PC lens flaw classification and identification method and system

Abstract

The invention relates to the field of optical material defect detection, and discloses a PC lens defect classification and identification method and system, and the method comprises the following steps: asynchronously collecting lens multi-modal data through polarized light imaging, thermal imaging and an acoustic emission sensor, and constructing a five-order asymmetric tensor containing polarization angle, spectrum, time, space and defect features; performing dynamic dimension reduction on the tensor by using an improved Tucker decomposition method, and extracting a core tensor and a factor matrix; establishing a polarization angle-time correlation coding model based on inverse problem solution of a light field state equation, deducing defect distribution characteristics and optimizing a core tensor; mapping the optimized tensor to the QUBO Hamiltonian of a quantum annealing machine, and carrying out optimization solution; and finally, fusing the multi-mode confidence coefficients of polarized light, thermal imaging and acoustic emission signals, generating defect classification labels and inverting geometric and mechanical parameters of the defects. The lens flaw detection efficiency and the classification accuracy are remarkably improved, and the method is suitable for industrial-grade high-precision quality control scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of optical material defect detection, and specifically to a method and system for PC lens defect classification and recognition. Background Art

[0002] During the production and use of PC lenses, the detection of surface and internal defects (such as cracks, bubbles, foreign object embedding, etc.) is a key link to ensure product quality. Traditional detection techniques mostly rely on single-modal sensor data (such as optical imaging or thermal imaging), and it is difficult to comprehensively capture the multi-dimensional physical characteristics of defects. Optical imaging has high sensitivity to surface micro-cracks, but has a weak response to internal defects or deep structure changes; thermal imaging can reflect the abnormal heat conduction inside the material, but is easily interfered by environmental temperature fluctuations, and the inversion accuracy of defect geometric parameters is limited. In addition, existing methods generally adopt simple feature splicing or static weighted fusion strategies when processing multi-modal data, ignoring the spatio-temporal correlation and non-linear coupling relationship between data with different physical characteristics, resulting in the detection results being easily affected by noise, and the classification accuracy significantly decreasing in complex defect scenarios.

[0003] In terms of defect parameter inversion, existing technologies mostly rely on empirical formulas or simplified physical models, and it is difficult to accurately establish the quantitative relationship between defect geometric features (such as depth, curvature) and the distribution of light field and thermal field. Especially for asymmetric defects or composite defects, systematic errors are often introduced due to over-simplification of the model during the inversion process. At the same time, the computational complexity of high-dimensional data is high, and traditional optimization algorithms (such as gradient descent, simulated annealing) are prone to falling into local optimal solutions when dealing with non-convex optimization problems, and cannot meet the dual requirements of real-time and accuracy in industrial scenarios. The above limitations make existing technologies often have problems such as high missed detection rate, low reliability of parameter inversion, and poor environmental adaptability when facing complex defect detection tasks, restricting the further improvement of the PC lens quality control level. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a method and system for PC lens defect classification and recognition, which solves the problems of insufficient multi-modal data fusion, low defect parameter inversion accuracy, and insufficient high-dimensional optimization solution efficiency in the existing PC lens defect detection technology.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for PC lens defect classification and recognition, including the following steps:

[0006] Asynchronously collect multi-source heterogeneous data of the lens, including polarized light imaging data, thermal imaging data, and acoustic emission signals;

[0007] Fuse the multi-source heterogeneous data to construct a fifth-order asymmetric tensor of polarization angle, spectrum, time, space, and defect features;

[0008] Dynamically reduce the dimension of the fifth-order asymmetric tensor using improved Tucker decomposition, and extract the core tensor and the corresponding factor matrices;

[0009] Combined with the extracted core tensor, based on the inverse problem solving method of the light field state equation, establish a polarization angle-time correlation encoding model, deduce the distribution characteristics of defects on and inside the lens surface, calculate the physical parameters of the defect area, and generate an optimized core tensor;

[0010] Map the optimized core tensor to the QUBO Hamiltonian of the quantum annealing machine, and adopt a segmented temperature control strategy to optimize the solution process and adjust the calculation path;

[0011] Combined with the obtained optimization results, fuse the multi-modal confidence levels of polarized light, thermal imaging, and acoustic emission signals, generate defect classification labels, and invert the geometric and mechanical parameters of the defects.

