OLED lighting rate prediction method based on improved transfer learning
By improving the transfer learning method, utilizing the self-attention-driven VGG16 network and cross-domain alignment technology, dynamically unfreezing convolutional blocks and suppressing noise, the problems of training instability and catastrophic forgetting in OLED brightness prediction are solved, thereby improving prediction accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-09
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies for predicting the brightness of OLED devices suffer from problems such as unstable training, insufficient cross-domain alignment, slow model convergence, and catastrophe forgetting, making it difficult to achieve efficient and stable predictions, especially in small sample scenarios.
An improved transfer learning method is adopted. By constructing a self-attention driven VGG16 network, combined with a lightweight Adapter structure, DANN+MMD cross-domain alignment and EWC regularization, the convolutional blocks are dynamically unfrozen to suppress noise and perform adaptive training, thereby achieving feature distribution alignment and parameter stability.
It significantly improves the accuracy and efficiency of OLED brightness prediction, solves the problems of training instability and catastrophic forgetting in small sample scenarios, and achieves higher prediction accuracy and generalization ability.
Smart Images
Figure CN121811213A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of OLED, in particular to an OLED lighting rate prediction method based on improved transfer learning. BACKGROUND
[0002] OLED devices are highly sensitive to humidity and oxygen, and have relatively poor environmental stability; the hygroscopic agent in the packaging system has a key influence on the device lifetime and lighting rate.
[0003] Traditional lighting rate evaluation highly depends on repeated physical tests, which is long in cycle, high in cost and strong in subjectivity, and it is difficult to form a high-throughput, quantitative engineering method; with the development of deep learning, CNN network shows good potential in material micro-nano image representation, but direct training or transfer with general visual model often faces the following difficulties: 1. Small sample constraint and high cost of label acquisition, which easily leads to overfitting and unstable convergence; 2. Domain distribution difference (domain gap) is significant - there is statistical distribution drift among magnification, lighting conditions, sample batches and process paths of hygroscopic agent micro-nano images, which makes the general pre-training features and target task representation mismatched; 3. Static fine-tuning strategy lacks perception of task correlation of different layers, which easily causes catastrophic forgetting or insufficient transfer, and the generalization ability is limited; Hygroscopic agent micro-nano structure images have problems such as coexistence of multi-scale structure and long-range dependence, obvious imaging heterogeneity, inconsistent features caused by imaging noise changes, and sample imbalance, although the introduction of self-attention mechanism in network structure can enhance the modeling ability of long-range dependence, but if there is no matching transfer learning strategy to deal with small sample and cross-domain problems, the overall prediction performance is still difficult to achieve the stability and robustness required by engineering application. SUMMARY
[0004] In view of the deficiencies of the existing method, the present application solves the problems of unstable training, insufficient cross-domain alignment, slow model convergence and catastrophic forgetting of the existing transfer learning strategy.
[0005] The technical solution adopted by the present application is: an OLED lighting rate prediction method based on improved transfer learning comprises the following steps: Step 1, constructing source domain set and target domain set of OLED hygroscopic agent micro-nano structure images and corresponding OLED lighting rate; As a preferred embodiment of the present application, the OLED image is subjected to image gray scale conversion, histogram equalization, image contrast enhancement, DnCNN denoising, size normalization and data augmentation.
[0006] Step 2, inputting the source domain set into the standard VGG16 network; training the weight parameters by using the standard VGG16 network; Step three, migrate the training weight parameters to the VGG16 network embedded with self-attention, calculate the average attention intensity of each convolutional block, compare the average attention intensity with the unfreezing threshold, and determine whether to freeze the weight parameters of the corresponding convolutional block; As a preferred embodiment of the present application, whether to freeze the weight parameters of the corresponding convolutional block comprises: First, traverse the target domain set Sample, calculate the average attention intensity of each convolutional block, the formula is:
[0007] Second, set the unfreezing threshold If , unfreeze the corresponding convolutional block to participate in gradient update; otherwise, if , freeze the convolutional block and keep the parameters unchanged. Finally, during the training process, after each training period, recalculate the layers in the convolutional block and update the unfreezing state.
