a preset threshold based on the target driving duty cycle and transient thermal risk index corresponding to each semiconductor light-emitting diode, determine the target starting phase of each semiconductor light-emitting diode, and generate a timing control signal for each semiconductor light-emitting diode that includes the corresponding target starting phase and the corresponding target driving duty cycle.

[0029] A driving module is used to drive the LED backlight array according to the timing control signals.

[0030] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for optimizing the brightness uniformity of LED backlights based on AI algorithms.

[0031] This invention provides a method for optimizing the brightness uniformity of LED backlights based on AI algorithms. It offers the following advantages:

[0032] 1. This invention inputs the physical topology diagram of the backlight into an embedded graph neural network for inference, and combines the optical superposition equation and the thermodynamic diffusion equation to constrain the prediction results. This scheme simultaneously completes brightness error iteration and thermal risk update during inference, compensates for diffusion crosstalk between adjacent light sources and local thermodynamic dynamic drift, and solves the technical problems of uneven brightness and thermal attenuation in traditional control mechanisms.

[0033] 2. This invention performs descending sorting of local areas based on transient thermal risk index and assigns corresponding initial phase offset to semiconductor light-emitting diodes to generate timing control signals. This mechanism forces adjacent components with heat accumulation risk to be turned on out of time in the time domain, cuts off the physical conditions that cause a surge in heat flux density caused by multiple points being turned on at the same frequency, and eliminates the transient thermal superposition peak of high-density backlight arrays.

[0034] 3. This invention pauses the forward drive current in the vertical blanking interval of the display system and applies a reverse bias voltage to synchronously collect the reverse junction capacitance and leakage current, thereby calculating the physical state parameters. At this time, the domain slicing acquisition scheme realizes the low-level non-destructive extraction of physical state parameters, eliminates the hardware wiring cost of external thermal sensors, and eliminates the interference of detection actions on the actual light emission visual uniformity of the screen.

[0035] 4. This invention maps the spatial distribution of semiconductor light-emitting diodes into nodes of a topological graph and constructs connecting edges based on physical distances to allocate dual weights for optical diffusion and thermal conduction. This graph structure construction process transforms the array of light-emitting components into a mathematical structure containing the laws of optical and thermal cross-interference.

[0036] 5. This invention calculates the prediction error between the synthesized brightness prediction value and the target brightness after obtaining the prediction value, and uses a numerical optimization algorithm to iteratively adjust the target driving duty cycle along the negative gradient direction of the error function. This closed-loop feedback optimization architecture realizes the adaptive cancellation of spatial optical crosstalk. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0038] Figure 2 This is a schematic diagram of the timing and sampling logic of the present invention;

[0039] Figure 3 This is a schematic diagram of the node connection and weight relationship of the present invention;

[0040] Figure 4 This is a flowchart of the feature update and physical constraint calculation process of the present invention;

[0041] Figure 5 This is a schematic diagram of the driving phase misalignment control timing of the present invention;

[0042] Figure 6 This is a block diagram of the modular architecture of the present invention. Detailed Implementation

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see the appendix Figure 1 To be continued Figure 6This invention provides an AI-based method for optimizing the brightness uniformity of LED backlights, which includes the following steps:

[0045] Obtain the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array;

[0046] Based on the spatial distribution of each semiconductor light-emitting diode, the physical state parameters, and the target brightness, a physical topology diagram of the backlight is constructed.

[0047] The physical topology map of the backlight is input into a preset physical law embedded graph neural network for network inference, and the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode are output. In the network inference process, the physical law embedded graph neural network combines the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculation on the prediction results output by the network.

[0048] Based on the target driving duty cycle and transient thermal risk index corresponding to each of the semiconductor light-emitting diodes, the target starting phase of each semiconductor light-emitting diode is determined, and a timing control signal containing the corresponding target starting phase and the corresponding target driving duty cycle is generated for each of the semiconductor light-emitting diodes.

[0049] The LED backlight array is driven according to the timing control signals described above.

[0050] The optimization method in this embodiment operates in a display driving control system that includes high-density semiconductor light-emitting diodes (LEDs). The hardware environment of this system includes an image processing front-end, a driving control chip, and an LED backlight module composed of arrayed semiconductor light-emitting diodes. The output pins of the driving control chip are electrically connected to the control channels of each semiconductor light-emitting diode, and are used to output timing control signals with specific time lengths and start and end times.

