Quantum-inspired active learning-based defect detection sample labeling method

Abstract

This invention relates to the field of sample labeling technology, and more particularly to a quantum-inspired active learning-based defect detection sample labeling method. This method can reduce the amount of manual labeling and lower labeling costs. Specifically, by balancing sample representativeness and uncertainty through a unified process design, it first obtains a set of central clusters based on Hamiltonian operator clustering, then calculates the degree of quantum entanglement to screen representative sample sets, and combines virtual-time evolution and the Schrödinger equation to process the remaining unlabeled sample set. The fusion result yields the sample set to be labeled. This eliminates the need to distinguish between budget scenarios, can adaptively optimize sample selection, improve labeling efficiency, and effectively solve the problem of strategy uniformity. Furthermore, by mapping samples to quantum states through self-supervised learning, optimizing clustering accuracy based on Hamiltonian operators, quantifying sample associations using quantum entanglement, and capturing uncertainty with virtual-time evolution and the Schrödinger equation, the quality of labeled samples can be improved, overcoming the representational limitations of classical frameworks.

Description

Technical Field

[0001] This invention relates to the field of sample labeling technology, and in particular to a quantum-inspired active learning method for defect detection sample labeling. Background Technology

[0002] In the field of organic light-emitting diode (OLED) defect detection, a massive number of inspection samples are generated during the OLED product manufacturing process. Manually labeling all samples would consume a lot of time, manpower, and costs. Therefore, active learning has become a key technology to solve this problem. Active learning can train a high-performance defect detection model by selecting a small number of key samples for labeling.

[0003] In terms of existing technologies, Guy Hacohen et al., in their paper "Active Learning on a Budget: Opposite Strategies Suit High and Low Budgets" published at the 2022 International Conference on Machine Learning (ICML), proposed two classic active learning sampling strategies: in low-budget scenarios, they prioritize the annotation of samples with the least uncertainty in the model to quickly improve model performance; in high-budget scenarios, they select samples that best represent the data distribution structure to optimize the model's generalization ability. This existing technology achieves active learning through clustering and density sampling, revealing the key dependence between strategy effectiveness and budget size.

[0004] However, this existing technology has significant drawbacks. On the one hand, its strategy is singular, using independent strategies in high- and low-budget scenarios, lacking a unified theoretical framework to balance the representativeness and uncertainty of the samples, and failing to achieve adaptive adjustment of the strategy. In practical applications, the low-budget stage not only requires uncertain samples to quickly start model training, but also requires an appropriate amount of representative samples to ensure data coverage. On the other hand, the high-budget stage, while consolidating the coverage of representative samples, still requires uncertain samples to further improve the model accuracy. Existing technologies cannot balance both aspects, resulting in the selected samples being either too conservative or too aggressive.

[0005] On the other hand, the existing technology is entirely based on the classical machine learning framework and fails to utilize the properties of quantum state superposition and quantum entanglement in quantum mechanics, thus failing to more precisely characterize the complex relationships and uncertainties between samples. The core reason for the above defects is that traditional active learning methods have limited ability to characterize the intrinsic structure of data and lack a unified theoretical system that can simultaneously handle the two key factors of representativeness and uncertainty, resulting in poor annotation efficiency and model performance when faced with massive OLED detection samples.

[0006] It is evident that existing technologies still need improvement and enhancement. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the present invention aims to provide a quantum-inspired active learning method for labeling defect detection samples. This method breaks through the limitations of existing technologies in terms of both strategy framework and characterization capability, providing a more efficient and accurate solution for labeling OLED defect detection samples.

[0008] The first aspect of this invention provides a quantum-inspired active learning-based defect detection sample labeling method, comprising: acquiring an original defect detection sample set; performing quantum state encoding on the original defect detection sample set through self-supervised learning to obtain an initial labeled sample set and an initial unlabeled sample set, and setting iterative hyperparameters; clustering the initial unlabeled sample set based on a predefined Hamiltonian operator to obtain a central cluster set and sample sets for each cluster; calculating the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set; filtering each cluster sample set based on the degree of quantum entanglement to obtain a representative sample set and a remaining unlabeled sample set; and processing the remaining unlabeled sample set based on imaginary time evolution and the Schrödinger equation. The sample set is filtered, and the filtering results are merged with the representative sample set to obtain the sample set to be labeled. The sample set to be labeled is labeled based on the preset defect classification criteria to obtain the labeled sample set. The initial labeled sample set and the initial unlabeled sample set are updated based on the labeled sample set to obtain the updated labeled sample set and the updated unlabeled sample set. The updated unlabeled sample set is then used to replace the initial unlabeled sample set. The process returns to the step of clustering the initial unlabeled sample set based on the predefined Hamiltonian operator to obtain the set of central clusters and the sample sets of each cluster. The iteration stopping condition is determined based on the set iteration hyperparameters. If the condition is met, the updated labeled sample set is output as the final labeled set.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the step of obtaining the original sample set for defect detection, encoding the original sample set for defect detection into quantum states through self-supervised learning to obtain an initial labeled sample set and an initial unlabeled sample set, and setting iterative hyperparameters includes: obtaining the original sample set for defect detection, which includes multiple original samples; inputting each original sample into a self-supervised learning function, extracting sample features through a simCLR architecture to map each original sample into a quantum state, wherein the quantum state satisfies a normalization condition; integrating the quantum states of all original samples to obtain a set of quantum states, and initializing the labeled sample set and the unlabeled sample set to obtain an initial labeled sample set and an initial unlabeled sample set, wherein the initial labeled sample set is empty, and the initial unlabeled sample set is the set of quantum states; and setting iterative hyperparameters, which are related to the size of the original sample set for defect detection, and include a total number of iteration rounds, a number of labeled samples per round, a clustering convergence threshold, and a maximum number of clustering iterations.

