Display screen dead pixel fault prediction detection method and system

Abstract

The invention discloses a display screen dead pixel fault prediction detection method and system. The method comprises the following steps: collecting a visible light image, a short wave infrared image, an ultraviolet fluorescence image, a driving current time sequence signal and surface temperature distribution data of a display screen; performing dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set; inputting the preprocessed multi-modal data set into a space-time fusion network, extracting space-time correlation characteristics in a dead pixel diffusion process through the space-time fusion network, and generating a pixel-level dead pixel mask of a current frame and a diffusion probability thermodynamic diagram of a future time period based on the space-time correlation characteristics; and according to the dead pixel mask and the diffusion probability thermodynamic diagram, dynamically adjusting detection parameters to adapt to target equipment, and outputting a probabilistic diagnosis result of dead pixel causes in combination with physical relevance of multi-modal data.

Description

Technical Field

[0001] The present invention belongs to the technical field of display screen manufacturing, and relates to a method and system for predicting and detecting display screen dead pixel faults. Background Art

[0002] During the manufacturing and use of display screens, the accurate detection and cause diagnosis of dead pixel faults are the core links to ensure display quality. Traditional detection methods usually rely on single-modal data for static analysis, and it is difficult to comprehensively capture the diffusion law of dead pixels over time and their dynamic correlation with circuit and material aging. In the prior art, single-modal based detection algorithms are vulnerable to environmental noise interference and lack the ability to synergistically analyze multi-source heterogeneous data, resulting in high false detection rates and missed detection rates under complex working conditions. In addition, most methods do not consider the influence of degradation noise during the long-term use of display screens, and the model robustness is insufficient, making it difficult to simulate the evolution characteristics of dead pixels in real aging scenarios. In terms of cross-modal data processing, existing solutions often ignore the accuracy requirements of spatio-temporal alignment, and the time synchronization deviation and spatial registration error of multi-sensor data are likely to cause feature fusion distortion, further reducing the detection reliability. At the same time, traditional methods have limited ability to predict the diffusion trend of dead pixels, cannot generate a pixel-level future diffusion probability heat map, and the diagnosis results are mostly based on a single threshold determination, lacking physical correlation analysis and probabilistic evaluation of multiple causes, and it is difficult to meet the real-time detection and maintenance decision-making requirements of high-resolution display screens. Summary of the Invention

[0003] The present invention aims to provide a method and system for predicting and detecting display screen dead pixel faults, which improve the detection robustness through multi-modal data collaborative analysis, dynamic noise injection and cross-modal alignment, extract spatio-temporal correlation features of dead pixel diffusion through a spatio-temporal fusion network to generate a mask and a heat map, and achieve efficient prediction detection and reliable maintenance of display screen dead pixel faults through dynamic parameter adjustment and probabilistic diagnosis.

[0004] To achieve the above object, the present invention adopts the following technical solutions: A method for predicting and detecting display screen dead pixel faults, comprising the following steps: Collect visible light images, short-wave infrared images, ultraviolet fluorescence images, drive current time series signals and surface temperature distribution data of the display screen; Perform dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set; Input the preprocessed multi-modal data set into a spatio-temporal fusion network, extract spatio-temporal correlation features during the dead pixel diffusion process through the spatio-temporal fusion network, and generate a pixel-level dead pixel mask for the current frame and a diffusion probability heat map for a future period based on the spatio-temporal correlation features; According to the dead pixel mask and the diffusion probability heat map, dynamically adjust the detection parameters to adapt to the target device, and combine the physical relevance of the multi-modal data to output a probabilistic diagnosis result of the cause of the dead pixel.

[0005] Further, the dynamic noise injection includes: Applying brightness attenuation noise based on a physical degradation model to the visible light image, and the brightness attenuation equation of the model is ; where is the brightness value of the pixel at coordinates (x, y) at time t, is the initial brightness value of the pixel at coordinates (x, y), is the attenuation coefficient, and the value range is 0.1 to 0.3, is the time constant, and the value range is 100 to 500 hours, is the non-linear attenuation factor, and the value range is 1.2 to 1.8; Superimposing Weibull distribution noise on the drive current timing signal, and its probability density function is ; where x is the current pulse amplitude, is the shape parameter, and the value range is 1.5 to 2.5, is the scale parameter, and the value range is 0.1 to 0.3.

[0006] Further, the cross-modal alignment includes: Time synchronization, taking the screen vertical synchronization signal as the time reference, and aligning the timestamps of the visible light image, the short-wave infrared image, the ultraviolet fluorescence image, the drive current timing signal, and the surface temperature distribution data; Spatial reference point extraction, based on the time-synchronized short-wave infrared image, extracting the circuit trace intersection points as spatial reference points, and the number of the spatial reference points is not less than 4 groups; Spatial registration, based on the extracted spatial reference points, solving the affine transformation matrix by the least squares method, and registering the visible light image and the ultraviolet fluorescence image to the coordinate system of the short-wave infrared image to achieve sub-pixel level spatial alignment.

[0007] Further, inputting the preprocessed multi-modal data set into a spatio-temporal fusion network, extracting spatio-temporal correlation features in the dead pixel diffusion process through the spatio-temporal fusion network, and generating a pixel-level dead pixel mask for the current frame and a diffusion probability heat map for a future period based on the spatio-temporal correlation features includes the following steps: Time evolution feature extraction, based on the preprocessed multi-modal data set, extracting the historical evolution features of the attenuation rate of the dead pixel brightness over time and the diffusion direction through causal time modeling; Spatial distribution analysis: Based on the extracted historical evolution features, locate the spatial distribution area of bad pixels in the current frame, and analyze the morphological correlation and microscopic texture differences between bad pixels and surrounding pixels. Cross-modal feature fusion: Fuse the obtained spatial distribution features with the driving current timing signal and surface temperature distribution data, calculate the correlation weights between optical features and electrical features through the cross-attention mechanism, and generate a pixel-level bad pixel mask and a diffusion probability heat map.