[0051] In the aforementioned high-density hardware physical architecture, on the one hand, the light illuminating a single LED will spread outwards in a divergent manner through the optical film material in the backlight module, forming a halo and superimposing on adjacent areas; on the other hand, the heat released by the densely arranged diodes during continuous operation will be conducted laterally through the backlight substrate, changing the junction temperature of surrounding components, and thus directly causing the degradation of electro-optical conversion efficiency.

[0052] To accurately quantify and intervene in the aforementioned complex physical field coupling effects, this embodiment pre-establishes a macroscopic physical superposition model for the actual observed brightness of the backlight. The actual observed brightness in this basic superposition model is determined by the following formula:

[0053] ;

[0054] in: Coordinates of the light-emitting surface of the backlight module The actual observed brightness at the point is obtained by scanning the light-emitting surface with a high-precision optical color analyzer during the factory calibration stage. This parameter is used to characterize the objective light intensity ultimately received by the human eye or the panel. The total number of semiconductor light-emitting diodes that contribute to light emission is determined based on the backlight design drawings and circuit wiring logic. For the first The actual luminous intensity of a semiconductor light-emitting diode under its current operating temperature and driving conditions is non-linearly mapped to the driving current, driving duty cycle and its own thermal decay. For the first The optical dot spread function of a semiconductor light-emitting diode in spatial coordinates The weighting coefficient at the point reflects the crosstalk dispersion law of a single point light source in space. It is obtained by Gaussian surface fitting calibration based on the two-dimensional spatial light intensity distribution data of a single component after it is lit at the factory stage.

[0055] Traditional display driving mechanisms rely solely on open-loop lookup tables based on the pixel grayscale of the input image, completely ignoring diffusion crosstalk between adjacent light sources. and the actual luminous intensity due to thermodynamics Dynamic drift. This causes backlights to be prone to uneven brightness, local overexposure, or severe lifespan degradation at the boundaries of images with large areas of high brightness or strong contrast.

[0056] The system first extracts physical state parameters that characterize the current decay and heat status of components from the underlying hardware in a non-destructive manner. Combined with the target brightness required for the image content, the entire physical space is abstracted into a physical topology map carrying features.

[0057] Subsequently, this scheme inputs the topology graph into a specific graph neural network. This network integrates the parameters characterizing optical crosstalk and the law of thermal conduction into its weight update mechanism. By combining physical equations for constraint calculation, the network directly infers the target driving duty cycle that can accurately compensate for photothermal attenuation and simultaneously outputs the transient thermal risk index.

[0058] Finally, the drive control chip rearranges the originally synchronously turned-on LEDs in the time domain based on the acquired thermal risk index and duty cycle. By assigning independent target start phases, adjacent components with a risk of heat buildup are forced to be staggered in the microsecond time axis.

[0059] In this embodiment, traditional junction temperature measurement usually relies on external thermistors and other additional sensors, which not only increases the wiring complexity and material cost of the hardware system, but also, due to the limitation of the heat conduction path, the measurement results often have obvious spatial errors and time lags.

[0060] In this embodiment, the process of acquiring the physical state parameters is deeply integrated into the existing display timing control mechanism. Specifically, this step is configured to be executed within the vertical blanking interval of the display system to which the LED backlight array belongs. Since the optical scanning of one frame of image has just ended within the vertical blanking interval, and the display data of the next frame has not yet been loaded into the working register, the driving system actively suspends the output of forward driving current to each of the semiconductor light-emitting diodes during this stage. The application of this time-domain slicing technology completely avoids the interference of the detection action on the actual screen brightness and visual uniformity.

[0061] During the off-peak period when the forward drive current is turned off, the drive system applies a constant reverse bias voltage to each semiconductor light-emitting diode in the LED backlight array through an integrated detection circuit. For GaN-based LEDs, the reverse bias voltage is set to a range of -3V to -5V, preferably -4V, to ensure that a stable and measurable reverse leakage current can be generated without triggering avalanche breakdown. At the same time, a high-frequency AC small signal with a peak-to-peak value of 50mV and a frequency of 1MHz is superimposed on the DC bias reference. The corresponding static reverse leakage current is collected through a low-pass filter branch using a sample-and-hold circuit, and the corresponding dynamic reverse junction capacitance is collected synchronously through a high-pass filter impedance analysis branch.