[0010] Optionally, in a second implementation of the first aspect of the present invention, the step of clustering the initial unlabeled sample set based on a predefined Hamiltonian operator to obtain a set of central clusters and a set of samples for each cluster includes: calculating the number of samples selected corresponding to the current iteration round based on the current iteration round and the number of labeled samples in each round; performing a random selection operation on the initial unlabeled sample set to construct an initial set of cluster centers, the initial set of cluster centers including multiple initial cluster centers, the number of initial cluster centers being consistent with the number of samples selected; performing iterative optimization clustering on the initial unlabeled sample set based on the predefined Hamiltonian operator and the initial set of cluster centers; determining whether the clustering iteration stopping condition is met based on the clustering convergence threshold and the maximum number of clustering iterations, and if so, outputting the set of central clusters and the set of samples for each cluster.

[0011] Optionally, in a third implementation of the first aspect of the present invention, the step of calculating the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set, and performing a screening process on each cluster sample set based on the degree of quantum entanglement to obtain a representative sample set and a remaining unlabeled sample set, includes: obtaining a preset sorting criterion; performing a sorting operation on the unlabeled samples in each cluster sample set based on the preset sorting criterion to obtain a sorted sample set for each cluster; calculating the degree of quantum entanglement between each unlabeled sample in each sorted sample set and the central cluster set using von Neumann entropy; and performing a screening operation on each sorted sample set based on the degree of quantum entanglement to construct a representative sample set and a remaining unlabeled sample set, wherein the number of unlabeled samples in the representative sample set is less than the number of labeled samples in each round.

[0012] Optionally, in a fourth implementation of the first aspect of the present invention, the step of filtering the remaining unlabeled sample set based on imaginary time evolution and the Schrödinger equation, and fusing the filtering results and the representative sample set to obtain the sample set to be labeled, includes: calculating the degree of quantum entanglement between each unlabeled sample in the remaining unlabeled sample set and the central cluster set using von Neumann entropy; sorting the remaining unlabeled sample set based on the degree of quantum entanglement to obtain a sorted unlabeled sample set; calculating the discard probability of each unlabeled sample in the sorted unlabeled sample set according to the Boltzmann distribution based on a preset inverse temperature parameter; and filtering the sorted unlabeled sample set based on the discard probability, fusing the filtering results and the representative sample set to obtain the sample set to be labeled.

[0013] Optionally, in a fifth implementation of the first aspect of the present invention, the step of annotating the sample set to be annotated based on a preset defect classification standard to obtain an annotated sample set includes: acquiring the original OLED imaging data corresponding to the sample set to be annotated; integrating the sample set to be annotated and its corresponding original OLED imaging data to obtain data to be annotated by experts; acquiring annotation information based on the data to be annotated by experts, wherein the annotation information is determined based on a preset defect classification standard, and the annotation information is the defect category label of each unannotated sample in the sample set to be annotated; and integrating the annotation information with the sample set to be annotated to obtain the annotated sample set.

[0014] Optionally, in the sixth implementation of the first aspect of the present invention, after the output updated labeled sample set is used as the final labeled set, the method further includes: constructing an OLED defect detection model to be trained, and iteratively training the model to be trained based on the final labeled set to obtain a trained OLED defect detection model; inputting the updated unlabeled sample set into the trained OLED defect detection model to obtain defect detection results corresponding to the updated unlabeled sample set.

[0015] In the technical solution of this invention, a unified process design is used to balance the representativeness and uncertainty of samples. First, a set of central clusters is obtained based on Hamiltonian operator clustering. Second, the degree of quantum entanglement is calculated to screen the representative sample set. The remaining unlabeled sample set is processed by combining virtual time evolution and the Schrödinger equation. The fusion result is the sample set to be labeled. There is no need to distinguish between budget scenarios. The sample selection can be adaptively optimized, improving labeling efficiency and effectively solving the problem of strategy monotony. In addition, samples are mapped to quantum states through self-supervised learning. The clustering accuracy is optimized based on Hamiltonian operator. Quantum entanglement is used to quantize sample association. Uncertainty is captured by virtual time evolution and the Schrödinger equation, which can improve the quality of labeled samples and solve the representation limitations of classical frameworks. This technical solution is suitable for OLED massive sample scenarios, which can reduce the amount of manual labeling and reduce labeling costs, making the accuracy and generalization ability of the OLED defect detection model trained later stronger. Attached Figure Description

[0016] Figure 1 The flowchart illustrates the quantum-inspired active learning-based defect detection sample labeling method provided in this embodiment of the invention. Detailed Implementation

[0017] This invention provides a quantum-inspired active learning-based defect detection sample annotation method. In this invention, the terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion; for example, a process or method comprising a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes or methods.

[0018] This application discloses a quantum-inspired active learning-based defect detection sample annotation method. For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the quantum-inspired active learning-based defect detection sample labeling method in this invention includes:

[0019] 101. Obtain the original sample set for defect detection, encode the original sample set for defect detection into quantum states through self-supervised learning to obtain the initial labeled sample set and the initial unlabeled sample set, and set the iterative hyperparameters;

[0020] In this embodiment, a massive amount of raw defect detection samples generated during the OLED production process are first acquired. These raw defect detection samples include both defective and non-defective samples, and exist in the form of high-resolution imaging data. Subsequently, a self-supervised learning technique is used to encode the quantum states of the raw defect detection samples. Essentially, this maps the high-dimensional imaging features of the raw defect detection samples to low-dimensional quantum states with quantum mechanical properties, thereby achieving a fine characterization of the complex relationships between the samples. Finally, based on the data scale and labeling requirements, key hyperparameters are set to ensure the orderly progress of iterations, and the initialization of the label set is completed.

