OLED equipment fault diagnosis system based on artificial intelligence

Abstract

The invention discloses an OLED equipment fault diagnosis system based on artificial intelligence, and relates to the technical field of fault diagnosis, and the system comprises the following steps: capturing OLED pixel-level brightness distribution in real time through a microscopic camera, synchronously collecting driving current, voltage and spectral emission data, and aligning all spatio-temporal data by taking a driving signal rising edge as a reference; a defect area is segmented and positioned by adopting a self-adaptive threshold value, topological features are extracted, and a defect distribution tensor containing spatial attributes is constructed; fusing space-time and electro-optical parameters to construct a fourth-order degradation tensor, and coupling a space attenuation weight and a time sensitive function through a space-time convolution operator; microdefect evolution is analyzed by using a 3D-CNN branch, a macroscopic degradation track is learned by using an LSTM branch, and dual-scale features are interactively fused through a mask matrix; and calculating fault probability distribution, generating a fault positioning thermodynamic diagram and residual life prediction, and outputting a diagnosis report. According to the method, the tensor degradation model based on spatial-temporal feature fusion is constructed, so that the sensitivity of the system to hidden faults is enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and particularly relates to an OLED equipment fault diagnosis system based on artificial intelligence. BACKGROUND

[0002] An organic light-emitting diode (OLED) is a solid-state light-emitting device based on organic semiconductor materials. Its core structure is a sandwich design: a plurality of organic thin films (including a hole transport layer, a light-emitting layer, and an electron transport layer) are deposited between two electrodes (an anode / cathode). Under the driving of an electric field, electrons and holes recombine excitons in the light-emitting layer, and photons are released through radiative decay to achieve self-luminescence. Unlike traditional LCD technology, OLED has the characteristics of self-luminescence, ultra-thin flexibility (can be made on a plastic substrate), high contrast (theoretically infinite), wide viewing angle, and fast response, and is widely used in high-end display panels (mobile phones, televisions) and lighting fields.

[0003] Current OLED equipment fault diagnosis systems mainly identify visible defects (such as dark spots and bright lines) through image processing or analyze the macro changes of electrical parameters (such as driving voltage and current efficiency). However, at the early stage of device aging, microscopic hidden faults (such as interface deterioration of the organic layer inside the pixel and sub-pixel level lattice distortion) often only cause nanoscale material structure changes. At this time, neither optical visible defects nor significant deviations of electrical parameters are formed. This cross-scale fault evolution characteristic causes the traditional method to have a monitoring blind area, and the microscopic abnormal signal is overwhelmed by noise, and the macro parameter change lags behind the actual damage process. The existing technology lacks a mechanism for real-time nonlinear coupling of microscopic defect evolution and macro electrical and optical characteristic degradation, making it difficult to capture early hidden faults in time, shortening the intervention maintenance window period. To solve the above problems, the following solutions are proposed. SUMMARY

[0004] The purpose of the present application is to provide an OLED equipment fault diagnosis system based on artificial intelligence, which can capture hidden fault signals that cannot be identified by single-mode monitoring by constructing a spatiotemporal feature fusion tensor degradation model and designing a spatiotemporal convolution operator with physical meaning to realize cross-scale nonlinear coupling, thereby solving the problem of early fault missed detection caused by the disconnection of micro-macro signals in the prior art.

[0005] To solve the above technical problems, the present application is realized by the following technical scheme:

[0006] The present application is an OLED equipment fault diagnosis system based on artificial intelligence, which comprises a data acquisition module, a defect feature extraction module, a degradation tensor modeling module, a double-branch diagnosis network, a feature interaction module, and a fault output module.

[0007] The data acquisition module, defect feature extraction module, degradation tensor modeling module, double-branch diagnostic network, feature interaction module, and fault output module are sequentially connected.