[0008] Furthermore, the weight calculation formula of the cross-attention mechanism is: ; where is the query vector corresponding to the i-th optical feature, which is derived from the feature map of visible light, short-wave infrared, or ultraviolet fluorescence images. is the key vector corresponding to the j-th electrical feature, which is derived from the timing feature vector converted from the driving current timing signal. N is the total number of sampling points of the timing signal, and T is the matrix transpose operator. is the attention weight between the optical feature at the i-th spatial position and the electrical feature at the j-th time point. is the natural exponential function.

[0009] Furthermore, the dynamic adjustment of detection parameters to adapt to the target device includes the following steps: Detection window step size adaptation: Dynamically adjust the moving step size of the detection window according to the pixel density of the target device. Small step sizes are used for high-resolution devices, and large step sizes are used for low-resolution devices. Confidence threshold dynamic setting: Based on the statistical features of the diffusion probability heat map, calculate the dynamic threshold through the mean and standard deviation. Multi-frame consistency verification: Perform a logical AND operation on the bad pixel masks and heat maps of multiple consecutive frames to filter out stable bad pixels that persist in multiple frames and suppress transient noise.

[0010] Furthermore, the probabilistic diagnostic result of the bad pixel cause is output by combining the physical relevance of multi-modal data. Among them, the determination conditions for the probabilistic diagnostic result of the bad pixel cause are: Circuit overload determination: When the amplitude of the driving current pulse exceeds the current determination threshold and the local temperature rise exceeds 5°C above the ambient temperature, a circuit overload alarm is triggered. Physical damage determination: If internal crack features are detected in the short-wave infrared image and the ultraviolet fluorescence intensity is lower than 30% of the normal area, it is determined as physical damage. Material aging determination: When the brightness decay rate exceeds 0.15% per hour and the chromaticity shift exceeds 8, it is marked as material aging.

[0011] Further, after probabilistically diagnosing the cause of dead pixels by combining the physical relevance of multimodal data, the following steps are also included: Combining the multimodal features of the circuit overload determination, the physical damage determination, and the material aging determination into a feature vector f, where the features include the driving current pulse amplitude, local temperature rise data, short-wave infrared crack features, ultraviolet fluorescence intensity, and chromaticity shift; Based on the feature vector f, calculate the posterior probability of each cause type through the following formula: ; where, is the posterior probability that the cause type of the dead pixel is c under the condition of the given multimodal feature combination f, is the cause type of the dead pixel. Under the condition that the known cause type of the dead pixel is c, is the likelihood probability of observing the feature vector f, is the prior probability of the cause type c of the dead pixel, that is, the initial probability distribution of the fault type when there is no any feature information, is a traversal variable of the cause type of the dead pixel, representing the sum of all possible values of c, is the summation operation for all possible cause types c' of the dead pixel, is under the condition that the known cause type of the dead pixel is , the likelihood probability of observing the feature vector , is the prior probability of the cause type of the dead pixel.

[0012] A display screen dead pixel fault detection system for executing the display screen dead pixel fault detection method described above, includes: A data acquisition module, used to synchronously acquire the visible light image, short-wave infrared image, ultraviolet fluorescence image, driving current timing signal, and surface temperature distribution data of the display screen; A preprocessing module, connected to the data acquisition module, used to perform dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multimodal data set; A spatio-temporal fusion network module, connected to the preprocessing module, used to extract the historical evolution features of the decay rate and diffusion direction of the dead pixel brightness over time, locate the spatial distribution area of the dead pixel in the current frame and analyze the morphological relevance between the dead pixel and surrounding pixels, and fuse the spatio-temporal features with the driving current timing signal and surface temperature data to generate a pixel-level dead pixel mask and a diffusion probability heat map; A dynamic adjustment module, connected to the spatio-temporal fusion network module, used to dynamically adjust the detection window step size and confidence threshold according to the pixel density of the target device to adapt to the detection requirements of devices with different resolutions; The cause diagnosis module, connected to the dynamic adjustment module, is used to output a probabilistic diagnosis result of the cause of the dead pixel by combining the physical relevance of multi-modal data.

[0013] Advantages of the present invention: The present invention constructs a comprehensive detection basis by synchronously collecting multi-modal data (visible light, short-wave infrared, ultraviolet images, current signals, and temperature data), enhances the adaptability of the model to real noise scenarios through dynamic noise injection, eliminates the time and space deviations of multi-source data through cross-modal spatio-temporal alignment, and ensures the accuracy of feature fusion; The spatio-temporal fusion network extracts spatio-temporal correlation features of dead pixel diffusion, generates accurate pixel-level masks and heat maps for predicting future diffusion trends, dynamically adjusts the detection window step size and confidence threshold to adapt to different resolution devices, and realizes probabilistic diagnosis by combining the physical relevance between multi-modal data, significantly improving the detection efficiency and accuracy. At the same time, it provides forward-looking maintenance guidance, providing an efficient and reliable full-process solution for display dead pixel detection. Description of the Drawings

[0014] Figure 1 It is a schematic diagram of the method steps of the present invention.

[0015] Figure 2 It is a schematic diagram of the cross-modal alignment steps of the present invention.

[0016] Figure 3 It is a schematic diagram of the steps of generating a pixel-level dead pixel mask and a diffusion probability heat map of the present invention.

[0017] Figure 4 It is a schematic diagram of the steps of dynamically adjusting detection parameters to adapt to the target device of the present invention. Detailed Embodiments

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit the present invention; the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above description of the drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of the present invention or the above drawings are used to distinguish different objects and not to describe a specific order.