[0021] 102. Based on the predefined Hamiltonian operator, cluster the initial unlabeled sample set to obtain the set of central clusters and the sample sets of each cluster;

[0022] In this embodiment, the natural division of unlabeled samples is achieved by simulating the central force field aggregation characteristics in quantum mechanics. The predefined Hamiltonian operator is the energy operator in quantum mechanics, and its core function is to quantify the correlation strength between the sample and the cluster center. The clustering process is guided by the defined Hamiltonian operator to gather samples with similar characteristics in the quantum state space into clusters, and finally outputs a stable set of central clusters (the core quantum states of each cluster) and a sample subset corresponding to each cluster, providing a grouping basis for subsequent sample screening.

[0023] 103. Calculate the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set, and perform screening processing on each cluster sample set based on the degree of quantum entanglement to obtain a representative sample set and the remaining unlabeled sample set;

[0024] In this embodiment, the degree of quantum entanglement is a physical quantity in quantum mechanics that characterizes the strength of the correlation between particles. This step uses von Neumann entropy to transform it into a correlation measure between the sample and the central cluster set. The higher the degree of entanglement of the sample, the more representative it is of the characteristics of its cluster and the overall data distribution. Based on the degree of quantum entanglement, samples are selected from each cluster to form a representative sample set covering the core data distribution area. The remaining unselected samples constitute the remaining unlabeled sample set.

[0025] 104. Based on imaginary time evolution and the Schrödinger equation, the remaining unlabeled sample set is screened, and the screening results and representative sample set are merged to obtain the sample set to be labeled;

[0026] In this embodiment, the imaginary time evolution characteristic of the Schrödinger equation can simulate the uncertainty distribution of samples. By substituting the degree of quantum entanglement as the Hamiltonian into the Schrödinger equation, the probability of discarding a sample is derived. The lower the degree of entanglement (i.e., the further away from the central cluster) the sample is, the higher the uncertainty and the greater the probability of being retained. The selected uncertain samples are fused with the representative sample set to form a set of samples to be labeled that has both data coverage and model optimization value.

[0027] 105. Based on the preset defect classification criteria, the sample set to be labeled is labeled to obtain the labeled sample set;

[0028] In this embodiment, the preset defect classification standard is a standardized defect judgment rule formulated for OLED products, including category definitions such as no defects, bright spot defects, dark spot defects, line defects, and surface defects. Technical personnel (experts) with professional knowledge make manual judgments based on the preset defect classification standard and the original OLED imaging data corresponding to the sample set to be labeled, and finally form a labeled sample set containing the correlation between sample quantum states and defect labels.

[0029] 106. Update the initial labeled sample set and the initial unlabeled sample set based on the labeled sample set to obtain the updated labeled sample set and the updated unlabeled sample set. Replace the initial unlabeled sample set with the updated unlabeled sample set. Return to execute the clustering of the initial unlabeled sample set based on the predefined Hamiltonian operator to obtain the set of central clusters and the sample sets of each cluster.

[0030] In this embodiment, the labeled sample set is added to the initial labeled sample set to expand the labeled samples; at the same time, labeled samples are removed from the unlabeled sample set to update the range of the unlabeled sample set; then the updated unlabeled sample set is used as the input for the next iteration, returning to the clustering step to start a new loop, and the quality of the labeled sample set is gradually optimized through multiple iterations; the core logic of this iterative loop is that each round optimizes the clustering and sampling strategy based on the labeling results of the previous round, so that the samples selected later are more in line with the model training requirements.

[0031] 107. Based on the set iteration hyperparameters, determine whether the iteration stopping condition is met. If it is met, output the updated labeled sample set as the final labeled set.

[0032] In this embodiment, the iteration stopping condition is that the current iteration round reaches the total number of iteration rounds. The iteration loop stops when either condition is met, or the unlabeled sample set is empty, meaning all original samples have been labeled. At this point, the updated labeled sample set covers the core features of the data distribution and the key samples required for OLED defect detection model optimization, and can be used as the final labeled set for subsequent training of the OLED defect detection model. The final labeled set includes key information such as sample quantum states, defect labels, and original sample indices, providing complete data support for the training of the OLED defect detection model.

[0033] This invention discloses a quantum-inspired active learning-based defect detection sample labeling method. Through a unified process design, it balances sample representativeness and uncertainty. First, it obtains a set of central clusters based on Hamiltonian operator clustering. Second, it calculates the degree of quantum entanglement to select a representative sample set. The remaining unlabeled sample set is then processed using virtual-time evolution and the Schrödinger equation. The fused results yield the sample set to be labeled. This method eliminates the need to differentiate between budget scenarios, adaptively optimizes sample selection, improves labeling efficiency, and effectively solves the problem of strategy uniformity. Furthermore, it maps samples to quantum states through self-supervised learning, optimizes clustering accuracy based on Hamiltonian operators, quantifies sample associations using quantum entanglement, and captures uncertainty using virtual-time evolution and the Schrödinger equation, thereby improving the quality of labeled samples and overcoming the representational limitations of classical frameworks. This technical solution is suitable for OLED massive sample scenarios, reducing manual labeling and lowering labeling costs, resulting in stronger accuracy and generalization ability of the subsequently trained OLED defect detection model.