[0062] After completing the digital quantization of the underlying electrical analog signals, the system converts the acquired electrical parameters into macroscopic physical state parameters according to a preset mapping function. The mapping function for calculating the junction temperature of each semiconductor LED at the current moment is as follows:

[0063] ;

[0064] in: Represents the calculated first... The junction temperature of a semiconductor light-emitting diode at the current moment, in degrees Celsius; This indicates the number of images collected in real time within the vertical blanking interval. The reverse leakage current value of a semiconductor light-emitting diode, in microamps; This indicates the standard reverse leakage current base value of this batch of semiconductor light-emitting diodes at the reference temperature, in microamps. This value is obtained from the semiconductor current-voltage characteristic curve in the component's factory test report and is pre-stored in the system. This represents the reference temperature constant corresponding to the standard reverse leakage current, for example, 25°C. This represents the logarithmic temperature sensitivity coefficient of leakage current; this coefficient is obtained in advance using the least squares fitting algorithm based on the multi-point temperature and leakage current corresponding scanning data under laboratory constant temperature chamber environment.

[0065] Simultaneously, to assess the long-term degradation of the luminous efficiency of the components, the mapping function for the aging degradation factor is calculated as follows:

[0066] ;

[0067] in: Indicates the first The aging degradation factor of a semiconductor light-emitting diode is used to characterize the proportion of irreversible decrease in the luminous efficiency of that node. This indicates the number of images collected in real time within the vertical blanking interval. The reverse junction capacitance value of a semiconductor light-emitting diode; This indicates the reference initial junction capacitance of the component at the beginning of its life cycle; this value is obtained according to the factory calibration test process and is read from the non-volatile memory by the control chip during the first power-on initialization. This represents the junction capacitance degradation conversion factor; this factor is pre-calibrated based on the mapping relationship between capacitance drift and absolute brightness decay of diodes of the same type within a standard accelerated aging test cycle.

[0068] In this embodiment, the system constructs a backlight physical topology map based on the spatial distribution of each semiconductor light-emitting diode and the extracted physical state parameters and target brightness, thereby abstracting the photothermal cross-coupling relationship of the backlight array in physical three-dimensional space into mathematical graph domain information that can be directly processed by graph neural networks.

[0069] In constructing the physical topology graph of the backlight, the system maps each semiconductor light-emitting diode in the LED backlight array to a node in the topology graph. For any node in the graph structure, the system combines the target brightness and the physical state parameters corresponding to that node to form the initial feature vector of that node.

[0070] Specifically, the first in the defined graph structure The feature vectors of the nodes at the initial time are as follows:

[0071] ;

[0072] in: Indicates the first Each semiconductor light-emitting diode serves as the initial feature vector corresponding to a node in the topological graph; Indicates the first The target brightness of each semiconductor light-emitting diode is obtained by the image processing module at the front end of the display system based on the pixel grayscale value of the current frame image in the corresponding local dimming area; This represents the junction temperature value at the current moment obtained from the aforementioned steps; This represents the aging degradation factor calculated in the aforementioned steps.

[0073] After establishing the nodes and their characteristic attributes, the system constructs the edges between the nodes based on the physical distance between each semiconductor light-emitting diode on the backlight substrate. Considering the spatial attenuation characteristics of the range of the physical field, the system uses a preset distance threshold as a criterion to establish edges only between two semiconductor light-emitting diode nodes whose physical distance is less than or equal to the distance threshold. This effectively eliminates invalid far-end weakly coupled calculation branches while ensuring the accuracy of the photothermal crosstalk model, significantly reducing the dimensionality of matrix operations.

[0074] After establishing edge connections, the system assigns dual physical attribute weights to each edge: optical diffusion weight and thermal conductivity weight. The optical diffusion weight characterizes the spatial crosstalk rate of light between adjacent semiconductor light-emitting diodes. The light dispersion through the diffuser and brightness enhancement film exhibits a Gaussian distribution, and the corresponding optical diffusion weight is determined by the following formula:

[0075] ;

[0076] in: Indicates from node To the node The optical diffusion weight; Indicates the first The semiconductor light-emitting diode and the first The absolute physical distance between the geometric centers of the semiconductor light-emitting diodes is determined by reading the factory printed circuit board wiring file of the backlight array. This represents the optical diffusion standard deviation coefficient of the backlight module; this coefficient characterizes the scattering and diffusion ability of each layer of optical film in the backlight module to light, and is obtained by fitting the light intensity spatial distribution curve obtained by performing a darkroom halo projection test on a single point light source during the factory calibration stage.