[0012] The present invention also provides a PC lens flaw classification and recognition system, including:

[0013] A data acquisition module for acquiring polarized light images, thermal imaging data, and acoustic emission signals of the object to be measured, and converting them into a multi-modal data matrix;

[0014] A tensor construction module for constructing a fifth-order asymmetric tensor based on the multi-modal data and performing normalization processing on the tensor;

[0015] An improved Tucker decomposition module for dynamically reducing the dimension of the fifth-order asymmetric tensor and extracting the optimized core tensor;

[0016] A polarization angle-time correlation encoding module for establishing a polarization angle-time correlation encoding model based on the inverse problem solving method of the light field state equation, deducing the distribution characteristics of defects, calculating the physical parameters of the defect area, and optimizing the core tensor;

[0017] A quantum optimization solution module for mapping the optimized core tensor to the QUBO Hamiltonian of the quantum annealing machine and adopting a segmented temperature control strategy to optimize the calculation path;

[0018] A multi-modal fusion analysis module for combining the optimization results, fusing the multi-modal confidence levels of polarized light, thermal imaging, and acoustic emission signals, generating defect classification labels, and inverting the geometric and mechanical parameters of the defects;

[0019] A display and storage module for displaying the defect detection results, storing the calculation data, and outputting the optimized defect classification and parameter information.

[0020] The present invention provides a PC lens flaw classification and recognition method and system. It has the following beneficial effects:

[0021] 1. The present invention collects multi-source heterogeneous data of polarized light imaging, thermal imaging and acoustic emission signals, which completely covers surface optical characteristics, internal heat conduction characteristics and mechanical response information, and solves the problem of easy omission or misdetection in single-modal detection. Based on the construction of a fifth-order asymmetric tensor, the unified expression of data with different physical characteristics is realized, and combined with the spatio-temporal alignment and weighted fusion strategy, the coverage range and recognition accuracy of defect detection are significantly improved. For example, micron-level surface scratches and deep bubble defects can be identified simultaneously.

[0022] 2. The present invention performs adaptive rank selection and dynamic dimensionality reduction on high-dimensional tensors through an improved Tucker decomposition method. While retaining the core defect features, the data dimension is reduced, and the interference of redundant information on subsequent analysis is reduced.

[0023] 3. Based on the solution of the inverse problem of the light field state equation and the polarization angle-time correlation coding model, the present invention directly correlates physical characteristics such as the conductivity gradient and stress concentration coefficient of defects with the light field distribution, and realizes the quantitative analysis of defect parameters.

[0024] 4. By mapping the core tensor to the QUBO Hamiltonian of the quantum annealing machine and designing a segmented temperature control strategy, the present invention solves the problem that traditional optimization algorithms are prone to fall into local optima in high-dimensional non-convex problems.

[0025] 5. Based on the adaptive weighting mechanism of multi-modal signal-to-noise ratio, the present invention dynamically adjusts the fusion weights of polarized light, thermal imaging and acoustic emission signals, and effectively suppresses the influence of single-modal data noise or outliers on the overall result. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a schematic flow chart of the method of the present invention;

[0027] Figure 2 is a schematic structural diagram of the system of the present invention.

[0028] Among them, 10. Data acquisition module; 20. Tensor construction module; 30. Improved Tucker decomposition module; 40. Polarization angle-time correlation coding module; 50. Quantum optimization solution module; 60. Multi-modal fusion analysis module; 70. Display and storage module. DETAILED DESCRIPTION OF THE INVENTION

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work shall fall within the protection scope of the present invention.

[0030] Please refer to the attached Figure 1 , the present invention provides a method for classifying and identifying PC lens defects. Aiming at defects such as surface scratches, bubbles, cracks, and foreign objects that may occur during the manufacturing process of PC lenses, through the multi-modal data fusion of polarized light imaging, thermal imaging, and acoustic emission signals, combined with a quantum optimization algorithm and a polarization angle-time correlation coding model, high-precision defect classification and parameter inversion are achieved.