[0019] References to "embodiments" in this specification mean that specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the invention. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0020] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0021] The present invention provides an attached Figures 1 to 4 , and the present invention aims to provide a method and system for predicting and detecting display screen dead pixel faults. By means of multi-modal data collaborative analysis, dynamic noise injection, and cross-modal alignment, the detection robustness is improved. The spatio-temporal correlation features of dead pixel diffusion are extracted by combining with a spatio-temporal fusion network to generate a mask and a heat map, and through dynamic parameter adjustment and probabilistic diagnosis, efficient prediction and detection of display screen dead pixel faults and reliable maintenance are achieved. Specific Embodiment 1 A method for predicting and detecting display screen dead pixel faults includes the following steps: S1. Collect visible light images, short-wave infrared images, ultraviolet fluorescence images, drive current timing signals, and surface temperature distribution data of the display screen; Specifically, visible light image data is used to collect visible light images of the display screen through a visible light camera to capture the brightness abnormal areas of dead pixels; Short-wave infrared image data is used to obtain short-wave infrared images of the internal circuit traces and crack features of the screen through a short-wave infrared camera; Ultraviolet fluorescence image data is used to capture ultraviolet fluorescence images of screen materials through an ultraviolet fluorescence imager to identify fluorescence attenuation caused by material aging; Drive current timing signal data is used to synchronously record drive current timing signals through a high-precision current sensor to monitor circuit overload pulses; Surface temperature distribution data is used to collect surface temperature distribution data of the screen through an infrared thermal imager to locate local temperature rise areas.

[0023] S2. Perform dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set; Specifically, dynamic noise injection includes applying noise that simulates long-term aging of the visible light image to enhance data robustness, and superimposing random noise on the drive current signal to simulate the real circuit interference scenario; cross-modal alignment includes using the vertical synchronization signal of the screen as the time reference to control the timestamp deviation of all sensor data within ±1 ms, and achieving pixel-level alignment of visible light, ultraviolet, and short-wave infrared images through spatial transformation based on the circuit trace features of the short-wave infrared image.

[0024] S3. Input the preprocessed multi-modal data set into the spatio-temporal fusion network, extract the spatio-temporal correlation features during the bad pixel diffusion process through the spatio-temporal fusion network, and generate a pixel-level bad pixel mask for the current frame and a diffusion probability heat map for the future time period based on the spatio-temporal correlation features. Specifically, input the preprocessed multi-modal data set into the spatio-temporal fusion network, where the network architecture of the spatio-temporal fusion network includes a time evolution branch, a spatial distribution branch, and a cross-modal fusion layer. The time evolution branch extracts the decay characteristics of the bad pixel brightness over time (such as the brightness change rate of 10-frame historical data) through a 3D convolutional layer. The spatial distribution branch analyzes the morphological correlation between bad pixels and surrounding pixels in the current frame (such as edge sharpness, texture consistency). The cross-modal fusion layer then performs weighted fusion of the time evolution features with current and temperature data to generate two types of outputs, as follows: a. Pixel-level bad pixel mask: Mark the exact positions of all suspicious bad pixels in the current frame. b. Diffusion probability heat map: Predict the probability distribution of bad pixel diffusion to adjacent regions within the next 24 hours.

[0025] S4. According to the bad pixel mask and the diffusion probability heat map, dynamically adjust the detection parameters to adapt to the target device, and combine the physical correlation of multi-modal data to output a probabilistic diagnosis result of the cause of bad pixels.

[0026] Specifically, generate a bad pixel mask through the spatio-temporal fusion network. This mask is a binary matrix that marks the exact positions of all confirmed bad pixels in the current frame. At the same time, obtain the diffusion probability heat map, which is a probability matrix that predicts the probability distribution of each pixel region developing into a bad pixel within the next 24 hours. Furthermore, based on the bad pixel mask, statistically calculate the density of the current bad pixel region, and dynamically calculate the detection window step size in combination with the proportion of high-risk regions in the diffusion heat map. For example, when the bad pixel density is 0.3% and the proportion of high-risk diffusion areas is 2%, adjust the detection step size from the baseline of 16 pixels to 12.8 pixels to balance detection accuracy and efficiency. Furthermore, multi-modal physical parameters such as the amplitude of the extracted current pulse, local temperature rise data, short-wave infrared crack characteristics, ultraviolet fluorescence intensity attenuation, brightness attenuation rate, and chromaticity shift are extracted. The initial probability weights of each cause are calculated through weighted association rules. Among them, the circuit overload weight is determined by the product of the current and temperature rise data, the physical damage weight is generated based on the crack length and fluorescence attenuation ratio, and the material aging weight is associated with the brightness attenuation rate and chromaticity shift value. Finally, it is normalized and output as a probability distribution. The example result is 72% for circuit overload, 22% for physical damage, and 6% for material aging.

[0027] This application enhances data robustness through dynamic noise injection and cross-modal spatio-temporal alignment, combines a spatio-temporal fusion network to extract the historical evolution characteristics of the brightness attenuation and diffusion direction of dead pixels, generates a high-precision pixel-level dead pixel mask and a future diffusion probability heat map, dynamically adjusts the detection window step size and confidence threshold to adapt to different resolution devices, and at the same time realizes probabilistic diagnosis of circuit overload, physical damage, and material aging based on the correlation of multi-modal physical parameters such as current, temperature rise, and cracks, significantly reducing the false detection rate and missed detection rate, improving the detection efficiency and the accuracy of cause judgment, and providing reliable prediction and decision support for display screen maintenance. Specific Embodiment 2 The dynamic noise injection includes: Applying brightness attenuation noise based on a physical degradation model to the visible light image, and the brightness attenuation equation of the model is ; where is the brightness value of the pixel at coordinates (x, y) at time t, is the initial brightness value of the pixel at coordinates (x, y), is the attenuation coefficient, with a value range of 0.1 to 0.3, is the time constant, with a value range of 100 to 500 hours, is the non-linear attenuation factor, with a value range of 1.2 to 1.8; Specifically, in this embodiment, the parameters are set as k = 0.2 (attenuation intensity), τ = 300 hours (time constant), = 1.5 (non-linear factor), and the initial pixel brightness is extracted from the original visible light image; The specific implementation steps are as follows: 1. Simulate the aging process of the display screen running continuously for 300 hours, and segment it according to the time step t = 24 hours; 2. Calculate the brightness attenuation value of each pixel every 24 hours according to the formula and superimpose it on the original image; 3. Generate a visible light image dataset containing brightness attenuation noise to simulate the dead pixel diffusion characteristics after long-term use.