[0077] Meanwhile, the thermal conductivity weight is used to characterize the reciprocal of the thermal flow resistance between adjacent semiconductor light-emitting diodes. The heat generated by the semiconductor light-emitting diodes is mainly conducted in a two-dimensional plane through the copper-clad substrate and heat sink backplate, and its thermal conductivity is inversely proportional to the spatial distance. The corresponding thermal conductivity weight is determined by the following formula:

[0078] ;

[0079] in: Represents a node With nodes The heat conduction weights between them; This represents the equivalent thermal conductivity of the backlight substrate material, a parameter obtained directly from the material property data sheet provided by the substrate supplier. As mentioned above, this represents the absolute physical distance between two nodes; The reference thermal resistance constant represents the lateral conduction of the substrate. This constant is obtained through pre-simulation calculation based on the dielectric layer thickness of the backlight substrate, the copper foil trace area, and the thermal interface material properties of the heat dissipation structure.

[0080] In this embodiment, the physical law-embedded graph neural network undergoes an offline training phase before executing the network inference. The specific training process includes: constructing a historical backlight driving dataset, which contains a large amount of physical state parameters, input target brightness, and corresponding actual luminous brightness distribution data under various operating conditions; and constructing a loss function, which includes not only the mean square error between the predicted brightness and the actual brightness but also a physical penalty term constructed based on the optical superposition equation and the thermodynamic diffusion equation. The specific loss function... The formula is as follows:

[0081] ;

[0082] Among them: the first item is the predicted brightness. Compared to true brightness The mean square error between them; the second term is the thermodynamic over-limit penalty term. The first term is a preset safe body temperature threshold, and a penalty gradient is only generated when the tissue temperature exceeds this threshold; the third term is a duty cycle smoothing regularization penalty term, which is used to limit drastic changes in the duty cycle of adjacent light-emitting diodes. and These are the preset physical penalty term weight hyperparameters. An adaptive moment estimation optimizer is used to calculate the gradient based on the loss function and perform backpropagation. The weights and biases of each layer of the graph convolution kernel in the graph neural network are iteratively updated until the loss function converges, thereby saving and obtaining the preset physical law embedded graph neural network model. After completing the structured mapping of the graph data, the system inputs the backlight physical topology map into the preset physical law embedded graph neural network for network inference, thereby outputting the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode.

[0083] The physical law embedded graph neural network includes a multi-layer network structure. In the single-layer feature update process of the physical law embedded graph neural network, the network aggregates the previous layer feature vector of the current node with the previous layer feature vector of the adjacent nodes in the backlight physical topology graph that are connected to the current node based on the optical diffusion weight and thermal conduction weight corresponding to each edge, so as to obtain the current layer feature vector of the current node.

[0084] Specifically, the first In the layer network The feature update aggregation formula for each node is as follows:

[0085] ;

[0086] in: and Representing nodes respectively In the Layer and first The feature vector of the layer; and Indicates the network in the 1st... Layer-learnable weight matrix; Represents a nonlinear activation function; This is a normalization constant used to eliminate scale bias caused by node degree. Specifically, The calculation formula is:

[0087] ;

[0088] in: and These represent nodes in the physical topology graph. and nodes The total number of valid edges. The above formula enables the forced injection of physical weights along the graph feature dimension.

[0089] During the network inference process, the system simultaneously calculates the predicted combined luminance value and the predicted junction temperature value for each semiconductor light-emitting diode. For luminance prediction, the network uses the optical superposition equation, which includes the optical diffusion weights, the target driving duty cycle, and an electro-optical conversion efficiency function with the physical state parameters as variables, to calculate the predicted combined luminance value. The specific optical superposition equation is as follows:

[0090] ;

[0091] in: Indicates the first Predicted combined brightness of individual semiconductor light-emitting diodes; Indicates the relationship between the physical topology graph and the first A node has a set of adjacent nodes connected by edges, containing nodes. itself; Represents a node For nodes The optical diffusion weights are obtained by fitting historical empirical data; Represents a node The output is the target drive duty cycle; This represents the electro-optical conversion efficiency function, whose variables include nodes. junction temperature value and aging degradation factor The mapping table of the function is pre-calibrated based on laboratory temperature cycling and aging test data of different batches of light-emitting diodes; This represents the reference maximum brightness of a single semiconductor light-emitting diode at its rated power, and is a factory-calibrated constant.