[0008] As a preferred embodiment of the present application, the VGG16 network embedded with self-attention is inserted between the convolutional layer and the pooling layer of each convolutional block.
[0009] As a preferred embodiment of the present application, the warm start gate is used to suppress source domain and target domain noise.
[0010] As a preferred embodiment of the present application, the VGG16 network embedded with self-attention is trained using ; The main loss is The adversarial domain alignment loss is The statistical distribution alignment loss is The parameter regularization loss is , , The loss weight coefficient is
[0011] As a preferred embodiment of the present application, the VGG16 network embedded with self-attention is evaluated using RMAE and R2 indicators.
[0012] As a preferred embodiment of the present application, the OLED luminance prediction system based on improved transfer learning comprises: a memory for storing instructions executable by a processor; and a processor for executing instructions to implement an OLED luminance prediction method based on improved transfer learning.
[0013] As a preferred embodiment of the present application, a computer readable medium storing computer program code, which, when executed by a processor, implements an OLED luminance prediction method based on improved transfer learning.
[0014] Advantages of the present application: 1. The present application first proposes an attention-driven hierarchical unfreezing mechanism, and cooperatively integrates a lightweight Adapter structure, a DANN+MMD cross-domain alignment method and an EWC regularization anti-forgetting technology to form a complete transfer learning optimization framework. 2. Based on the attention response of the target domain input at each feature layer, the unfreezing order of different layers is adaptively sorted and trained in stages, avoiding manual setting of the freezing / unfreezing strategy and improving the training stability under small samples.
[0015] 3. DANN + MMD cross-domain alignment strategy: In the transfer process, an adversarial domain classifier and a maximum mean difference (MMD) constraint are introduced simultaneously to enhance the alignment effect of source / target domain feature distribution and reduce the performance decline caused by domain shift; EWC regularization anti-forgetting mechanism: In the target domain fine-tuning stage, the elastic weight constraint (EWC) is applied to the important parameters of the source domain to introduce new task knowledge while maintaining the source domain knowledge, and to alleviate the problem of "catastrophic forgetting" in the transfer process.
[0016] 4. The present application first realizes the overall framework of "level adaptation + parameter lightweight + feature alignment + memory retention" in the material structure image prediction task, which significantly improves the precision and efficiency of small sample prediction. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 The present application is based on an improved transfer learning OLED luminance prediction method flowchart. Figure 2 The present application is a transfer learning structure schematic diagram. Figure 3 The present application is an improved transfer learning logic diagram. Figure 4 The present application is a training set prediction result diagram of each model under the RMAE evaluation index. Figure 5 The present application is a training set prediction result diagram of each model under the R2 evaluation index. Figure 6 The present application is a test set prediction result diagram of each model under the test set RAME evaluation index. Figure 7 The present application is a test set prediction result diagram of each model under the training set R2 evaluation index. DETAILED DESCRIPTION
[0018] The application will be further described below in conjunction with the accompanying drawings and examples, which are simplified schematic diagrams and only schematically show the basic structure of the application, and thus only show the components related to the application.
[0019] Existing transfer learning methods mostly focus on parameter inheritance and unified fine-tuning of general visual tasks, lacking systematic improvement for material micro-nano image scenes: for example, few methods use attention response as a dynamic signal for unfreezing and updating to distinguish the importance of each layer to the target task; it is also less common to incorporate domain-adaptive and parameter importance regularization (such as EWC) into the same training framework to simultaneously alleviate distribution shift and catastrophic forgetting; in the specific application of OLED luminance prediction, how to build a transferable, scalable, and interpretable learning paradigm for micro-nano images under the combined constraints of small sample size, cross-process, multiple magnifications, imaging noise, and data imbalance still lacks effective technical paths.