[0008] The working process of the diagnostic system is as follows:

[0009] Step S1, multi-source data synchronous acquisition: real-time capture of OLED pixel-level brightness distribution through a microscopic camera, synchronous acquisition of driving current, voltage, and spectral emission data, and alignment of all spatiotemporal data based on the rising edge of the driving signal;

[0010] Step S2, dynamic defect feature extraction: adaptive threshold segmentation is used to locate the defect area, and topological features such as defect position and area are extracted to construct a defect distribution tensor containing spatial attributes;

[0011] Step S3, tensor degradation modeling: fusion of spatiotemporal and electro-optical parameters to construct a fourth-order degradation tensor, coupling of spatial decay weights and time-sensitive functions through a spatiotemporal convolution operator to realize nonlinear fusion of cross-scale features;

[0012] Step S4, double-branch cross-scale diagnosis: use of a 3D-CNN branch to analyze micro-defect evolution and an LSTM branch to learn macro-degradation trajectories, and through a mask matrix to interactively fuse double-scale features;

[0013] Step S5, fault mode output and visualization: calculation of five types of fault probability distribution, generation of fault localization heat map and remaining life prediction, and output of an interpretable diagnostic report.

[0014] Further, the data acquisition module is used to synchronously acquire multi-source heterogeneous data, capture pixel-level brightness distribution through a microscopic camera, collect current and voltage time series through an electrical signal sensor, record emission spectrum through a spectrum analyzer, and align time stamps based on the driving signal;

[0015] The defect feature extraction module is used to real-time detect micro-defects, use a dynamic threshold segmentation algorithm to identify abnormal pixel regions, extract defect topological features through connected component analysis, and generate a four-channel defect distribution tensor;

[0016] The degradation tensor modeling module is used to construct a cross-scale degradation model, integrate spatiotemporal features into a fourth-order tensor, couple micro-defect evolution and macro-electro-optical property degradation through a spatiotemporal convolution operator, and output a feature tensor with physical constraints;

[0017] The double-branch diagnostic network is used to extract cross-scale features in parallel, a micro-branch uses a 3D-CNN to analyze defect cluster spatiotemporal evolution, a macro-branch uses an LSTM to learn electro-optical parameter degradation trajectories, and separated feature learning is realized;

[0018] The feature interaction module is used for fusing multi-scale diagnostic information, performing mask-weighted fusion of micro defect features and macro aging features, and generating a joint feature representation with spatial sensitivity in combination with defect position prior knowledge;

[0019] The fault output module is used for generating an interpretable diagnostic result, outputting a fault probability distribution through a classifier, generating a fault positioning heat map based on a gradient, and predicting a remaining service life in combination with a time-sensitive function.

[0020] Further, the step S1 of synchronously collecting multi-source data specifically includes the following steps:

[0021] Step S11: capturing spatial domain data of the OLED panel through a high-resolution microscopic camera to obtain a pixel-level RGB three-channel brightness distribution matrix;

[0022] Step S12: synchronously collecting electrical parameters in a device driving process, including millisecond-level precision driving current and voltage waveforms, and simultaneously recording spectral emission characteristics of each pixel point in a visible light band by using a spectrometer;

[0023] Step S13: establishing a time-space synchronization label: taking a rising edge of a device driving signal as a reference time origin, marking a unified millisecond-level time stamp on the microscopic image, electrical parameters and spectral response data, and constructing a time-space aligned multi-dimensional data set;

[0024] This step synchronously collects pixel-level brightness distribution, driving current and voltage and emission spectrum data of the OLED panel through a high-frame-rate microscopic camera, an electrical sensor and a spectrometer, aligns the time-space label with the driving signal as a reference, and constructs a millisecond-level precision multi-modal data set.

[0025] Further, the step S2 of dynamically extracting defect features specifically includes the following steps:

[0026] Step S21: detecting a defect region for the OLED brightness distribution data collected at each time point by calculating a second-order gradient amplitude of brightness change of each pixel position, and dynamically adjusting a threshold value compared with the amplitude over time, and the formula is:

[0027]

[0028] τ dyn (t)=α×e -βt +γ;

[0029] In the formula, B t (x,y) is a binary defect mask at time t and spatial position (x, y), t is a time sampling point index, is a Laplacian operator, L t(x, y) is the RGB luminance vector at time, location (x, y), ||·|| is the Euclidean norm, τ dyn (t) is a dynamic threshold function, a is an initial threshold coefficient, e is a natural constant, β is a time decay coefficient, t is a continuous time variable, γ is a lower limit of the steady-state threshold;