[0034] Further, in this embodiment of the invention, the step of obtaining the original sample set for defect detection, encoding the original sample set for defect detection into quantum states through self-supervised learning to obtain an initial labeled sample set and an initial unlabeled sample set, and setting iterative hyperparameters, includes:

[0035] 201. Obtain the original sample set for defect detection, which includes multiple original samples;

[0036] In this embodiment, the original sample set for defect detection comes from the detection equipment on the OLED production line. It includes samples generated by OLED products in different production stages and under different working conditions. It covers various known defect samples (such as bright spots, dark spots, line defects, etc.) as well as defect-free samples. The original samples are stored in digital imaging data format to ensure the integrity and identifiability of sample features.

[0037] 202. Input each original sample into a self-supervised learning function, extract sample features through the simCLR architecture, so as to map each original sample into a quantum state, wherein the quantum state satisfies the normalization condition;

[0038] In this embodiment, the original sample set for defect detection includes N samples, each sample Encode into quantum states through self-supervised learning. ,in Each original sample is input into a self-supervised learning function, which is:

[0039] ;

[0040] Among them, the parameters of the self-supervised learning function Unsupervised pre-training is required for optimization. The pre-training process aims to maximize the contrastive loss of samples, making the quantum states of similar samples closer and the quantum states of different samples more distant; the dimensionality of the quantum state... The dimension can be preset according to the complexity of the sample features, typically taking values ​​of 64, 128, or 256; the selection of the dimension needs to balance the feature representation capability and computational efficiency.

[0041] Quantum state encoding requires the use of the simCLR self-supervised learning architecture, with a tanh activation function added to the last layer of the network to ensure that the mapped quantum state satisfies the normalization condition. The simCLR architecture is a self-supervised learning framework based on contrastive learning, possessing powerful feature extraction capabilities and the ability to automatically learn the intrinsic feature representations of samples. By employing self-supervised learning within the simCLR architecture for quantum state encoding, it eliminates the need for manually labeled pre-training data, reducing reliance on initial annotations. Simultaneously, it automatically learns deep features of samples, resulting in more accurate feature representation and stronger generalization ability compared to traditional manual feature extraction methods. Furthermore, by incorporating a tanh activation function at the final layer of the simCLR architecture, the extracted features undergo nonlinear transformation and normalization, ensuring that the output quantum state satisfies [certain conditions]. , This is for quantum state inner product operations; through normalization, it lays the foundation for subsequent calculation of quantum entanglement degree and Hamiltonian operator clustering.

[0042] 203. Integrate the quantum states of all original samples to obtain a set of quantum states, and initialize the labeled sample set and the unlabeled sample set to obtain an initial labeled sample set and an initial unlabeled sample set, wherein the initial labeled sample set is empty and the initial unlabeled sample set is the set of quantum states;

[0043] In this embodiment, the quantum states of all samples are integrated into a unified set of quantum states to facilitate batch processing of subsequent clustering and screening operations. Initially, since no manual labeling has been performed, the labeled sample set is set to empty, and all combinations of sample quantum states are assigned to the unlabeled sample set, ensuring that the initial sample selection covers the entire data distribution. Let the current iteration round be... ,make For the first The labeled sample set of the round of iterations, let For the first The unlabeled sample set in the round of iteration.

[0044] 204. Set the iteration hyperparameters, which are related to the size of the original sample set for defect detection. The iteration hyperparameters include the total number of iteration rounds, the number of labeled samples per round, the clustering convergence threshold, and the maximum number of clustering iterations.

[0045] In this embodiment, the value of the iteration hyperparameter directly affects the annotation efficiency and the quality of the label set. It needs to be positively correlated with the size of the original defect detection sample set. That is, the larger the original defect detection sample set, the more iterations can be performed. The number of labeled samples per round can be adjusted according to the annotation cost. For example, when the original defect detection sample set includes 10,000 original samples, the number of labeled samples per round can be set. =20, Total number of iterations =10; By designing the correlation between hyperparameters and sample set size, the method has good adaptability and can be adapted to OLED defect detection sample sets of different sizes, thus improving the practicality and versatility of the method.

[0046] The total number of iteration rounds The total number of labeled samples is determined, and the number of labeled samples in each round is also determined. To control the workload of a single round of annotation, the clustering convergence threshold and the maximum number of clustering iterations ensure the stability and efficiency of the clustering process. The clustering convergence threshold is generally set to a value of [value missing]. The maximum number of clustering iterations is typically set to 50 to 200 to prevent the clustering process from getting stuck in an infinite loop.

[0047] Furthermore, in this embodiment of the invention, the clustering of the initial unlabeled sample set based on the predefined Hamiltonian operator to obtain a set of central clusters and sample sets of each cluster includes:

[0048] 301. Based on the current iteration round and the number of labeled samples per round, calculate the number of samples selected corresponding to the current iteration round;

[0049] In this embodiment, the number of samples selected is the first... The number of clusters in each iteration needs to be dynamically adapted to the iteration process, that is, the cluster size in each iteration needs to be dynamically adjusted; The number of clusters in the round of iteration is , For the current iteration round, The number of labeled samples per round; by increasing the number of samples per round. The sample division is gradually refined by using clusters. In the early stage of iteration, due to the small number of labeled samples and clusters, the focus is on capturing the macroscopic distribution of the data. In the later stage of iteration, due to the increase in labeled samples and clusters, the focus is on mining the microscopic features of the data, so that the clustering results are more in line with the progressive needs of model training.

[0050] 302. Randomly select samples from the initial unlabeled sample set to construct an initial cluster center set, wherein the initial cluster center set includes multiple initial cluster centers, and the number of initial cluster centers is consistent with the number of samples selected;

[0051] In this embodiment, a uniform random sampling method is used to select quantum states equal to the number of clusters from the unlabeled sample set as the initial cluster centers. This ensures that the distribution of the initial cluster centers in the quantum state space is random and uniform, avoiding the bias in clustering results caused by the initial cluster centers being concentrated in local areas, and laying a fair initial foundation for subsequent iterative optimization of clustering.