[0029] Add Weibull distribution noise to the drive current timing signal, and its probability density function is ; where x is the current pulse amplitude, is the shape parameter, with a value range of 1.5 to 2.5, is the scale parameter, with a value range of 0.1 to 0.3.

[0030] Specifically, in this embodiment, the parameters are set as = 2.0 (shape parameter, simulating the skewness of the pulse amplitude distribution), (scale parameter, controlling the noise intensity), and the sampling range of the current pulse amplitude x is 0.1 mA to 20 mA; The specific implementation steps are as follows: 1. Randomly sample time points in the original drive current signal; 2. Generate noise amplitudes according to the Weibull distribution and superimpose them on the current pulses at the sampling points; 3. Generate current timing data with random pulse interference to simulate abnormal fluctuations caused by circuit aging.

[0031] In this embodiment, brightness attenuation noise is injected to simulate the bad point brightness attenuation law after long-term use of the display screen, so that the training data covers the real aging scenario and improves the detection robustness of the model to progressive faults; Inject Weibull distribution noise to generate random pulses that conform to the statistical characteristics of circuit aging, and enhance the generalization ability of the model to current abnormal signals. Specific Embodiment Three The cross-modal alignment includes: S221. Time synchronization: Use the screen vertical synchronization signal as the time reference to align the timestamps of the visible light image, the short-wave infrared image, the ultraviolet fluorescence image, the drive current timing signal, and the surface temperature distribution data; S222. Spatial reference point extraction: Based on the time-synchronized short-wave infrared image, extract the circuit trace intersection points as spatial reference points, and the number of spatial reference points is not less than 4 groups; S223. Spatial registration: Based on the extracted spatial reference points, solve the affine transformation matrix by the least squares method, and register the visible light image and the ultraviolet fluorescence image to the coordinate system of the short-wave infrared image to achieve sub-pixel level spatial alignment.

[0033] Specifically, the specific steps of cross-membrane alignment in this embodiment are as follows: 1. Time synchronization: Input data: visible light image data, short-wave infrared image data, ultraviolet fluorescence image data, drive current timing signal data, surface temperature distribution data; Using the screen vertical synchronization signal (VSync) as the unified time reference, add time stamps accurate to the millisecond level to each frame of image and sensor data; Align low-frequency data (such as temperature data at 10Hz) to the high-frequency time axis (such as visible light images at 60fps) through an interpolation algorithm, with the time deviation controlled within ±1ms.

[0034] 2. Spatial reference point extraction: Input the short-wave infrared image data after time synchronization; Detect the intersection points of circuit traces, and use the Harris corner detection algorithm to extract high-contrast corner points at the trace intersections; among them, the screening conditions are that the distance between adjacent corner points ≥ 50 pixels and the corner response value > 0.8 (normalized intensity threshold); After screening, retain 6 groups of spatial reference points with a coordinate accuracy of ±0.5 pixels; 3. Spatial registration: Input visible light image data, ultraviolet fluorescence image data, short-wave infrared image data (reference coordinate system); Extract the reference point coordinates and obtain the pixel coordinates of 6 groups of reference points from the short-wave infrared image; Feature matching to locate the corresponding reference point positions in the visible light and ultraviolet images; Affine transformation matrix calculation: ; Solve the matrix parameters by the least squares method, and the objective function is to minimize the sum of the squares of the reference point coordinate residuals; Example matrix parameters, , , pixels, , , pixels; among them, 、 are the pixel coordinates of a certain point in the original image, 、 are the pixel coordinates of the corresponding point in the transformed image, is the scaling factor in the x-axis direction, is the shear factor of the y-axis with respect to the x-axis, is the translation amount in the x-axis direction, is the shear factor of the x-axis with respect to the y-axis, is the scaling factor in the y-axis direction, is the translation amount in the y-axis direction, and the homogeneous coordinate "1" is the fixed expansion term for homogeneous coordinate transformation, used to be compatible with matrix operations for translation transformation.

[0035] Image transformation: Apply the above affine matrix to visible light and ultraviolet images and register them to the short-wave infrared image coordinate system, with a registration error < 0.3 pixels (sub-pixel accuracy).

[0036] In this embodiment, time synchronization eliminates the time offset of multi-sensor data, ensuring the strict alignment of the timing signal and the image frame, and avoiding feature fusion distortion caused by time misalignment; spatial registration achieves precise spatial alignment of visible light, ultraviolet, and short-wave infrared images (error < 0.3 pixels) through sub-pixel affine transformation, providing geometric consistency guarantee for cross-modal feature fusion; fiducial point screening is based on the stable physical features of circuit trace intersections, ensuring the repeatability and robustness of fiducial points in different modal images. Specific Embodiment 4 Input the preprocessed multi-modal data set into the spatio-temporal fusion network, extract spatio-temporal correlation features during the bad pixel diffusion process through the spatio-temporal fusion network, and generate a pixel-level bad pixel mask for the current frame and a diffusion probability heat map for future time periods, including the following steps: S31. Extract time-evolution features: Based on the preprocessed multi-modal data set, extract the decay rate of the bad pixel brightness over time and the historical evolution features of the diffusion direction through causal time modeling. S32. Analyze spatial distribution: According to the extracted historical evolution features, locate the spatial distribution area of bad pixels in the current frame, and analyze the morphological correlation and micro-texture differences between bad pixels and surrounding pixels. S33. Cross-modal feature fusion: Fuse the obtained spatial distribution features with the driving current timing signal and surface temperature distribution data, calculate the correlation weights between optical features and electrical features through a cross-attention mechanism, and generate a pixel-level bad pixel mask and a diffusion probability heat map.