[0020] As shown in Figure 1 , 2 , 3, an OLED luminance prediction method based on improved transfer learning includes the following steps: Step 1: Obtain an OLED image. Preprocess the OLED image: including image grayscale conversion, histogram equalization, image contrast enhancement, DnCNN denoising, size normalization, and data augmentation.
[0021] Step 2: Establish a source domain set and a target domain set of moisture-absorbing agent micro-nano structure images and corresponding OLED luminance. Step 3: Input the source domain set into the standard VGG16 network; train the weight parameters using the standard VGG16 network. Step 4: Migrate the standard VGG16 network training weight parameters to the VGG16 network embedded with self-attention, calculate the average attention intensity of each convolution block, compare the average attention intensity with the unfreezing threshold, and determine whether to freeze the weight parameters of the corresponding convolution block. The VGG16 network embedded with self-attention (referred to as improved VGG16 network) includes: first convolution block, second convolution block, third convolution block, fourth convolution block, and fifth convolution block; the first convolution block and the second convolution block use two Conv layers, ReLU, self-attention layer, and MaxPoling cascaded; the third convolution block, the fourth convolution block, and the fifth convolution block use three Conv layers, ReLU, self-attention layer, and MaxPoling cascaded; The fifth convolution block is input into the classification layer after two fully connected layers; the classification layer uses a linear activation function. The detailed parameters of the VGG16 network embedded with self-attention can refer to the prior application "OLED Luminance Prediction Method Based on Moisture-absorbing Agent Micro-nano Structure"; The freezing and unfreezing process is as follows: First, the improved VGG16 network embedded in the self-attention layer is recorded as the first Convolution block, =1,2,3,4,5, traverse the target domain set Sample, calculate the average attention intensity of the convolution block, the formula is:
[0022] Where, is the average attention intensity index of the first Convolution block, and according to the unfreezing threshold Perform phased unfreezing; is the Query matrix of the first Sample in the self-attention layer of the first Convolution block; is the Key matrix of the first Sample in the self-attention layer of the first Convolution block; is the dimension of the Key matrix; The operation normalizes the dot product similarity of Query and Key to obtain the attention weight matrix; The operation takes the average of all spatial positions of the weight matrix.
[0023] Second, set the unfreezing threshold , judge of each convolution block: if , it is determined that the convolution block has high correlation with the current OLED brightness prediction task, and the corresponding convolution block is unfrozen to participate in gradient update; otherwise, if , freeze the convolution block and keep the parameters unchanged; that is, freeze and unfreeze the weight parameters of the convolution layer, pooling layer and attention layer in the convolution block.
[0024] Finally, dynamically update the stage, during the training process, every time a training cycle (Epoch) is completed, the layers in the convolution block will be recalculated and the unfreezing state will be updated; this dynamic adjustment mechanism can continuously optimize the participation state of the layers as the model learns the features of OLED micro-nano images, achieving the effect of "task-related layers learning first and irrelevant layers reducing interference".
[0025] According to the characteristics of OLED moisture absorbent micro-nano images, the advantages of the freezing and unfreezing strategy are as follows: 1. Accurately capture task-related features, quantify the degree of correlation between each layer and "micro-nano structure-brightness rate" through attention intensity, prioritize unfreezing layers that contribute significantly to the prediction of brightness rate (such as high layers that can capture the distribution pattern of micropores), avoid ineffective learning of irrelevant features (such as imaging noise and background interference), and improve the targeting of feature extraction. 2. Adaptable to small sample scenarios: The cost of obtaining experimental samples in the OLED field is high (requiring complex processes for preparation and testing). This mechanism only unfreezes the highly correlated layer, reducing the scale of parameters involved in the update and lowering the risk of overfitting under small sample conditions. It can still converge stably in scenarios with small sample datasets. Given the structural characteristics of VGG networks, the advantages of the freeze and unfreeze strategies are as follows: 1. Alignment with hierarchical feature abstraction logic: VGG's shallow layers (first to second convolutional blocks) extract low-level features such as edges and textures, while the deep layers (third to fifth convolutional blocks) extract semantic-level abstract features. This mechanism can distinguish between them by attention intensity: if the OLED brightness is strongly correlated with the global distribution of micropores (deep semantic features), the deep layers will be unfrozen first; if it is strongly correlated with local textures (shallow features), the shallow layers will be unfrozen first, thus achieving precise alignment with VGG's hierarchical feature abstraction logic.