[0052] 303. Based on the predefined Hamiltonian operator and combined with the initial cluster center set, perform iterative optimization clustering on the initial unlabeled sample set;

[0053] In this embodiment, iterative optimization clustering based on predefined Hamiltonian operators organically combines quantum mechanics principles with clustering algorithms. It achieves dynamic allocation of samples and iterative updating of cluster centers through energy expectation values. Compared with traditional clustering algorithms (such as K-Means), it can more accurately capture the complex correlations of samples in quantum state space, and the clustering results are more consistent with the inherent distribution structure of the data. The clustering process requires iterative loops of initial random selection of cluster centers, sample allocation, and cluster center updating until the convergence condition is met.

[0054] Predefined Hamiltonian operators For the first In the first iteration The Hamiltonian operator for each cluster is the key operator for quantifying the energy state of a sample within the cluster. For each unlabeled sample... Calculate fidelity :

[0055] ;

[0056] in, For the first In the first iteration The sample and the first The fidelity of each cluster center is used to measure the similarity between a sample and the cluster center. For the first In the first iteration The initial cluster center quantum state of each cluster, It is a cluster index, with a value range of 1 to 1. ;

[0057] The fidelity and expected energy value Negative correlation:

[0058] ;

[0059] in, For the first In the first iteration The nth sample pair The expected energy value of each cluster is used to determine the cluster affiliation of a sample;

[0060] During the iteration process, each sample is first assigned to the cluster with the smallest expected energy value (i.e., the highest fidelity). Then, the cluster center is updated based on the mean quantum state of all unlabeled samples within the cluster. This process is repeated until the clustering converges. During the sample assignment process, it is necessary to traverse each unlabeled sample and all initial cluster centers to calculate the corresponding expected energy value to ensure the accuracy of the sample assignment.

[0061] make Indicates the first Assigned to clusters in round iteration The sample set, for each cluster Calculate the new cluster center

[0062] ;

[0063] After the cluster center is updated, it needs to be normalized, that is...

[0064] ;

[0065] Ensure that the updated cluster center still satisfies the quantum state normalization condition; in each round of iterative optimization, the fidelity and energy expectation of the unlabeled sample and the cluster center need to be recalculated to achieve dynamic adjustment of the cluster structure.

[0066] 304. Based on the clustering convergence threshold and the maximum number of clustering iterations, determine whether the clustering iteration stopping condition is met. If it is met, output the set of central clusters and the sample set of each cluster.

[0067] In this embodiment, the clustering iteration stops when the mean of the expected energy values ​​of all unlabeled samples is less than a preset clustering convergence threshold, or when the number of iterations reaches a preset maximum number of clustering iterations. When either condition is met, it indicates that the cluster structure has become stable, and further iterations cannot significantly improve the clustering quality. At this point, the final set of central clusters (updated cluster center quantum states) and the sample set corresponding to each cluster are output. The random selection of the initial cluster centers and the control of the clustering convergence threshold avoid bias and instability in the clustering results, ensuring the reliability and efficiency of the clustering process, and providing a high-quality grouping basis for subsequent representative sample selection.

[0068] Further, in this embodiment of the invention, the step of calculating the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set, and then filtering each cluster sample set based on the degree of quantum entanglement to obtain a representative sample set and a remaining unlabeled sample set, includes:

[0069] 401. Obtain the preset sorting criteria, and sort the unlabeled samples in each cluster sample set based on the preset sorting criteria to obtain the sorted sample set of each cluster.

[0070] In this embodiment, the screening priority of each cluster is determined through a sorting operation. The preset sorting criteria are divided into primary criteria and secondary criteria. The primary criterion is "whether the cluster contains labeled samples", and clusters containing labeled samples are sorted first. The secondary criterion is "the number of samples in the cluster". When the primary criteria are the same, clusters with more samples are sorted first. Through the preset sorting criteria, it is ensured that subsequent sample screening focuses on high-value clusters.

[0071] 402. Calculate the degree of quantum entanglement between each unlabeled sample in each cluster sorted sample set and the central cluster set using von Neumann entropy;

[0072] In this embodiment, the degree of quantum entanglement is the core indicator for measuring the representativeness of a sample. The higher the correlation strength, the more representative the sample is of its cluster and the overall data distribution. By calculating the degree of quantum entanglement based on von Neumann entropy, the entanglement characteristics in quantum mechanics can be transformed into a quantitative indicator of sample representativeness. Compared with traditional representativeness measures such as density and distance, it can more accurately capture the complex correlation of samples in quantum state space and make the representativeness judgment more accurate.

[0073] The calculation process requires first constructing each sample density matrix :

[0074] ;

[0075] in, For the first In the round of iteration, the remaining unlabeled sample set is the first... The density matrix of each sample is used to fuse the association information between the sample and all cluster centers; For the first Round iterations use Hamiltonian operators to cluster the set of central clusters. For the first The quantum states of the remaining unlabeled samples, For the central cluster set The Middle The central quantum state of each cluster; based on Constructing the density matrix To ensure the timeliness of related information;

[0076] Then, solve for the density matrix. eigenvalues , density matrix No. 1 eigenvalue, , density matrix The total number of eigenvalues ​​is equal to the dimension of the quantum state. ;

[0077] Will Substitute into the von Neumann entropy formula to calculate the degree of entanglement :

[0078] ;

[0079] in, For the first In the first iteration The degree of entanglement between unlabeled samples and the central cluster is calculated; the degree of entanglement is calculated based on eigenvalues ​​to ensure the rigor and accuracy of the calculation results. The larger the sample size, the more representative the sample.