[0038] Specifically, the specific steps of this embodiment are as follows: S311. Input the preprocessed multi-modal data set (visible light, short-wave infrared, ultraviolet image sequences, driving current timing signal, surface temperature data), with a time window of 10 consecutive frames and a time interval of 2 seconds / frame. S312. Use a 3D causal convolution layer (kernel size 3×3×3) to slide along the time dimension and extract the bad pixel brightness decay rate and diffusion direction features. The time convolution operation is: ; where is the time-evolution feature value extracted at position (x, y) and time t, is the weight parameter of the 3D convolution kernel at spatial offset (i, j) and time offset k, It is the pixel value of the input multi-modal data at the position (x + i, y + j) and time t - k. Here, i and j are the offsets in the spatial dimension, with a value range of -1, 0, 1, and k is the offset in the time dimension, with a value range of 0, 1, 2. They are the spatial coordinates (x, y) and the time point t of the current feature map. The output feature is the historical evolution feature matrix (with dimensions H×W×C, where C = 64 channels), which includes the brightness attenuation rate (unit: % / hour) of each pixel point and the diffusion direction vector (angle 0°~360°). S321. Input the time evolution feature matrix (H×W×64). S322. Locate the bad pixel area through the spatial attention mechanism and calculate the morphological correlation between each pixel and its neighborhood (3×3 window). In this step, the morphological difference degree is calculated as: ; Among them, It is the morphological difference degree at the position (x, y), which measures the texture consistency between this point and its neighboring pixels. N is the total number of neighboring pixels, which is 3×3 = 9 here. It is the feature vector at the position (x, y) of the spatial distribution feature map. It is the feature vector of the neighboring pixel at the position (x + i, y + j). It is the L2 norm, which calculates the Euclidean distance between two feature vectors. Here, i and j are the spatial offsets, with a value range of -1, 0, 1, and x and y are the spatial coordinates of the current feature map. S323. Screen the candidate areas of bad pixels. Pixels with a morphological difference degree S(x, y)>0.8 and temperature data T(x, y)>45°C are marked as suspected bad pixels. S324. Output the result, the spatial distribution feature map (H×W×32), which includes the positions of bad pixels and the texture correlation with surrounding pixels. S331. Input the spatial distribution feature map (H×W×32), the driving current timing signal (1×1000-dimensional timing feature), and the surface temperature distribution data (H×W×1). S332. Encode the current timing signal into a timing context vector (32-dimensional) through LSTM. S333. Piece together the temperature data and the spatial distribution feature map pixel by pixel to generate an enhanced spatial feature of H×W×33. S334. Calculate the correlation weight between the optical feature (spatial distribution) and the electrical feature (timing context) through the cross-attention mechanism: Query vector It comes from the i-th position of the spatial feature map. Key vector from the j-th time point of the temporal context vector; S335. Two types of results are output after weighted fusion: Pixel-level bad pixel mask: a binary matrix (H×W×1), with a threshold of 0.5; Diffusion probability heat map: a probability matrix (H×W×1), with a value range of 0 to 1, predicting the diffusion probability for the next 24 hours.

[0039] In this embodiment, the time evolution modeling in this embodiment captures the decay trend of the bad pixel brightness through causal 3D convolution to improve the prediction accuracy of the diffusion direction. The spatial positioning accuracy is jointly screened by the morphological difference degree and the temperature threshold. The bad pixel positioning error is <2 pixels. The cross-modal fusion enhances the spatial features through the current time series and temperature data, reducing the overloading misjudgment rate. The coincidence degree between the heat map prediction and the actual maintenance record can be greater than 90%. Specific Embodiment Five The weight calculation formula of the cross-attention mechanism is: ; where is the query vector corresponding to the i-th optical feature, which is derived from the feature map of visible light, short-wave infrared, or ultraviolet fluorescence images, is the key vector corresponding to the j-th electrical feature, which is derived from the temporal feature vector converted from the driving current time series signal. N is the total number of sampling points of the time series signal, and T is the matrix transpose operator, is the attention weight between the optical feature at the i-th spatial position and the electrical feature at the j-th time point, is the natural exponential function.