[0030] Step S22: Connected component analysis is performed on the marked defect points, only continuous regions with an area exceeding 10 pixels are retained as effective defects; for each effective defect region, the center coordinate position, the covered pixel area, the boundary perimeter and the shape factor reflecting the shape irregularity are extracted;

[0031] This step uses a dynamic threshold segmentation algorithm to identify the microscopic defect regions in each frame of image, extracts the topological features (position / area / shape factor) of the defects, generates a defect feature tensor that fuses the spatial distribution and morphological attributes, and realizes the quantitative characterization of physical defects.

[0032] Further, the step S3 of tensor degeneration modeling specifically includes the following steps:

[0033] Step S31: A fourth-order spatio-temporal degeneration tensor structure is constructed, which covers all pixel positions of the display panel along the spatial dimension, continuously stacks feature data of multiple sampling time points along the time dimension, and contains six types of key parameters in the feature channel dimension: pixel area average luminance, current efficiency value, standard chroma coordinate value, defect distribution density, driving voltage offset and spectral redshift;

[0034] Step S32: For each spatio-temporal position in the tensor, the corresponding spatial weight factor and time weight factor are calculated respectively, and the spatial and time weight factors are cross-scale coupled with the original degeneration features, and the formula is:

[0035]

[0036] In the formula, is the output fused feature tensor, i is the spatial horizontal coordinate index, j is the spatial vertical coordinate index, k is the time axis index, is the feature vector of the degeneration tensor at position (i, j, k), is the tensor outer product operator, W s (i, j) is a spatial weight function, is a custom spatio-temporal convolution operator, W t (k) is a time weight function; i0, j0 is the current defect cluster center coordinate, σ s is a spatial Gaussian decay coefficient, tanh is a hyperbolic tangent activation function, D t (i, j) is the feature vector of the defect distribution tensor at (i, j); η is a time-sensitive factor, k is a discrete time index, tfail To predict the failure time point, is a discrete difference operator, V(k) is the driving voltage at time k, and I(k) is the driving current at time k;

[0037] This step builds a fourth-order space-time degradation tensor, integrating multiple-dimensional features such as brightness, electrical parameters, and defect evolution. Through an original space-time convolution operator (coupling spatial attenuation function and time-sensitive weight), the nonlinear fusion of microscopic defects and macro aging is realized, forming a cross-scale degradation model.

[0038] Further, the step S4, the double-branch cross-scale diagnosis specifically includes the following steps:

[0039] Step S41: input the space-time degradation tensor into the double-branch neural network for parallel processing: the micro-diagnosis branch adopts a three-dimensional convolution network structure, scans the local area through multiple layers of convolution kernels, and extracts the morphological evolution features of defect clusters over time; the macro-diagnosis branch adopts a long short-term memory network to analyze the degradation trajectory of global electro-optical characteristic parameters in the time series;

[0040] Step S42: realize cross-scale fusion by fusing the double-path outputs through a feature interaction module, and the formula is:

[0041] h fusion =ReLU(W micro h micro +W macro h macro -b th )⊙M sparse ;

[0042] In the formula, h fusion is the fusion feature vector, ReLU is the rectified linear unit activation function, W micro is the micro-branch weight matrix, h micro is the micro-feature output by the 3D-CNN, W macro is the macro-branch weight matrix, h macro is the macro-feature output by the LSTM, b th is the fusion threshold bias vector, is the Hadamard product, and M sparse is the defect position sparse mask matrix.

[0043] This step uses a double-branch neural network to extract local defect cluster evolution features (3D-CNN) and global electro-optical degradation trajectories (LSTM), respectively. Through a feature interaction module, the micro-abnormalities are dynamically associated with macro parameters to generate a fusion diagnosis feature vector under the constraint of a sparse mask.