[0092] After outputting the target drive duty cycle for the current round and obtaining the synthesized brightness prediction value, the network performs closed-loop calibration calculations. When the absolute value of the difference between the synthesized brightness prediction value and the target brightness in the input features is greater than a preset brightness error threshold, the network determines that the current drive state cannot achieve high-precision display restoration. At this time, based on the difference, a preset numerical optimization algorithm is used to iteratively adjust the target drive duty cycle output by the network until the absolute value of the difference is less than or equal to the preset brightness error threshold. The numerical optimization algorithm specifically adopts the gradient descent algorithm, which updates the duty cycle value along the negative gradient direction of the error function to achieve adaptive cancellation and compensation of spatial optical crosstalk. During the update process, the iteration step size of the duty cycle adopts a dynamic decay strategy: in the initial stage, a large preset step size is set, such as 0.05, to achieve fast convergence; when the absolute value of the difference shrinks to close to the brightness error threshold, the step size is automatically reduced to one-tenth of the current value to avoid oscillations near the optimal solution, thereby stably achieving adaptive cancellation and compensation of spatial optical crosstalk within a limited calculation period.

[0093] Simultaneously, the network uses the thermodynamic diffusion equation, which includes the heat conduction weights, the physical state parameters, and the target driving duty cycle, to synchronously calculate the predicted junction temperature for the next moment. The specific thermodynamic diffusion equation is as follows:

[0094] ;

[0095] in: Indicates the first Predicted junction temperature of a semiconductor light-emitting diode at the next moment; and They represent the first The node and the first The physical state parameters of each node at the current moment; This represents the transient thermal resistance coefficient, obtained based on test data of the heat dissipation characteristics of the backlight module material. This indicates the rated input power of a single semiconductor light-emitting diode; Used to characterize the proportion of non-luminous dissipation in the conversion of electrical power into thermal power; This represents the set of adjacent nodes whose spatial distance is within a preset thermal influence range; Represents a node With nodes The heat conduction weight between them.

[0096] After calculating the predicted junction temperature, the system determines the thermal safety boundary. When the predicted junction temperature at the next moment exceeds the preset safe junction temperature threshold (set between 85℃ and 105℃, for example, 100℃), it indicates that the physical arrangement of the current area is superimposed on the high-load image-driven content, which is about to trigger a local overheating risk. At this time, the system does not directly turn off the LEDs to avoid causing black spots on the screen, but instead increases the transient thermal risk index output by the corresponding semiconductor LED network according to a preset step size or preset ratio. The specific index update formula is as follows:

[0097] ;

[0098] in, The updated transient thermal risk index; The baseline risk index is the initial output of the network forward inference, and its value is normalized to be between [0,1]. The predicted junction temperature exceeds the limit; For safe junction temperature threshold; The risk surge amplification factor is preset (e.g., preferably 1.5). This updated formula ensures that the thermal risk index of nodes with larger overheating amplitudes increases non-linearly.

[0099] In this embodiment, the system determines the target start phase of each semiconductor light-emitting diode (LED) based on the target driving duty cycle and transient thermal risk index, and generates a timing control signal for each LED that includes the corresponding target start phase and the corresponding target driving duty cycle. By reconstructing the driving timing, adjacent light-emitting units that were originally lit synchronously are staggered on a microsecond-level time axis, thereby cutting off the physical conditions for transient superposition of thermal fields.

[0100] When determining the target starting phase of each semiconductor light-emitting diode (LED), the system first performs a global traversal of the transient thermal risk index of the entire board to determine whether the transient thermal risk index corresponding to each LED exceeds a preset trigger threshold. If the transient thermal risk index of the corresponding LED does not exceed the preset trigger threshold, it indicates that the area is currently within the safe heat dissipation margin, and the system's preset initial phase is directly used as the target starting phase of that LED.

[0101] If the transient thermal risk index of a corresponding LED exceeds a preset trigger threshold, the system initiates a local phase avoidance mechanism. First, the system extracts the LED that triggered the threshold along with adjacent LEDs whose spatial distance is less than or equal to a preset distance threshold, forming an independent target set. Specifically, the preset distance threshold can be defined as the number of physical connectivity hops in the topology. For example, the target set can be a local array consisting of eight LEDs within a 3×3 physical array region radiating outwards from the LED that triggered the threshold. This partitioning method covers the high-risk areas with the most concentrated heat conduction while avoiding computational overload caused by including global nodes in the sorting. Subsequently, within the target set, the system performs a descending sorting operation according to the transient thermal risk index value of each LED, thereby obtaining the ranking of the corresponding LED in the target set.