[0041] Specifically, this embodiment realizes the calculation of the correlation weight between the optical feature and the electrical feature through the cross-attention mechanism. The specific calculation steps are as follows: 1. Input the optical feature map and the electrical time series feature. The optical feature map comes from the feature maps of visible light, short-wave infrared, and ultraviolet fluorescence images, with a size of H×W×64 (height×width×number of channels). The feature vector at each spatial position i is , and the electrical time series feature is the temporal context vector encoded by the LSTM of the driving current time series signal. The total number of sampling points N = 1000, and the feature vector at each time point j is ; 2. Calculate the attention weight. The calculation formula is ; 3. Input the optical feature map and the electrical time series feature ; 4. For each spatial position i, calculate its attention weight with all time points j , generating a weight matrix ; 5. Weight the electro - time - sequence features by weights and sum them up to obtain a spatio - temporal context vector; ; 6. Generate a fused feature by combining the context vector with the optical features ; 7. Map the fused feature to the output dimension through a fully - connected layer to generate a bad - pixel mask and a diffusion - probability heat map. Specific Embodiment Six The dynamic adjustment of detection parameters to adapt to the target device includes the following steps: S41. Adapt the detection window step size. Dynamically adjust the moving step size of the detection window according to the pixel density of the target device according to the formula. High - resolution devices use a smaller step size to improve accuracy, and low - resolution devices use a larger step size to speed up; Specifically, the specific calculation embodiments of this step are as follows: 1. Input the pixel density (PPI) of the target device, such as an 8K screen (PPI = 450), a 4K screen (PPI = 300); 2. Adjust the step size The adjustment formula is: ; where the reference device PPI = 300, that is , the reference step size , The unit of is pixels, S is the adjusted moving step size of the detection window, is the actual pixel density of the target device; The calculation example is as follows: 8K screen (PPI = 450): ; where the unit of S is pixels, and the value result is rounded down; 2K screen (PPI = 150): ; High - resolution devices use a smaller step size to improve accuracy, and low - resolution devices use a larger step size to speed up;; S42. Dynamically set the confidence threshold. Based on the statistical features of the diffusion - probability heat map, calculate the dynamic threshold through the mean and standard deviation; Specifically, the specific calculation embodiments of this step are as follows: 1. Input the diffusion - probability heat map (matrix size H×W, value range 0 to 1); 2. Calculate the mean μ and standard deviation σ of the probabilities of all pixels in the heat map (for example, μ = 0.3, σ = 0.15); 3. Calculate the dynamic threshold, and the calculation formula is: ; where, is the confidence determination threshold after dynamic calculation, and k is the adjustment coefficient, which is dynamically set according to the signal-to-noise ratio (SNR) of the device; The parameter settings are as follows: For devices with high SNR (SNR > 30dB): k = 1.5, threshold = 0.3 + 1.5 × 0.15 = 0.525; For devices with low SNR (SNR < 15dB): k = 2.5, threshold = 0.3 + 2.5 × 0.15 = 0.675.

[0043] 4. Bad pixel determination. The bad pixel determination conditions are as follows: Probability ≥ threshold: Marked as a confirmed bad pixel; When threshold > probability ≥ threshold - 0.2: Marked as a suspected area, which requires manual review.

[0044] S43. Multi-frame consistency verification. Perform a logical AND operation on the bad pixel masks and heatmaps of multiple consecutive frames to filter out stable bad pixels that persist in multiple frames and suppress transient noise.

[0045] Specifically, the specific calculation implementation example of this step is as follows: 1. Input the bad pixel masks (binary matrices) and heatmaps (probability matrices) of 5 consecutive frames; 2. Perform a logical AND operation on each pixel of the 5 masks, and retain the pixels that are all 1 in the 5 frames; 3. Statistically analyze the pixel areas in the heatmap where the probability > 0.7 persists for 5 frames; 4. Output results: Stable bad pixels: Pixels that pass both the mask logical AND and heatmap persistence verification; Transient noise: Suspected areas that appear only in single frames or non-consecutive frames.

[0046] Specifically, the step size adjustment in this embodiment depends on device parameters. The PPI determines the detection step size to ensure the adaptability of the resolution and detection efficiency; the threshold dynamically responds to the data distribution, and the statistical features of the heatmap drive the threshold adjustment in real time to balance false detection and missed detection; multi-frame verification eliminates transient interference, and verification in the time dimension improves the reliability of the results and suppresses the influence of single-frame noise. Specific Embodiment Seven The probability-based diagnostic result of the bad pixel cause is output by combining the physical relevance of multi-modal data. Among them, the determination conditions for the probability-based diagnostic result of the bad pixel cause are as follows: Circuit overload determination. When the amplitude of the drive current pulse exceeds the current determination threshold and the local temperature rise exceeds the ambient temperature by 5°C, a circuit overload alarm is triggered; Specifically, the circuit overload determination conditions in this embodiment are as follows: ; Among them, is the amplitude of the driving current pulse, is the current determination threshold (which can be set to 10 mA), is the local temperature rise data (relative to the ambient temperature), is the temperature rise determination threshold (which can be set to 5 °C); Example: When the current is 12 mA (≥10 mA) and the temperature rise is 6 °C (≥5 °C), it is determined that the circuit is overloaded.

[0048] For physical damage determination, if internal crack features are detected in the short-wave infrared image and the ultraviolet fluorescence intensity is 30% lower than that of the normal area, it is determined as physical damage; Specifically, the physical damage determination condition in this embodiment is: ; wherein, is the length of the crack feature in the short-wave infrared image, is the crack length threshold (which can be set to 3 px), is the attenuation ratio of the ultraviolet fluorescence intensity, is the fluorescence attenuation threshold (which can be set to 30%); Example: When the crack length is 5 pixels (≥3 pixels) and the fluorescence attenuation is 40% (≥30%), it is determined as physical damage.

[0049] For material aging determination, when the brightness attenuation rate exceeds 0.15% per hour and the chromaticity shift exceeds 8, it is marked as material aging.

[0050] Specifically, the material aging determination condition in this embodiment is: ; wherein, is the brightness attenuation rate, is the brightness attenuation threshold (which can be set to 0.15% / h), is the CIELab chromaticity shift amount, is the chromaticity shift threshold (which can be set to 8); Example: When the brightness attenuation is 0.2% per hour (≥0.15%) and ΔE = 9 (≥8), it is determined as material aging. Specific Embodiment VIII After outputting the probabilistic diagnosis result of the bad pixel cause by combining the physical relevance of the multi-modal data, the following steps are further included: S5. Combine the multi-modal features of the circuit overload determination, the physical damage determination, and the material aging determination into a feature vector f, and the features include the amplitude of the driving current pulse, the local temperature rise data, the short-wave infrared crack feature, the ultraviolet fluorescence intensity, and the chromaticity shift; Specifically, the feature vector is constructed as ; S6. Calculate the posterior probability of each cause type based on the feature vector f through the following formula: ; where, is the posterior probability that the cause type of the bad point is c under the condition of the given multi-modal feature combination f, is the cause type of the bad point. Under the condition that the cause type of the bad point is known to be c, is the likelihood probability of observing the feature vector f, is the prior probability of the cause type c of the bad point, that is, the initial probability distribution of the fault type when there is no feature information, is the traversal variable of the cause type of the bad point, indicating the sum of all possible values of c, is the summation operation for all possible cause types c' of the bad point, is under the condition that the cause type of the bad point is known to be the likelihood probability of observing the feature vector is the cause type of the bad point the prior probability of.