[0044] Further, the step S5, the fault mode output and visualization specifically includes the following steps:

[0045] Step S51: input the fusion feature vector into the diagnosis layer, calculate the probability distribution of the fault type through the softmax classifier, and the formula is:

[0046]

[0047] c∈{dark spot, bright line, color deviation, efficiency decay, catastrophic failure};

[0048] wherein, is the given tensor is the probability of the fault type c, θ c is the classification weight vector of the fault type c, T is the vector transposition operator, h fusion is the fusion feature vector, j is the fault type index, and C is the total number of fault types.

[0049] Step S52: generate a pixel-level fault positioning heat map based on the feature gradient, intuitively label the abnormal area of the panel, and the fault positioning heat map is:

[0050]

[0051] Meanwhile, the remaining life is calculated by combining the predicted fault time point and the current time difference, and a complete diagnosis report containing the specific fault type probability, spatial positioning heat map and quantitative life prediction is output.

[0052] The remaining life prediction formula is:

[0053]

[0054] wherein, RUL is the remaining useful life, t fail is the predicted fault time point, t current is the current time point, κ is the life correction coefficient, is the time gradient operator, h macro is the macro feature vector, and ||·||2 is the L2 norm.

[0055] This step calculates the probability distribution of five typical fault types based on the fusion feature, simultaneously generates a fault positioning heat map (gradient sensitivity analysis) and a remaining life prediction value (fusion failure time point and degradation gradient), and outputs an interpretable diagnosis report.

[0056] The present application has the following beneficial effects:

[0057] 1. The application couples the micro-defect evolution and the macro-electro-optical parameter degradation non-linearly by constructing a four-order space-time degradation tensor; the adaptive dynamic threshold detection in the spatial domain combined with the electrical property derivative weight in the time domain can capture the early abnormal signal that cannot be detected by traditional single-mode analysis; when the local area appears nanoscale lattice distortion, the weak current efficiency fluctuation caused by it will be amplified by the space-time convolution operator in the tensor model, and the spectral redshift feature at the corresponding position will be formed; this cross-scale modeling mechanism enhances the sensitivity of the system to hidden faults, so that the diagnostic window period is advanced, providing conditions for preventive maintenance.

[0058] 2. The application adopts a collaborative architecture of a double-branch neural network and a feature interaction module; the micro-branch focuses on the morphological evolution law of the defect cluster, and the macro-branch analyzes the degradation trajectory of the overall panel electro-optical property, and the two are dynamically fused through a sparse mask matrix with physical constraints; this design can effectively distinguish between real fault signals and environmental noise interference: if the instantaneous fluctuation of the driving voltage is not accompanied by a change in the spatial correlation of the defect area, it will be judged as noise by the interaction module; otherwise, when the micro-defect expansion and the macro-current efficiency derivative mutation meet the space-time coupling condition, a high-confidence alarm is triggered; this design can avoid the risk of misdiagnosis.

[0059] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0061] Figure 1 The structure diagram of the OLED equipment fault diagnosis system based on artificial intelligence of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] Please refer to Figure 1As shown, the present application is an OLED device fault diagnosis system based on artificial intelligence, the diagnosis system comprises a data acquisition module, a defect feature extraction module, a degradation tensor modeling module, a double-branch diagnosis network, a feature interaction module and a fault output module;

[0064] The data acquisition module, the defect feature extraction module, the degradation tensor modeling module, the double-branch diagnosis network, the feature interaction module and the fault output module are connected in sequence.

[0065] The working process of the diagnosis system is as follows:

[0066] Step S1, multi-source data synchronous acquisition: real-time capture of OLED pixel-level brightness distribution through a microscopic camera, synchronous acquisition of driving current, voltage and spectral emission data, and alignment of all space-time data based on the rising edge of the driving signal;

[0067] Step S2, dynamic defect feature extraction: adaptive threshold segmentation is used to locate the defect area, topological features are extracted, and a defect distribution tensor containing spatial attributes is constructed;

[0068] Step S3, tensor degradation modeling: four-order degradation tensors are constructed by fusing space-time and electro-optical parameters, and space attenuation weights and time sensitive functions are coupled through space-time convolution operators to realize nonlinear fusion of cross-scale features;

[0069] Step S4, double-branch cross-scale diagnosis: 3D-CNN branch is used to analyze micro-defect evolution, and LSTM branch is used to learn macro-degradation trajectory, and double-scale features are interactively fused through a mask matrix;

[0070] Step S5, fault mode output and visualization: calculate the probability distribution of five types of faults, generate fault positioning heat map and remaining life prediction, and output an interpretable diagnosis report.