[0026] 2. Reduce training costs and improve efficiency: VGG networks have a large number of parameters, and full fine-tuning is prone to overfitting and time-consuming in small-sample OLED scenarios; this mechanism only unfreezes highly relevant layers, which can reduce the number of parameters that need to be updated by more than 60%.
[0027] The distributional differences between the source domain (general image) and the target domain (micro / nano image) introduce noise, resulting in incorrect gradients or unstable feature update signals at the beginning of training. Without noise suppression, the model may prematurely learn imaging noise, texture artifacts, or domain-specific details in the target domain, affecting its ability to extract general features. To suppress noise, a warm-start gating method is employed. (2) This also improves the learning rate per layer of convolutional blocks in the VGG16 network. Shallow layers are updated in small steps first, while deeper layers are adapted first; among them, It is the Sigmoid function, which acts as a smooth "gating function" with an output value between 0 and 1 to avoid abrupt changes; Warm-up factor over time Increase The control gate function gradually becomes more sensitive from the initial "passivation" to suppress unstable updates in the early stages of training; No. The layer in the first tThe gating coefficient of the wheel represents the degree of participation in the training at this time of the layer, and is between 0-1; the higher the attention intensity, the later the training, and the closer the gating value to 1; is the basic learning rate, and an initial value is globally set; is the first layer in the first t round of training; is the actual learning rate of the layer in the first round of training, which is determined by multiplying the basic learning rate by the gating value; the gating threshold value
[0028] is 0.6-0.8, which is used to distinguish between strongly related and weakly related layers.
[0029] , wherein, is used to drive the source domain and the target domain feature distribution to be aligned in an antagonistic manner, and is added to the total loss as a training regular term; is a source domain sample, and is from general visual data (ImageNet training set); is a target domain sample, and is from an OLED micro-nano image (small sample domain); F is a feature extractor; D is a domain discriminator is a feature extractor; F tries to deceive the domain discriminator, so that the source domain and the target domain features cannot be distinguished; the domain discriminator D tries to distinguish the two domains to form an antagonistic game; is an expectation (expectation loss), which is realized by mini-batch approximation.
[0030] In order to alleviate the feature distribution difference between the source domain and the target domain, the maximum mean discrepancy (MMD) loss is introduced to statistically align the source / target features:
[0031] , wherein, is used to constrain the source domain and the target domain feature distribution to be aligned, and is added to the total loss as a training regular term; n is the number of samples; is the first sample of the source domain; is the first j sample of the target domain; is a kernel function implicit mapping.
[0032] This loss promotes the consistency of features across domains by minimizing the statistical distance of source / target feature distributions.
[0033] Elastic Weight Constraint (EWC) is introduced to prevent catastrophic forgetting.
[0034] EWC is used to constrain important parameters from deviating too much from the original value during target task training. Weight coefficient; control the influence strength of EWC regularization term in the total loss. Parameter index, a constraint is calculated for each parameter of the neural network. The optimal value of the p th parameter in the source task (or pre-trained model); Indicates the importance of the parameter p To the source task. The greater the value, the more important it is and the less it can be changed at will.
[0035] The present application only imposes EWC constraints on the Adapter module and deep parameters, while retaining the knowledge of the source task and adapting to the target task.
[0036] The overall target is obtained by integration, and the joint loss calculation part is:
[0037] Among them, The main loss (such as MSE) of the lighting rate prediction task; Adversarial domain alignment loss; Statistical distribution alignment loss; Parameter regularization to prevent forgetting; , , Respectively, the weight coefficient of the corresponding loss.
[0038] The improved VGG16 network is back propagated through the joint loss function, while realizing prediction task fitting, feature distribution alignment and parameter stability constraint, so as to obtain better generalization ability in small sample target domain.