[0052] More specifically, the calculation process example is as follows: 1. Prior probability, the initial value is set as P(circuit overload) = 0.5, P(physical damage) = 0.3, P(material aging) = 0.2; Likelihood probability, obtained by statistical analysis of historical maintenance data, the example is as follows: Circuit overload category: Among 1000 historical data, the frequency of the feature f appearing is 0.08; Physical damage category: Among 800 historical data, the frequency of the feature f appearing is 0.12; Material aging category: Among 500 historical data, the frequency of the feature f appearing is 0.02.

[0053] 2. Calculate the numerator term: P(f∣circuit overload)P(circuit overload)=0.08×0.5 = 0.04; P(f∣physical damage)P(physical damage)=0.12×0.3 = 0.036; P(f∣material aging)P(material aging)=0.02×0.2 = 0.004; 3. Calculate the denominator term: ; 4. Calculate the posterior probability: , , . ​

[0054] Specifically, the posterior probability directly reflects the contribution degree of each cause. For example, the physical damage probability of 45% guides the priority of repairing the crack area. Specific Embodiment Nine A display screen dead pixel fault detection system for implementing the display screen dead pixel fault prediction and detection method as described above, comprising: A data acquisition module for synchronously acquiring visible light images, short-wave infrared images, ultraviolet fluorescence images, driving current timing signals, and surface temperature distribution data of the display screen; A preprocessing module connected to the data acquisition module for performing dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set; A spatio-temporal fusion network module connected to the preprocessing module for extracting historical evolution features of the decay rate and diffusion direction of the dead pixel brightness over time, locating the spatial distribution area of the dead pixel in the current frame and analyzing the morphological correlation between the dead pixel and surrounding pixels, and fusing the spatio-temporal features with the driving current timing signals and surface temperature data to generate a pixel-level dead pixel mask and a diffusion probability heat map; A dynamic adjustment module connected to the spatio-temporal fusion network module for dynamically adjusting the detection window step size and confidence threshold according to the pixel density of the target device to adapt to the detection requirements of devices with different resolutions; A cause diagnosis module connected to the dynamic adjustment module for outputting a probabilistic diagnosis result of the dead pixel cause by combining the physical relevance of the multi-modal data.

[0056] The present invention can be used in many general or special computer system environments or configurations.

[0057] For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, small computers, large computers, distributed computing environments including any of the above systems or devices, and so on.

[0058] The present invention can be described in the general context of computer-executable instructions executed by a computer, such as program modules.

[0059] Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network.

[0060] In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0061] Specifically, those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0062] It should be understood that although the steps in the flowchart of the accompanying drawings are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and they can be executed in other orders.

[0063] Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same moment, but can be executed at different moments, and their execution order is not necessarily sequential. Instead, they can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0064] Obviously, the above-described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. The accompanying drawings show the preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present invention more thorough and comprehensive.

[0065] Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing specific embodiments, or perform equivalent replacements for some of the technical features. Any equivalent structures made by using the specification and drawings of the present invention, directly or indirectly applied to other related technical fields, are equally within the scope of the patent protection of the present invention.

Claims

1. A method for predicting and detecting display screen dead pixel faults, characterized in that, It includes the following steps: Collect visible light images, short-wave infrared images, ultraviolet fluorescence images, drive current timing signals, and surface temperature distribution data of the display screen; Perform dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set; Input the preprocessed multi-modal data set into a spatio-temporal fusion network, extract spatio-temporal correlation features during the bad pixel diffusion process through the spatio-temporal fusion network, and generate a pixel-level bad pixel mask for the current frame and a diffusion probability heat map for a future period based on the spatio-temporal correlation features; According to the bad pixel mask and the diffusion probability heat map, dynamically adjust the detection parameters to adapt to the target device, and output a probabilistic diagnosis result of the bad pixel cause in combination with the physical correlation of the multi-modal data.

2. The method for predicting and detecting dead pixel faults of a display screen according to claim 1, wherein, The dynamic noise injection includes: Apply luminance attenuation noise based on a physical degradation model to the visible light image, and the luminance attenuation equation of the model is ; Among them, is the brightness value of the pixel with coordinates (x, y) at time t, is the initial brightness value of the pixel with coordinates (x, y), is the attenuation coefficient, and its value range is from 0.1 to 0.3, is the time constant, and its value range is from 100 to 500 hours, is the non-linear attenuation factor, and its value range is from 1.2 to 1.8; Superimpose Weibull distribution noise on the driving current timing signal, and its probability density function is, ; where x is the amplitude of the current pulse, is the shape parameter, with a value range of 1.5 to 2.5, is the scale parameter, with a value range of 0.1 to 0.

3.

3. The method for predicting and detecting dead pixel faults of a display screen according to claim 1, wherein, The cross-modal alignment includes: Time synchronization, using the screen vertical synchronization signal as the time reference to align the timestamps of the visible light image, the short-wave infrared image, the ultraviolet fluorescence image, the drive current timing signal, and the surface temperature distribution data; Spatial reference point extraction, based on the time-synchronized short-wave infrared image, extract the circuit trace intersection points as spatial reference points, and the number of spatial reference points is not less than 4 groups; Spatial registration, based on the extracted spatial reference points, solve the affine transformation matrix by the least squares method, and register the visible light image and the ultraviolet fluorescence image to the coordinate system of the short-wave infrared image to achieve sub-pixel level spatial alignment.