[0071] The data acquisition module is used for synchronous acquisition of multi-source heterogeneous data, the pixel-level brightness distribution is captured through a microscopic camera, the current voltage time sequence is collected through an electrical signal sensor, the emission spectrum is recorded through a spectrum analyzer, and the time stamp is aligned with the driving signal;

[0072] The defect feature extraction module is used for real-time detection of micro-defects, a dynamic threshold segmentation algorithm is used to identify abnormal pixel regions, topological features of defects are extracted through connected component analysis, and a four-channel defect distribution tensor is generated;

[0073] The degradation tensor modeling module is used for constructing a cross-scale degradation model, integrating space-time features into a four-order tensor, coupling micro-defect evolution and macro-electro-optical property degradation through a space-time convolution operator, and outputting a feature tensor with physical constraints;

[0074] The dual-branch diagnostic network is used for extracting cross-scale features in parallel, a micro-branch is used for analyzing spatial-temporal evolution of defect clusters by using a 3D-CNN, and a macro-branch is used for learning a degradation trajectory of electro-optical parameters by using an LSTM, so that separated feature learning is realized.

[0075] The feature interaction module is used for fusing multi-scale diagnostic information, performing mask-weighted fusion of micro-defect features and macro-aging features, and generating a joint feature representation with spatial sensitivity in combination with defect position prior knowledge.

[0076] The fault output module is used for generating an interpretable diagnostic result, outputting a fault probability distribution through a classifier, generating a fault positioning heat map based on a gradient, and predicting a remaining useful life in combination with a time-sensitive function.

[0077] Step S1, the multi-source data synchronous acquisition specifically includes the following steps:

[0078] Step S11: spatial domain data of the OLED panel is captured by a high-resolution microscopic camera (sampling rate ≥ 60 fps), and a pixel-level RGB three-channel brightness distribution matrix is obtained (m, n are panel pixel sizes);

[0079] Step S12: time domain data is synchronously acquired:

[0080] Electrical parameters: driving current I(t), voltage V(t) (sampling frequency 1 kHz);

[0081] Spectrum response: emission spectrum S(x, y, λ, t) of each pixel at wavelength λ ∈ [380, 780] nm;

[0082] Step S13: time and space synchronization label is established: time stamp t of all data is aligned based on the rising edge of the device driving signal k = k × Δt, (Δt = 1 ms).

[0083] Step S2, dynamic defect feature extraction specifically includes the following steps:

[0084] Step S21: spatial defect detection:

[0085] For each frame L t Adaptive threshold segmentation is applied:

[0086]

[0087] τ dyn (t) = α × e -βt + γ;

[0088] In the formula, B t (x, y) is a binary defect mask at time t and spatial position (x, y), t is a time sampling point index, Laplace operator, L t (x,y) is the RGB intensity vector at time, location (x, y), ||·|| is the Euclidean norm, τ dyn (t) is the dynamic threshold function, a is the initial threshold coefficient, e is the natural constant, β is the time decay coefficient, t is the continuous time variable, γ is the lower limit of the steady-state threshold;

[0089] Step S22: extracting defect topological features:

[0090] Calculating the connected domain set

[0091] Generating defect distribution tensor The channel dimension includes: position coordinates, area, perimeter, shape factor.