[0031] As Figure 1 shown, the method for classifying and identifying PC lens defects may include the following steps:

[0032] S1. Asynchronously collect polarized light imaging, thermal imaging, and acoustic emission signal data to obtain defect information on the surface and inside of the lens;

[0033] S2. Perform fusion on multi-source heterogeneous data to construct a fifth-order asymmetric tensor containing polarization angle, spectrum, time, space, and defect features;

[0034] S3. Use an improved Tucker decomposition method to perform dynamic dimensionality reduction on the fifth-order asymmetric tensor, and extract the core tensor and factor matrices;

[0035] S4. Based on the inverse problem solving method of the light field state equation, combined with the core tensor, establish a polarization angle-time correlation coding model, deduce the distribution characteristics of lens defects, and calculate physical parameters;

[0036] S5. Map the optimized core tensor to the QUBO Hamiltonian of the quantum annealing machine, and adopt a segmented temperature control strategy to optimize the solution process;

[0037] S6. Fusion the multi-modal confidence levels of the optimized polarized light, thermal imaging, and acoustic emission signals, generate defect classification labels, and invert the geometric and mechanical parameters of the defects.

[0038] The following is a detailed description of each step in the method of the present invention, comprehensively elaborating on the specific implementation principles, technical details, and processes for each step.

[0039] For step S1, in this embodiment, aiming at the defect detection requirements of PC lenses, multi-modal sensing technology is used to asynchronously collect data with different physical characteristics in order to obtain complete information on surface and internal defects. The core of this step is to use three data sources of polarized light imaging, thermal imaging, and acoustic emission signals to respectively obtain the optical, thermal, and mechanical response information of the lens, thereby providing multi-modal data support for subsequent defect classification and defect parameter inversion.

[0040] During the data acquisition process, the system adopts an asynchronous acquisition strategy, that is, different sensors perform data acquisition in their respective independent time domains, and the acquired data is calibrated through timestamps. Since the working mechanisms and data sampling rates of different sensors are different, the asynchronous acquisition method can avoid the problem of multimodal data mismatch caused by timing errors in the synchronous acquisition process, and at the same time improve the adaptability and real-time performance of the system.

[0041] To ensure the accuracy of data fusion, this embodiment requires that the polarized light imaging data, thermal imaging data, and acoustic emission signals satisfy the time synchronization constraint conditions, that is:

[0042]

[0043] Among them,

[0044] Δt k is the maximum time deviation of each modal data at the k-th sampling;

[0045] τ comp (θ) is the time compensation function corresponding to the polarization angle θ;

[0046] is the derivative of the time compensation function with respect to the polarization angle;

[0047] T s is the minimum sampling period;

[0048] ε is the synchronization deviation control threshold.

[0049] The physical meaning of this constraint is that for the imaging processes with different polarization angles, there may be time offsets, so it is necessary to correct the acquired data through the time compensation function τ comp (θ) to ensure that all modal data satisfy strict time synchronization during fusion. Specifically, the implementation methods of the time synchronization constraint conditions include:

[0050] Adopt a high-precision timestamp synchronization mechanism to attach timestamps to the sampled data of different sensors, and use the hardware clock to uniformly calibrate the data;

[0051] In the data post-processing stage, correct the time error of the polarized light imaging data through the time compensation function τ comp (θ) to make it satisfy the synchronization constraint;

[0052] Through the constraint of the minimum sampling period T s to ensure that the matching accuracy of different modal data in the time domain does not exceed the set deviation threshold ε.