4. The method for predicting and detecting the dead pixel fault of a display screen according to claim 1, characterized in that The step of inputting the preprocessed multi-modal data set into a spatio-temporal fusion network, extracting spatio-temporal correlation features during the bad pixel diffusion process through the spatio-temporal fusion network, and generating a pixel-level bad pixel mask for the current frame and a diffusion probability heat map for a future period based on the spatio-temporal correlation features includes the following steps: Time evolution feature extraction, based on the preprocessed multi-modal data set, extract the decay rate of the bad pixel brightness over time and the historical evolution features of the diffusion direction through causal time modeling; Spatial distribution analysis, according to the extracted historical evolution features, locate the spatial distribution area of the bad pixels in the current frame, and analyze the morphological correlation and micro-texture differences between the bad pixels and the surrounding pixels; Cross-modal feature fusion, fuse the obtained spatial distribution features with the drive current timing signal and the surface temperature distribution data, calculate the correlation weights of the optical features and the electrical features through the cross-attention mechanism, and generate a pixel-level bad pixel mask and a diffusion probability heat map.

5. The method for predicting and detecting dead pixel faults of a display screen according to claim 4, wherein, The weight calculation formula of the cross-attention mechanism is as follows: ; Among them, is the query vector corresponding to the i-th optical feature, which is derived from the feature map of visible light, short-wave infrared, or ultraviolet fluorescence images. is the key vector corresponding to the j-th electrical feature, which is derived from the timing feature vector converted from the drive current timing signal. N is the total number of sampling points of the timing signal, and T is the matrix transpose operator. is the attention weight between the optical feature at the i-th spatial position and the electrical feature at the j-th time point. is the natural exponential function.

6. The method for predicting and detecting the dead pixel fault of a display screen according to claim 1, characterized in that, The step of dynamically adjusting the detection parameters to adapt to the target device includes the following steps: Detection window step size adaptation, dynamically adjust the moving step size of the detection window according to the pixel density of the target device according to the formula, use a small step size for high-resolution devices, and a large step size for low-resolution devices; Confidence threshold dynamic setting, based on the statistical features of the diffusion probability heat map, calculate the dynamic threshold through the mean and standard deviation; Multi-frame consistency verification, perform a logical AND operation on the bad pixel masks and heat maps of multiple consecutive frames, and filter out the stable bad pixels that persist in multiple frames to suppress transient noise.

7. A method for predicting and detecting display screen dead pixel faults according to claim 6, characterized in that, The step of outputting a probabilistic diagnosis result of the bad pixel cause in combination with the physical correlation of the multi-modal data, wherein the determination condition of the probabilistic diagnosis result of the bad pixel cause is: Circuit overload determination: When the amplitude of the driving current pulse exceeds the current determination threshold and the local temperature rise exceeds the ambient temperature by 5°C, a circuit overload alarm is triggered. Physical damage determination: If internal crack features are detected in the short-wave infrared image and the ultraviolet fluorescence intensity is lower than 30% of the normal area, it is determined as physical damage. Material aging determination: When the brightness attenuation rate exceeds 0.15% per hour and the chromaticity shift exceeds 8, it is marked as material aging.

8. The method for predicting and detecting the dead pixel fault of a display screen according to claim 7, characterized in that, After the probabilistic diagnosis result of the cause of the dead pixel is output by combining the physical relevance of the multi-modal data, the following steps are further included: Combining the multi-modal features of the circuit overload determination, the physical damage determination, and the material aging determination into a feature vector f, where the features include the amplitude of the driving current pulse, local temperature rise data, short-wave infrared crack features, ultraviolet fluorescence intensity, and chromaticity shift. Based on the feature vector f, calculate the posterior probability of each cause type through the following formula: ； where, is the posterior probability that the cause type of the bad pixel is c under the condition of a given multi-modal feature combination f, is the cause type of the bad pixel. Under the condition that the cause type of the bad pixel is known to be c, is the likelihood probability of observing the feature vector f, is the prior probability of the cause type c of the bad pixel, that is, the initial probability distribution of the fault type when there is no feature information, is the traversal variable of the cause type of the bad pixel, indicating the summation over all possible values of cc, is the summation operation over all possible cause types c' of the bad pixel, is under the condition that the cause type of the bad pixel is known to be and is the likelihood probability of observing the feature vector is the prior probability of the cause type of the bad pixel.

9. A display screen dead pixel fault detection system for performing the display screen dead pixel fault prediction and detection method according to any one of claims 1 to 8, characterized in that, Including: A data acquisition module for synchronously acquiring the visible light image, short-wave infrared image, ultraviolet fluorescence image, driving current time series signal, and surface temperature distribution data of the display screen. A preprocessing module connected to the data acquisition module for performing dynamic noise injection and cross-modal alignment on the data to generate a preprocessed multi-modal data set. A spatio-temporal fusion network module connected to the preprocessing module for extracting the historical evolution features of the attenuation rate and diffusion direction of the dead pixel brightness over time, locating the spatial distribution area of the dead pixel in the current frame and analyzing the morphological relevance between the dead pixel and the surrounding pixels, and fusing the spatio-temporal features with the driving current time series signal and surface temperature data to generate a pixel-level dead pixel mask and a diffusion probability heat map. A dynamic adjustment module connected to the spatio-temporal fusion network module for dynamically adjusting the detection window step size and confidence threshold according to the pixel density of the target device to adapt to the detection requirements of devices with different resolutions. A cause diagnosis module connected to the dynamic adjustment module for outputting a probabilistic diagnosis result of the cause of the dead pixel by combining the physical relevance of the multi-modal data.

Images