[0092] Step S3, tensor degradation modeling specifically includes the following steps:

[0093] Step S31: constructing a spatiotemporal degradation tensor

[0094] Stacking the features of T sampling time points along the time axis;

[0095] The feature channel dimension includes: intensity mean, current efficiency, color coordinates (u ' ,v ' ), defect density, driving voltage offset, spectral redshift;

[0096] Step S32: defining a coupled tensor operation, generating a fused high-dimensional feature expression through tensor outer product operation and a custom spatiotemporal convolution operation:

[0097]

[0098] In the formula, is the output fused feature tensor, i is the spatial horizontal coordinate index, j is the spatial vertical coordinate index, k is the time axis index, is the feature vector of the degradation tensor at position (i, j, k), is the tensor outer product operator, W s (i,j) is the spatial weight function, is the custom spatiotemporal convolution operator, W t (k) is the time weight function; i0, j0 is the current defect cluster center coordinate, σ s is the spatial Gaussian decay coefficient, tanh is the hyperbolic tangent activation function, D t (i,j) is the feature vector of the defect distribution tensor at (i,j); η is the time sensitivity factor, k is the discrete time index, t fail is the predicted failure time point, For discrete difference operators, V(k) is the driving voltage at time k, and I(k) is the driving current at time k.

[0099] Step S4, the bi-branch cross-scale diagnosis specifically includes the following steps:

[0100] Step S41: ... Input a two-branch neural network:

[0101] Branch 1 (Microscopic Diagnosis): 3D convolutional network extracts local defect cluster evolution features;

[0102] Branch 2 (Macroscopic Diagnosis): The LSTM network learns the global electro-optic property degradation trajectory;

[0103] Step S42: The feature interaction module achieves cross-scale fusion.

[0104] h fusion =ReLU(W micro h micro +W macro h macro -b th )⊙M sparse ;

[0105] In the formula, h fusion To fuse feature vectors, ReLU is the rectified linear unit activation function, W micro Let h be the micro-branch weight matrix. micro W represents the microscopic features output by 3D-CNN. macro For the macro-branch weight matrix, h macro For the macroscopic characteristics of the LSTM output, b th Let M be the fusion threshold bias vector, ⊙ be the Hadamard product, and M be the fusion threshold bias vector. sparse This is a sparse mask matrix for the defect location.

[0106] Step S5, Fault Mode Output and Visualization, specifically includes the following steps:

[0107] Step S51: The diagnostic layer calculates the fault probability:

[0108]

[0109] c∈{dark spot, bright line, color shift, efficiency degradation, catastrophic failure};

[0110] In the formula, For a given tensor The probability of fault type c, θ c Let h be the classification weight vector for fault type c, T be the vector transpose operator, and h be the classification weight vector for fault type c. fusion The feature vector is fused, where j is the fault type index and C is the total number of fault types;

[0111] Step S52: generating a report:

[0112] Fault location heat map:

[0113]

[0114] Remaining useful life prediction:

[0115]

[0116] where RUL is the remaining useful life, t fail is the predicted failure time point, t current is the current time point, and K is a life correction coefficient, is the time gradient operator, h macro is the macro feature vector, and ||·||2 is the L2 norm.

[0117] One specific application of this embodiment is:

[0118] Implementation background: the device under test is a 12-inch OLED vehicle display screen (resolution 2560x1440, pixel density 203PPI); the aging condition is continuous work for 500 hours in an 85℃ / 85%RH environmental chamber; the driving mode is constant current driving (I=20mA), and the gray value changes according to a sine wave (period 15s);

[0119] Implementation steps:

[0120] Step S1, data acquisition and synchronization

[0121] Spatial domain data acquisition: capture the real-time luminance distribution of the center area (1024x768 pixels) of the display screen;

[0122] Obtain the RGB three-channel luminance matrix sequence:

[0123]

[0124] Time domain data synchronization:

[0125] Electrical parameters: record the driving voltage V(t) and current I(t) waveforms

[0126] V(t) = {3.2, 3.3, 3.25, …} V, I(t) = {19.8, 20.1, 19.9, …} mA;

[0127] Spectral parameters: collect the full-screen average spectrum S(λ, t) every 30 seconds;

[0128] Synchronization mechanism: use the rising edge of the driving signal as the reference to align all data timestamps:

[0129] t k = k x 0.001 seconds, k = 0, 1, 2, …, 1, 800,000;

[0130] Step S2, defect feature extraction

[0131] Dynamic defect detection:

[0132] Take aging coefficients a = 0.12, b = 0.018, g = 0.04;