[0039] Learning rate and training schedule, adopt two-stage schedule: Phase I (stable alignment): only train adapter + deep + attention layer + domain discriminator; shallow layer is frozen, and shallow layer learning rate low =0; Phase II (progressive unfreezing): according to From high to low, unfreeze layer by layer, , and gradually increase Complete fine-grained alignment; Full use of hierarchical adaptive weight decay: More gentle regularization for "important layers"; is the base weight decay value (global initial value); is the adjustment coefficient; where important layers are layers with high attention intensity and large contribution to the light-on rate prediction task; adaptive weight decay decreases, i.e. more gentle regularization, to maintain its representation ability; unimportant layers maintain strong regularization to prevent overfitting; Small sample robustness and imbalance correction. For the imbalance of the light-on rate interval, the quantile weighted MSE is used, and the formula is:
[0040] Where, is the label quantile; Higher weight is given to rare intervals; together with the "adapter-only training" of stage I, it alleviates small sample overfitting.
[0041] Training and optimization, Bayesian optimization of hyperparameters, using Bayesian search key parameters: Table 1 Hyperparameters
[0042] The relative mean absolute error (RMAE), the coefficient of determination (R2), and the mean absolute error (MAE) are used to measure and evaluate the prediction performance of the network model; the calculation formula of the relative mean absolute error (RMAE) is:
[0043] The calculation formula of the coefficient of determination (R2) is:
[0044] Where, is the predicted value, is the true value, is the average value of the true value.
[0045] The improved transfer learning model of the application is tested using three evaluation indicators: as shown in Figure 4 , the model shows the best results, with an average R2 score of 0.83 for the training set and 0.77 for the test set; as shown in Figure 6 , the model also has the lowest RMAE score, with a training score of 6.48 and a test score of 7.02; the convergence speed of the model R2 training set and test set is shown in Figure 5 and Figure 7 .
[0046] With the above ideal embodiments according to the present application as the inspiration, through the above description, relevant staff can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the content of the specification, and must be determined according to the scope of the claims.
Claims
1. A method for predicting OLED brightness rate based on improved transfer learning, characterized in that, Includes the following steps: Step 1: Construct the micro / nano structure image of the OLED hygroscopic agent and the corresponding source and target domain sets for the OLED brightness rate; Step 2: Input the source domain set into the standard VGG16 network; train the weight parameters using the standard VGG16 network. Step 3: Transfer the training weight parameters to the VGG16 network with embedded self-attention, calculate the average attention intensity of each convolutional block, compare the average attention intensity with the unfreeze threshold, and determine whether to freeze the weight parameters of the convolutional block.
2. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, The weight parameters for whether to freeze the convolutional block include: First, traverse the target domain set. Calculate the average attention intensity for each convolutional block from samples. ; Secondly, set a thawing threshold. ,like If so, then unfreeze the corresponding convolutional block to allow it to participate in gradient update; otherwise, if Then freeze the convolution block, keeping the parameters unchanged; Finally, during training, after each training cycle, the layers in the convolutional block are recalculated and the unfrozen state is updated.
3. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, use Train a VGG16 network with embedded self-attention.
4. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, The process involves performing grayscale conversion, histogram equalization, contrast enhancement, DnCNN denoising, size normalization, and data augmentation on OLED images.
5. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, The VGG16 network with embedded self-attention inserts a self-attention layer between the convolutional and pooling layers in each convolutional block.
6. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, Warm-start gating is used to suppress noise in both the source and target domains.
7. The OLED brightness prediction method based on improved transfer learning according to claim 1, characterized in that, The VGG16 network with self-attention embedded using RMAE and R2 metrics was evaluated.
8. An OLED brightness rate prediction system based on improved transfer learning, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the OLED brightness prediction method based on improved transfer learning as described in any one of claims 1-7.
9. A computer-readable medium storing computer program code, characterized in that, The computer program code, when executed by a processor, implements the OLED brightness prediction method based on any one of claims 1-7.