[0102] After completing the local sorting, the system calculates the starting phase offset of the corresponding semiconductor light-emitting diode based on the sorting rank. The specific calculation formula is as follows:

[0103] ;

[0104] in: Indicates the first The initial phase offset calculated from each semiconductor light-emitting diode. Indicates the first The sorting rank obtained by arranging the semiconductor light-emitting diodes in the target set in descending order, the value of which consists of discrete integers. This represents the preset unit clock phase offset. This offset is preset by the internal phase-locked loop crystal oscillator frequency and minimum addressing clock period of the underlying hardware driver control chip.

[0105] After calculating the offset, the system superimposes the starting phase offset with a preset initial phase to obtain the theoretical starting phase. To avoid the reconstructed driving timing exceeding the clock boundary of a single display cycle, the system performs a summation check on the theoretical starting phase and the conduction duration corresponding to the target driving duty cycle. If the total summation duration exceeds a preset PWM display driving cycle duration, a modulo operation is performed on the theoretical starting phase for that PWM display driving cycle duration, causing the overflow portion of the driving waveform to be folded and reloaded to the beginning of that cycle, thereby finally obtaining the valid target starting phase for the corresponding semiconductor light-emitting diode.

[0106] After completing all spatial dimensionality reduction and temporal rearrangement operations, the system drives the LED backlight array according to the timing control signals. Specifically, the driving process involves the system sending each timing control signal to the driving control chip on the backplane substrate via a high-speed serial interface. Upon receiving the command, the driving control chip precisely controls the turn-on time of the driving channel of the corresponding semiconductor light-emitting diode in the LED backlight array based on the target start phase parsed from each timing control signal. Simultaneously, based on the target driving duty cycle in each timing control signal, it controls the conduction duration of the field-effect transistor inside the driving channel, achieving precise projection of the corresponding luminous intensity.

[0107] After the driving instructions for the current display cycle are executed, the system immediately enters the display preparation stage of the next frame and returns to the step of obtaining the physical state parameters of each semiconductor light-emitting diode in the LED backlight array, thereby forming a continuous closed-loop control data stream to ensure the long-term consistency and thermal safety of the backlight display state.

[0108] This embodiment also provides an LED backlight brightness uniformity optimization system based on AI algorithms. The system includes a data acquisition module, a topology graph construction module, an inference module, a signal generation module, and a driving module. The data acquisition module can be implemented by a microcontroller circuit equipped with an analog-to-digital converter, used to acquire the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array. This module interacts directly with the underlying display driver hardware, collecting electrical parameters within the vertical blanking interval that does not interfere with normal display, and through conversion mapping, digitizing the physical state data of the underlying hardware and transmitting it to the upper-layer business logic operation domain.

[0109] The topology graph construction module communicates with the data acquisition module to construct a physical topology graph of the backlight based on the spatial distribution of each semiconductor light-emitting diode and the transmitted physical state parameters and target brightness. This module is responsible for performing the structured transformation of the data, converting discrete observation parameters into a spatial graph data structure that includes initial feature vectors of nodes and accompanying optical diffusion weights and thermal conduction weights, thus establishing the mathematical boundaries of the internal physical effects of the backlight.

[0110] The inference module receives graph domain data output by the topology graph construction module. Specifically, it can be deployed in the display main control SoC chip with built-in tensor processing unit. It is used to input the backlight physical topology graph into a preset physical law embedded graph neural network for network inference. In the inference calculation pipeline, the module performs physical constraint calculation and error judgment on the prediction results output by the forward propagation of the network according to the preset optical superposition equation and thermodynamic diffusion equation, thereby outputting the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode after iterative optimization calibration of physical laws.

[0111] The signal generation module is connected to the output of the inference module. It is used to determine the target start phase of each semiconductor light-emitting diode based on the target driving duty cycle and transient thermal risk index corresponding to each semiconductor light-emitting diode. This module resolves the thermal risk concentrated in the spatial domain into a microsecond time domain opening misalignment by performing local safety warning, target set descending sorting and phase offset superposition calculation, and generates timing control signals for each semiconductor light-emitting diode, which contain the corresponding target start phase and target driving duty cycle.