[0080] 403. Based on the degree of quantum entanglement, a screening operation is performed on the sorted sample sets of each cluster to construct a representative sample set and a remaining unlabeled sample set, wherein the number of unlabeled samples in the representative sample set is less than the number of labeled samples in each round;

[0081] In this embodiment, the most representative samples are selected from the high-priority clusters. The selection operation starts from the first cluster after sorting and selects the samples with the highest degree of quantum entanglement from each cluster in turn, until the number of selected samples is less than the number of labeled samples in each round. The selected samples constitute the representative sample set, and the unselected samples constitute the remaining unlabeled sample set, reserving space for subsequent screening of uncertain samples; only one representative sample is selected for each cluster to ensure that the representative sample set covers different clusters and improves data coverage.

[0082] Furthermore, in this embodiment of the invention, the step of filtering the remaining unlabeled sample set based on imaginary time evolution and the Schrödinger equation, and fusing the filtering results with the representative sample set to obtain the sample set to be labeled, includes:

[0083] 501. Calculate the degree of quantum entanglement between each unlabeled sample in the remaining unlabeled sample set and the central cluster set using von Neumann entropy;

[0084] In this embodiment, the calculation method for the degree of quantum entanglement between each unlabeled sample and the central cluster set in the remaining unlabeled sample set is consistent with step 402, ensuring the consistency and rigor of representativeness and uncertainty determination, making the mathematical basis of dual-objective screening coherent, and avoiding screening result conflicts caused by different measurement methods.

[0085] 502. Based on the degree of quantum entanglement, sort the remaining unlabeled sample set to obtain an unlabeled sorted sample set;

[0086] In this embodiment, the sorting logic is "sorting by the degree of quantum entanglement from smallest to largest". Since the smaller the degree of entanglement, the higher the uncertainty of the sample, that is, the sample deviates from the core distribution, the samples with the smaller degree of entanglement are sorted first, which makes it easier to prioritize the screening of high uncertainty samples in the future and improve the marginal contribution of sample labeling to model optimization.

[0087] 503. Based on the preset inverse temperature parameter, calculate the discard probability of each unlabeled sample in the unlabeled sorted sample set according to the Boltzmann distribution;

[0088] In this embodiment, to simulate the distribution characteristics of sample uncertainty, an imaginary time form of the Schrödinger equation is introduced. Imaginary time evolution can transform the probability distribution of quantum states into a stable distribution of the lowest energy state, which can adapt to the requirement of screening samples with low entanglement (high uncertainty). The imaginary time form of the Schrödinger equation is:

[0089] ;

[0090] in, This is the imaginary time parameter, used to simulate the probability distribution evolution of quantum states; virtual time The quantum state at a given moment; Let Hamiltonian be the degree of quantum entanglement of the samples. This equation is used to quantify the correlation between the energy state of a sample and uncertainty. Its core function is to describe the evolution of the quantum state with imaginary time, and its form determines the trend of subsequent probability amplitude changes.

[0091] The solution to the Schrödinger equation is:

[0092] ;

[0093] in, The initial quantum state, corresponding to the initial quantum state of the sample. It is an exponential operator used to describe the evolution of a quantum state over imaginary time;

[0094] Based on the solution to the Schrödinger equation, the evolved quantum state expression can be obtained, and the relationship between the evolved probability amplitude and the initial probability amplitude is as follows:

[0095] ;

[0096] in, virtual time After evolution, the first The quantum state probability amplitude of each sample; At the initial moment, the first The probability amplitude of each sample; the exponent term is Since the probability amplitude of a quantum state is the core intermediate quantity for measuring the likelihood of a sample being retained, and the rejection probability is positively correlated with the degree of entanglement, meaning that the more representative the sample, the more likely it is to be rejected; therefore, the Hamiltonian is set to be the negative value of the sample's degree of entanglement, so that the exponential term of the evolved probability amplitude is positively correlated with the degree of entanglement and the imaginary time, meaning that the higher the degree of entanglement of a sample, the faster its probability amplitude grows with the evolution of the imaginary time, laying the foundation for the formation of a high rejection probability for highly entangled samples in the future;

[0097] The squared modulus of the probability amplitude represents the initial probability of the sample. However, the differences in probability amplitudes among different samples need to be normalized to form a mathematically sound probability distribution. Normalization is achieved by summing the squared moduli of the probability amplitudes of all samples, ensuring that the sum of the probabilities in the final distribution is 1. After normalization, the Boltzmann distribution is obtained.

[0098] ;

[0099] in, , where is the inverse temperature parameter, and Z is the partition function, Z= This is used to ensure that the sum of the discard probabilities meets the probability distribution requirements; For unlabeled sample set Any sample in;

[0100] To further enhance the sensitivity of the discard probability to the degree of entanglement, an inverse temperature parameter is introduced. , To transform the normalized probability distribution into a Boltzmann distribution; inverse temperature parameter The value of depends on the current iteration round. Dynamic adjustment ( By dynamically adjusting the inverse temperature parameter, this method achieves adaptive discard probability to the iteration process. It allows switching of selection priorities at different budget stages without manual intervention, solving the problem of manual strategy adjustment required in existing technologies and improving the automation and adaptability of the method. In the early stages of iteration, the inverse temperature parameter... The smaller the temperature parameter, the lower the discard probability, and the lower the sensitivity of the entanglement level, thus prioritizing the retention of representative samples; in the later stages of iteration, the inverse temperature parameter... Larger, highly entangled samples (low uncertainty) have a higher probability of being discarded, so uncertain samples should be retained first.