[0053] In this embodiment, the main function of polarized light imaging is to identify surface defects of the lens, such as cracks, scratches, and stress concentration areas, by using the polarization characteristics of light. Specifically, a polarized light microscopy imaging device is used to collect optical images at different polarization angles (e.g., 0°, 45°, 90°), and the polarization light transmittance is calculated through Malus' law. The mathematical expression of Malus' law is as follows:

[0054] I = I0cos 2 (θ)

[0055] where I is the light intensity after passing through the polarizer, I0 is the incident light intensity, and θ is the angle of the polarizer. Based on the above relationship, a mapping relationship between the polarization angle and the light transmittance can be constructed, and the optical contrast change in different regions of the lens surface can be extracted by combining image processing algorithms to enhance the visibility of defects such as cracks and scratches.

[0056] In addition, in order to detect possible bubbles, delaminations, or material defects inside the lens, this embodiment uses thermal imaging technology for non-contact detection. By applying a short-term thermal excitation (such as pulsed laser or resistive heating) to the lens, the temperature of different regions inside the lens can undergo transient changes, and an infrared thermal imager is used to record the change of its temperature distribution over time. The basic principle of thermal imaging can be described by the heat conduction equation:

[0057]

[0058] where T is the temperature, t is the time, α is the thermal diffusivity, is the Laplace operator. Since there are differences in the thermal diffusivity of different materials or defect regions, by analyzing the temperature change curve, the defect regions inside the lens can be identified, and the type and depth of the defects can be further inferred.

[0059] In the acoustic emission signal acquisition part, this embodiment uses a high-frequency acoustic emission sensor to record the acoustic emission signals generated by the lens under mechanical loading. These signals are usually excited by the expansion of defects, interfacial friction, or stress release, and have strong time series characteristics. The feature extraction of acoustic emission signals can adopt the wavelet transform method. By decomposing the signal spectrum, the characteristic frequencies related to defects are extracted. The wavelet transform can be expressed as:

[0060]

[0061] where x(t) is the acoustic emission signal, ψ is the wavelet basis function, and a and b represent the scale and time offset respectively. By calculating the energy distribution of the acoustic emission signal at different scales, different types of defects can be effectively distinguished, and characteristic information related to the defect size and morphology can be provided.

[0062] In this embodiment, all the collected data are time-synchronized by a high-precision clock, and an interpolation algorithm is used to align the data in time sequence for subsequent data fusion processing. In addition, the data acquisition process is affected by external noise, so a multi-stage filtering technique (such as Kalman filtering or adaptive noise filtering) is used to denoise the original data and improve the data quality.

[0063] In summary, this step realizes the acquisition of multi-modal information on the surface and inside of the PC lens through the asynchronous acquisition of polarized light imaging, thermal imaging, and acoustic emission signals, and uses a time compensation function and time synchronization constraint conditions to ensure the time sequence matching of the data, providing high-quality original data for subsequent defect classification and defect parameter inversion.

[0064] For step S2, in this embodiment, after obtaining the polarized light imaging data, thermal imaging data, and acoustic emission signals, it is necessary to fuse these heterogeneous data to construct a fifth-order asymmetric tensor containing multi-dimensional information. Since there are significant differences in the physical characteristics, data dimensions, and time resolutions of different modal data, steps such as data preprocessing, alignment, and mapping are required to ensure that the data can be uniformly expressed and can accurately describe the defect characteristics of the PC lens.

[0065] First, before data fusion, it is necessary to standardize the formats of different data sources. The polarized light imaging data is usually stored in the form of a grayscale image or a color image, and the polarization angle corresponding to each pixel point can be used as additional channel information; the thermal imaging data is expressed in the form of a temperature matrix and usually contains dynamic information about the temperature change over time; the acoustic emission signal is time series data and needs to be transformed into a time-frequency diagram through time-frequency transformation to be consistent with the other two types of data in the spatial dimension.

[0066] To ensure the spatial alignment of the data, this embodiment uses an interpolation method to normalize the data of all modalities to have the same resolution. At the same time, since the sampling times of different data sources may be different, it is necessary to align the data based on the time compensation function in step S1 to ensure that the sampling times of each modal data are within the allowable synchronization error range.