[0112] The driving module is connected to the underlying multi-channel constant current driving IC through a serial peripheral interface bus. It is used to drive the LED backlight array according to the timing control signals issued by the signal generation module. As the end of the closed-loop link, this module directly communicates with the hardware driver chip of the display system through the bus interface. It parses and converts the digitized multi-dimensional timing control signals into physical level actions that control the switching state of the field-effect transistors of each independent driving channel.

[0113] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements all the steps in the aforementioned embodiment of the LED backlight brightness uniformity optimization method based on AI algorithms. The computer-readable storage medium can be a non-volatile storage medium, such as a read-only memory, magnetic disk, or optical disk, or it can be a volatile random access memory.

[0114] In practice, the hardware entity executing the aforementioned computer program is an electronic device containing a processor and memory, such as a high-end display main control board or a dedicated image processing system-on-a-chip. The memory internally stores the weight matrix of the physical laws embedded in the graph neural network, historical experience mapping tables of various physical parameters, and the program instruction set for executing each control step. The processor reads instructions from the memory and executes calculations through the system's high-speed front-side bus, coordinating the underlying sample-and-hold circuitry and serial communication interface in real time.

Claims

1. A method for optimizing the brightness uniformity of LED backlights based on AI algorithms, characterized in that, Includes the following steps: Obtain the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array; Based on the spatial distribution of each semiconductor light-emitting diode, the physical state parameters, and the target brightness, a physical topology diagram of the backlight is constructed. The physical topology map of the backlight is input into a preset physical law embedded graph neural network for network inference, and the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode are output. In the network inference process, the physical law embedded graph neural network combines the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculation on the prediction results output by the network, and iteratively adjusts the target driving duty cycle based on the calculated brightness prediction error, and updates the transient thermal risk index based on the junction temperature prediction value. Based on the target driving duty cycle and transient thermal risk index corresponding to each semiconductor light-emitting diode, a descending sorting and phase offset are performed on the local area where the transient thermal risk index exceeds a preset threshold, the target starting phase of each semiconductor light-emitting diode is determined, and a timing control signal containing the corresponding target starting phase and the corresponding target driving duty cycle is generated for each semiconductor light-emitting diode. The LED backlight array is driven according to the timing control signals described above.

2. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 1, characterized in that, The acquisition of the physical state parameters of each semiconductor light-emitting diode in the LED backlight array includes: In the vertical blanking range of the display system to which the LED backlight array belongs, the output of positive drive current to each of the semiconductor light-emitting diodes is paused; A reverse bias voltage is applied to each semiconductor light-emitting diode in the LED backlight array, and the corresponding reverse junction capacitance and reverse leakage current are collected. Based on a preset mapping function, the junction temperature and aging degradation factor of each semiconductor light-emitting diode at the current moment are calculated, and the junction temperature and aging degradation factor are used as the physical state parameters.

3. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 1, characterized in that, The construction of a backlight physical topology map based on the spatial distribution of each of the semiconductor light-emitting diodes, the physical state parameters, and the target brightness includes: Each semiconductor light-emitting diode in the LED backlight array is taken as a node in the physical topology diagram of the backlight, and the target brightness and physical state parameters corresponding to each semiconductor light-emitting diode are taken as the initial feature vector of each node. The edges between the nodes are constructed based on the physical distance between each of the semiconductor light-emitting diodes, and optical diffusion weights and thermal conduction weights are assigned to the edges; wherein, the optical diffusion weights are used to characterize the spatial crosstalk rate of light between adjacent semiconductor light-emitting diodes, and the thermal conduction weights are used to characterize the reciprocal of the thermal flow impedance between adjacent semiconductor light-emitting diodes.

4. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 3, characterized in that, The physical law embedded graph neural network includes a multi-layer network structure; in the single-layer feature update process of the physical law embedded graph neural network, based on the optical diffusion weight and thermal conduction weight corresponding to each edge, the previous layer feature vector of the current node is aggregated with the previous layer feature vector of the adjacent nodes in the backlight physical topology graph that are connected to the current node by an edge to obtain the current layer feature vector of the current node.

5. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 3, characterized in that, The process of combining the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculations on the prediction results of the network output, iteratively adjusting the target driving duty cycle based on the calculated brightness prediction error, and updating the transient thermal risk index based on the junction temperature prediction value includes: During the network inference process, the predicted combined brightness of each semiconductor light-emitting diode and the predicted junction temperature at the next moment are calculated. The synthesized brightness prediction value is calculated according to the preset optical superposition equation, wherein the variables of the optical superposition equation include the optical diffusion weight, the target driving duty cycle, and the electro-optical conversion efficiency determined by the physical state parameters. When the absolute value of the difference between the synthesized brightness prediction value and the target brightness is greater than a preset brightness error threshold, the difference is taken as the brightness prediction error, and a preset numerical optimization algorithm is used to iteratively adjust the target drive duty cycle output by the network until the absolute value of the difference is less than or equal to the preset brightness error threshold; the junction temperature prediction value at the next moment is calculated according to the preset thermodynamic diffusion equation, wherein the variables of the thermodynamic diffusion equation include the heat conduction weight, the physical state parameters, and the target drive duty cycle; When the predicted junction temperature value at the next moment exceeds the preset safe junction temperature threshold, the transient thermal risk index output by the corresponding semiconductor light-emitting diode network is increased by a preset step size or a preset ratio to complete the update.

6. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 1, characterized in that, The step of performing descending sorting and assigning phase offsets to local regions where the transient thermal risk index exceeds a preset threshold, and determining the target starting phase of each semiconductor light-emitting diode, includes: Determine whether the transient thermal risk index corresponding to each of the semiconductor light-emitting diodes exceeds the trigger threshold, which is the preset threshold; If so, the corresponding semiconductor light-emitting diode and its adjacent semiconductor light-emitting diodes with a spatial distance less than or equal to a preset distance threshold are combined to form a target set, which is the local region; within the target set, each semiconductor light-emitting diode is sorted in descending order according to its transient thermal risk index to obtain the ranking of the corresponding semiconductor light-emitting diode in the target set; Based on the sorting order and the preset unit clock phase offset, the starting phase offset of the corresponding semiconductor light-emitting diode is calculated. The target starting phase of the corresponding semiconductor light-emitting diode is obtained by superimposing the starting phase offset with the preset initial phase.

7. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 6, characterized in that, Also includes: If the transient thermal risk index of the corresponding semiconductor light-emitting diode does not exceed the trigger threshold, then the preset initial phase is directly used as the target starting phase of the semiconductor light-emitting diode.

8. The method for optimizing the brightness uniformity of LED backlights based on AI algorithms according to claim 1, characterized in that, The step of driving the LED backlight array according to each of the timing control signals includes: Each of the timing control signals is sent to the drive control chip, and the turn-on time of the corresponding semiconductor light-emitting diode in the LED backlight array is controlled based on the target start phase in each of the timing control signals, and the conduction time of the corresponding semiconductor light-emitting diode is controlled according to the target drive duty cycle in each of the timing control signals. After the current display cycle ends, return to the step of obtaining the physical state parameters of each semiconductor light-emitting diode in the LED backlight array.

9. A system for optimizing the brightness uniformity of an LED backlight based on an AI algorithm, characterized in that, include: The data acquisition module is used to acquire the physical state parameters and target brightness of each semiconductor light-emitting diode in the LED backlight array; The topology construction module is used to construct a backlight physical topology based on the spatial distribution of each of the semiconductor light-emitting diodes, as well as the physical state parameters and target brightness. The inference module is used to input the physical topology map of the backlight into a preset physical law embedded graph neural network for network inference, and output the target driving duty cycle and transient thermal risk index of each semiconductor light-emitting diode; wherein, during the network inference process, the physical law embedded graph neural network combines the preset optical superposition equation and thermodynamic diffusion equation to perform physical constraint calculation on the prediction results output by the network, and iteratively adjusts the target driving duty cycle based on the calculated brightness prediction error, and updates the transient thermal risk index based on the junction temperature prediction value; The signal generation module is used to perform descending sorting and allocate phase offset for local areas where the transient thermal risk index exceeds a preset threshold based on the target driving duty cycle and transient thermal risk index corresponding to each semiconductor light-emitting diode, determine the target starting phase of each semiconductor light-emitting diode, and generate a timing control signal for each semiconductor light-emitting diode that includes the corresponding target starting phase and the corresponding target driving duty cycle. A driving module is used to drive the LED backlight array according to the timing control signals.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the LED backlight brightness uniformity optimization method based on the AI ​​algorithm as described in any one of claims 1 to 8.

Images