[0101] The discard probability is calculated using the Boltzmann distribution to ensure the rationality and standardization of the probability distribution; the discard probability is positively correlated with the degree of entanglement, therefore, the discard probability of each sample is... The calculation method is as follows:

[0102] ;

[0103] By taking Hamiltonian as And obtained in imaginary time evolution

[0104] ;

[0105] The partition function Z serves as a normalization factor, ensuring that the discard probability remains within the range [0, 1]. The exponential term retains the degree of entanglement and... The correlation between these factors ensures that the discard probability of highly entangled samples is always higher than that of low-entangled samples, ultimately achieving the goal of accurately screening uncertain samples through the evolutionary laws of quantum mechanics.

[0106] 504. Based on the discard probability, perform a filtering operation on the unlabeled sorted sample set, and fuse the filtering results with the representative sample set to obtain the sample set to be labeled;

[0107] In this embodiment, the filtering operation is performed by generating uniformly distributed random numbers. If If the random number is greater than the discard probability, then the unlabeled sample is retained; the selected uncertain samples are merged with the aforementioned representative sample set. If the number of samples after merging is insufficient... Then, randomly supplement the remaining unlabeled samples to ensure that the number of samples to be labeled is exactly the number of samples labeled in each round. The introduction of appropriate randomness through random number determination enhances the diversity of sample selection. The fusion strategy of representative and uncertain samples ensures that the sample set to be labeled has both data coverage and model optimization value. The samples labeled in each round not only consolidate the model's learning of the core distribution, but also make up for the model's lack of understanding of the marginal distribution, which greatly improves the labeling efficiency and model training effect.

[0108] Furthermore, in this embodiment of the invention, the step of labeling the sample set to be labeled based on a preset defect classification standard to obtain a labeled sample set includes:

[0109] 601. Obtain the original OLED imaging data corresponding to the sample set to be labeled, and integrate the sample set to be labeled and its corresponding original OLED imaging data to obtain the data to be labeled by experts;

[0110] In this embodiment, the sample set to be labeled contains the quantum states of unlabeled samples. Expert labeling needs to be based on the physical defect characteristics of the unlabeled samples. Therefore, it is necessary to obtain the original OLED imaging data corresponding to each sample to be labeled by associating it with the sample index. The quantum states and the original imaging data are integrated to form the data to be labeled by experts, ensuring that experts can view the quantum state characteristics and original physical characteristics of the samples at the same time, avoiding labeling errors caused by incomplete information and improving labeling accuracy.

[0111] 602. Obtain annotation information based on the data to be annotated by experts, wherein the annotation information is determined based on a preset defect classification standard, and the annotation information is the defect category label of each unannotated sample in the sample set to be annotated;

[0112] In this embodiment, experts need to possess professional knowledge of OLED defect detection and be familiar with preset defect classification standards. The defect classification standards include the definitions, characteristics, and judgment thresholds of categories such as no defects, bright spot defects, dark spot defects, line defects, and surface defects. By viewing the original OLED imaging data in the data to be labeled by the experts, and combining their own professional knowledge with the preset defect classification standards, they determine a unique defect category label for each sample, thus forming labeling information.

[0113] 603. Integrate the annotation information with the sample set to be annotated to obtain the annotated sample set;

[0114] In this embodiment, the defect category labels in the annotation information are associated and integrated with the corresponding quantum states in the sample set to be labeled. Each sample's quantum state corresponds to a unique defect label, forming a labeled sample set. The standard sample set is the core data for updating the labeled sample set, providing labeled training data for subsequent OLED defect detection model training.

[0115] Furthermore, in this embodiment of the invention, after updating the labeled sample set as the final labeled set, the method further includes:

[0116] 701. Construct an OLED defect detection model to be trained, and iteratively train the model to be trained based on the final label set to obtain a trained OLED defect detection model.

[0117] In this embodiment, the OLED defect detection model to be trained can be a deep learning model suitable for image classification tasks, such as Convolutional Neural Network (CNN), Transformer, ResNet, etc. The structure of the OLED defect detection model needs to be designed according to the complexity of sample features and the requirements for detection accuracy. The training process uses the final labeled set as training data, takes the sample quantum state as input, and the defect category label as output. The model parameters of the OLED defect detection model to be trained are iteratively optimized through the backpropagation algorithm until the detection accuracy of the OLED defect detection model to be trained on the validation set reaches a preset threshold (such as 95%), and the trained model is obtained. The cross-entropy loss function is selected as the loss function in the training process, and the Adam optimizer is selected as the optimizer. The loss value of the model and the accuracy of the validation set need to be recorded during the training process to facilitate the analysis of the model training effect. If the model overfits, regularization techniques can be used for optimization.

[0118] In this embodiment, model training based on the final labeled set makes full use of the key samples selected by active learning. Only a small number of labeled samples are needed to train a high-performance OLED defect detection model, which greatly reduces the labeling cost and solves the problem of difficulty in labeling massive samples in OLED defect detection.

[0119] 702. Input the updated unlabeled sample set into the trained OLED defect detection model to obtain the defect detection results corresponding to the updated unlabeled sample set;

[0120] In this embodiment, the updated unlabeled sample set consists of the remaining unlabeled samples after the iteration terminates, which typically constitute the vast majority of the original defect detection sample set. The quantum states of the updated unlabeled sample set are input into the trained OLED defect detection model. The OLED defect detection model, through the learned mapping relationship between quantum states and defect categories, outputs the defect category prediction result for each sample to complete the automated detection of OLED defects. The defect category prediction result includes information such as sample index, predicted defect category, and prediction confidence. The automated inference process enables rapid detection of massive unlabeled samples, with detection efficiency far exceeding that of manual detection. This meets the large-scale, real-time detection requirements of OLED production lines, improving production efficiency and quality control levels.