[0133] Calculate the detection threshold at 120 hours (t = 432000 seconds):

[0134] τ dyn = 0.12 x e -0.018×432 + 0.04 = 0.061;

[0135] Perform Laplace edge detection on the brightness matrix:

[0136]

[0137] Defect topology analysis:

[0138] Detect the main defect area c1:

[0139] Position coordinates: (x = 512, y = 384);

[0140] Area: 28 pixels;

[0141] Shape factor: 0.42 (irregular polygon);

[0142] Generate defect distribution tensor:

[0143] Step S3, degradation tensor modeling

[0144] Construct the spatiotemporal degradation tensor:

[0145] Intercept t = 400-450 hours data (T = 180,000 sampling points);

[0146] Tensor structure As follows:

[0147]

[0148] STTDM fusion calculation:

[0149] Set parameters: s s = 45, n = 0.12, t fail = 480 hours;

[0150] Calculate the spatial weight matrix:

[0151] Computing the time weight function:

[0152] Performing tensor fusion operation:

[0153]

[0154] Step S4, cross-scale fault diagnosis

[0155] Dual-branch network processing:

[0156] Micro-branch: 3D CNN processes local region (64x64x30000x6);

[0157] Macro-branch: LSTM processes global features (1024x768x180000x6 after average pooling);

[0158] Feature interaction fusion: generate defect mask M sparse (512, 384 positions are 1);

[0159] Computing fusion features:

[0160]

[0161] Step S5, output of diagnosis result

[0162] Fault probability calculation:

[0163]

[0164] Visualized diagnosis report:

[0165] Fault location heat map:

[0166]

[0167] Remaining life prediction:

[0168] In the description of the present specification, the description referring to the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0169] The preferred embodiments of the application disclosed above are only to facilitate the elucidation of the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments described. Obviously, many modifications and variations can be made in light of the teachings above. The description is chosen and described in order to best explain the principles of the application and its practical application to thereby enable others skilled in the art to best utilize the application and get the best results from the application. The application is only limited by the claims and their full scope and equivalents.

Claims

1. An OLED device fault diagnosis system based on artificial intelligence, characterized in that, The diagnostic system includes a data acquisition module, a defect feature extraction module, a degradation tensor modeling module, a bi-branch diagnostic network, a feature interaction module, and a fault output module. The data acquisition module, defect feature extraction module, degradation tensor modeling module, dual-branch diagnostic network, feature interaction module, and fault output module are connected in sequence. The workflow of the diagnostic system is as follows: Step S1: Multi-source data synchronous acquisition: Real-time capture of OLED pixel-level brightness distribution using a microscope camera, synchronous acquisition of driving current, voltage and spectral emission data, and alignment of all spatiotemporal data with the rising edge of the driving signal as the reference. Step S2, Dynamic Defect Feature Extraction: Adaptive threshold segmentation is used to locate the defect region, extract topological features, and construct a defect distribution tensor containing spatial attributes; Step S3, Tensor Degradation Modeling: Construct a fourth-order degradation tensor by fusing spatiotemporal and electro-optic parameters, and achieve cross-scale feature nonlinear fusion by coupling spatial decay weights and time-sensitive functions through spatiotemporal convolution operators; Step S4, Dual-branch cross-scale diagnosis: 3D-CNN branch is used to analyze the evolution of micro-defects, LSTM branch is used to learn the macro-degradation trajectory, and dual-scale features are fused through mask matrix interaction; Step S5, Fault Mode Output and Visualization: Calculate the fault probability distribution, generate a fault location heatmap and remaining life prediction, and output an interpretable diagnostic report.

2. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, The data acquisition module is used to synchronously acquire multi-source heterogeneous data, capture pixel-level brightness distribution through a microscope camera, collect current and voltage timing data through an electrical signal sensor, record emission spectrum through a spectrum analyzer, and align timestamps with drive signals. The defect feature extraction module is used to detect micro defects in real time, uses a dynamic threshold segmentation algorithm to identify abnormal pixel regions, extracts defect topological features through connected component analysis, and generates a four-channel defect distribution tensor. The degradation tensor modeling module is used to construct a cross-scale degradation model, which integrates spatiotemporal features into a fourth-order tensor. It couples the evolution of microscopic defects with the degradation of macroscopic electro-optic properties through a spatiotemporal convolution operator, and outputs a feature tensor that integrates physical constraints. The dual-branch diagnostic network is used to extract cross-scale features in parallel. The micro-branch uses 3D-CNN to analyze the spatiotemporal evolution of defect clusters, and the macro-branch uses LSTM to learn the electro-optic parameter degradation trajectory, thus realizing separate feature learning. The feature interaction module is used to fuse multi-scale diagnostic information, perform masked weighted fusion of micro-defect features and macro-aging features, and generate a spatially sensitive joint feature representation by combining prior knowledge of defect location. The fault output module is used to generate interpretable diagnostic results, output fault probability distribution through a classifier, generate a fault location heatmap based on gradient, and predict the remaining service life by combining a time-sensitive function.

3. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, Step S1, multi-source data synchronous acquisition, specifically includes the following steps: Step S11: Capture the spatial domain data of the OLED panel using a high-resolution microscope camera to obtain a pixel-level RGB three-channel brightness distribution matrix; Step S12: Synchronously acquire electrical parameters during the device driving process, including millisecond-level precision driving current and voltage waveforms, and simultaneously use a spectrometer to record the spectral emission characteristics of each pixel in the visible light band. Step S13: Establish spatiotemporal synchronization tags: Using the rising edge of the device drive signal as the reference time origin, assign a unified millisecond-level timestamp to the microscopic images, electrical parameters, and spectral response data to construct a spatiotemporally aligned multidimensional dataset.

4. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, Step S2, dynamic defect feature extraction, specifically includes the following steps: Step S21: By calculating the second-order gradient amplitude of the brightness change at each pixel position, defect area detection is performed on the OLED brightness distribution data collected at each time moment, and the threshold for comparison with the amplitude is dynamically adjusted over time. Step S22: Perform connected component analysis on the marked defect points, and retain only continuous regions with an area of ​​more than 10 pixels as valid defects; for each valid defect region, extract its center coordinates, the area of ​​the covered pixels, the boundary perimeter, and the morphological factor reflecting the irregularity of the shape.

5. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, Step S3, tensor degradation modeling, specifically includes the following steps: Step S31: Construct a fourth-order spatiotemporal degradation tensor structure. This structure covers all pixel positions of the display panel along the spatial dimension and continuously stacks feature data from multiple sampling times along the temporal dimension. The feature channel dimension includes six key parameters: average brightness of the pixel area, current efficiency value, standard chromaticity coordinate value, defect distribution density, driving voltage offset, and spectral redshift. Step S32: For the feature vector at each spatiotemporal location in the tensor, calculate its corresponding spatial weight factor and temporal weight factor, and couple the spatial and temporal weight factors with the original degenerate features across scales.

6. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, Step S4, the dual-branch cross-scale diagnosis, specifically includes the following steps: Step S41: Input the spatiotemporal degradation tensor into a parallel-processing dual-branch neural network: the microscopic diagnosis branch adopts a three-dimensional convolutional network structure, which scans the local area through multiple convolutional kernels to extract the morphological evolution characteristics of defect clusters over time; the macroscopic diagnosis branch adopts a long short-term memory network to analyze the degradation trajectory of global electro-optical characteristic parameters over time. Step S42: Achieve cross-scale fusion by fusing the dual outputs through the feature interaction module.

7. The OLED device fault diagnosis system based on artificial intelligence according to claim 1, characterized in that, Step S5, fault mode output and visualization, specifically includes the following steps: Step S51: Input the fused feature vector into the diagnostic layer and calculate the probability distribution of the fault type using the softmax classifier; Step S52: Generate a pixel-level fault location heatmap based on feature gradients to visually mark abnormal areas on the panel; at the same time, combine the predicted fault time point with the current time difference to calculate the remaining lifespan and output a complete diagnostic report containing the probability of specific fault types, spatial location heatmap, and quantified lifespan prediction.

Images