[0121] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A quantum-inspired active learning-based defect detection sample annotation method, characterized in that, include: Obtain the original sample set for defect detection, encode the original sample set for defect detection into quantum states through self-supervised learning to obtain an initial labeled sample set and an initial unlabeled sample set, and set the iterative hyperparameters; Based on the predefined Hamiltonian operator, the initial unlabeled sample set is clustered to obtain a set of central clusters and sample sets of each cluster. Calculate the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set. Based on the degree of quantum entanglement, filter each cluster sample set to obtain a representative sample set and the remaining unlabeled sample set. Based on imaginary time evolution and the Schrödinger equation, the remaining unlabeled sample set is screened, and the screening results and representative sample set are merged to obtain the sample set to be labeled. Specifically, the degree of quantum entanglement between each unlabeled sample in the remaining unlabeled sample set and the central cluster set is calculated using von Neumann entropy; Based on the degree of quantum entanglement, the remaining unlabeled sample set is sorted to obtain an unlabeled sorted sample set; based on a preset inverse temperature parameter, the discard probability of each unlabeled sample in the unlabeled sorted sample set is calculated according to the Boltzmann distribution; Based on the discard probability, the unlabeled sorted sample set is filtered, and the filtering results and the representative sample set are combined to obtain the sample set to be labeled. The sample set to be labeled is labeled based on the preset defect classification criteria to obtain the labeled sample set; The initial labeled sample set and the initial unlabeled sample set are updated based on the labeled sample set to obtain the updated labeled sample set and the updated unlabeled sample set. The initial unlabeled sample set is replaced with the updated unlabeled sample set. The process then returns to the execution of the clustering of the initial unlabeled sample set based on the predefined Hamiltonian operator to obtain the set of central clusters and the sample sets of each cluster. Based on the set iteration hyperparameters, determine whether the iteration stopping condition is met. If it is met, output the updated labeled sample set as the final labeled set.

2. The quantum-inspired active learning-based defect detection sample labeling method according to claim 1, characterized in that, The process of obtaining the original sample set for defect detection involves encoding the original sample set into quantum states through self-supervised learning to obtain an initial labeled sample set and an initial unlabeled sample set, and setting iterative hyperparameters, including: Obtain a defect detection raw sample set, which includes multiple raw samples; Each original sample is input into a self-supervised learning function, and sample features are extracted through the simCLR architecture to map each original sample to a quantum state, wherein the quantum state satisfies the normalization condition. Integrate the quantum states of all original samples to obtain a set of quantum states, and initialize the labeled sample set and the unlabeled sample set to obtain an initial labeled sample set and an initial unlabeled sample set. The initial labeled sample set is empty, and the initial unlabeled sample set is the set of quantum states. Set iteration hyperparameters, which are related to the size of the original sample set for defect detection. The iteration hyperparameters include the total number of iteration rounds, the number of labeled samples per round, the clustering convergence threshold, and the maximum number of clustering iterations.

3. The quantum-inspired active learning-based defect detection sample labeling method according to claim 2, characterized in that, The method, based on a predefined Hamiltonian operator, clusters the initial unlabeled sample set to obtain a set of central clusters and sample sets for each cluster, including: Calculate the number of samples selected for the current iteration round based on the current iteration round and the number of labeled samples in each round; A random selection operation is performed on the initial unlabeled sample set to construct an initial cluster center set, which includes multiple initial cluster centers, and the number of initial cluster centers is consistent with the number of samples selected. Based on the predefined Hamiltonian operator and combined with the initial cluster center set, the initial unlabeled sample set is iteratively optimized and clustered. Based on the clustering convergence threshold and the maximum number of clustering iterations, determine whether the clustering iteration stopping condition is met. If it is met, output the set of central clusters and the sample sets of each cluster.

4. The quantum-inspired active learning-based defect detection sample labeling method according to claim 2, characterized in that, The process involves calculating the degree of quantum entanglement between each unlabeled sample in each cluster sample set and the central cluster set, and then filtering each cluster sample set based on the degree of quantum entanglement to obtain a representative sample set and the remaining unlabeled sample set, including: Obtain a preset sorting criterion, and sort the unlabeled samples in each cluster sample set based on the preset sorting criterion to obtain the sorted sample set of each cluster. The degree of quantum entanglement between each unlabeled sample in each cluster sorted sample set and the central cluster set is calculated using von Neumann entropy. Based on the degree of quantum entanglement, a screening operation is performed on the sorted sample sets of each cluster to construct a representative sample set and a remaining unlabeled sample set. The number of unlabeled samples in the representative sample set is less than the number of labeled samples in each round.

5. The quantum-inspired active learning-based defect detection sample labeling method according to claim 1, characterized in that, The annotation process, based on a preset defect classification standard, is performed on the sample set to be annotated to obtain an annotated sample set, including: Obtain the original OLED imaging data corresponding to the sample set to be labeled, and integrate the sample set to be labeled and its corresponding original OLED imaging data to obtain the data to be labeled by experts; Annotation information is obtained based on the data to be annotated by experts. The annotation information is determined based on a preset defect classification standard. The annotation information is the defect category label of each unannotated sample in the sample set to be annotated. The labeled information is integrated with the sample set to be labeled to obtain the labeled sample set.

6. The quantum-inspired active learning-based defect detection sample labeling method according to claim 1, characterized in that, After updating the labeled sample set as the final labeled set, the output also includes: An OLED defect detection model to be trained is constructed, and the model to be trained is iteratively trained based on the final label set to obtain a trained OLED defect detection model. The updated unlabeled sample set is input into the trained OLED defect detection model to obtain the defect detection results corresponding to the updated unlabeled sample set.

Images