[0067] After the data format is standardized, a fifth-order asymmetric tensor can be formally constructed. Let be the constructed tensor, where:

[0068] P represents the polarization angle dimension, which depends on the acquisition angle of the polarized light imaging data;

[0069] S represents the number of spectral channels, which is usually determined by the wavelength distribution of the polarized light and thermal imaging data;

[0070] T represents the time step, which is used to record the change of data in the time domain;

[0071] L represents the spatial resolution, corresponding to the number of pixels.

[0072] D represents the defect feature dimension, including features from polarized light, thermal imaging, and acoustic emission data, such as contrast, temperature gradient, acoustic emission frequency, etc.

[0073] The core of constructing this tensor lies in the fusion strategy of different modal data. In this embodiment, a method combining feature concatenation and weighted fusion is adopted to avoid a single modality having an excessive impact on the overall feature distribution while ensuring data integrity. Specifically, the features of polarized light, thermal imaging, and acoustic emission signals are fused in the following way:

[0074]

[0075] Among them,

[0076] represents the elements of the constructed tensor, covering spatial, temporal, spectral, and defect feature information;

[0077] is the feature matrix of polarized light data;

[0078] F t (i, j, t, λ) is the feature matrix of thermal imaging data;

[0079] F s (i, j, t, λ) is the feature matrix of acoustic emission data;

[0080] w p 、w t 、w s are the weighted coefficients of the three modalities respectively, and are adaptively adjusted by the signal-to-noise ratio.

[0081] The physical meaning of the above formula is that when constructing the tensor, the contribution degrees of different data sources are considered, and a more reliable data is given a higher weight through an adaptive weighting mechanism. For example, when the signal-to-noise ratio of polarized light imaging is high, w p will increase accordingly to enhance the contribution of this modality, while for a modality with a large amount of noise, its weight will be reduced to reduce the interference to the overall data.

[0082] In this embodiment, in order to further reduce the problem of feature-level mismatch of different modal data, a feature mapping strategy based on the kernel method is also introduced. Specifically, the Gaussian kernel function is used to map polarized light, thermal imaging, and acoustic emission signals to a high-dimensional space respectively to reduce the non-linear differences between data. The Gaussian kernel function is defined as follows:

[0083]

[0084] Among them, x and y respectively represent feature vectors of different modalities, and σ is the kernel width parameter. Through kernel mapping, data of different modalities can have stronger comparability in the feature space, thereby improving the effectiveness of data fusion.

[0085] After data fusion is completed, the constructed fifth-order tensor is used as the input for subsequent step S3 for Tucker decomposition and feature dimensionality reduction. In summary, in this step, through operations such as data preprocessing, time synchronization, weighted fusion, and feature mapping, a fifth-order asymmetric tensor is constructed, ensuring the unified expression of data of different modalities and laying a foundation for subsequent defect classification and parameter inversion.

[0086] For step S3, in this embodiment, for the constructed fifth-order asymmetric tensor it is dynamically dimensionally reduced by an improved Tucker decomposition method. Tucker decomposition is a tensor decomposition technique widely used in high-dimensional data processing, which can decompose a high-dimensional tensor into the product of a core tensor and multiple factor matrices. Specifically, in this embodiment, by means of step-by-step modal decomposition, core features are extracted from the high-dimensional space to reduce redundant information, thereby realizing the dimensionality reduction of the tensor.

[0087] First, the fifth-order asymmetric tensor is preprocessed by normalization. The purpose of normalization is to unify the scale differences between different data dimensions, ensure the balanced contribution of data of different modalities, and avoid a certain modality having an excessive impact on the overall dimensionality reduction result due to a large value. This preprocessing step enables the data on each dimension to be processed at the same scale, thereby improving the stability and convergence speed of the decomposition process.

[0088] Next, a step-by-step rank adaptive selection method based on nuclear norm constraint is used to determine the target ranks R1, R2, R3, R4, R5 of the core tensor, and these target ranks respectively correspond to the polarization angle, spectrum, time, space, and defect feature dimensions. Specifically, the target ranks are adaptively selected through the following optimization objective:

[0089]

[0090] Among them, represents the nuclear norm of the core tensor , is the factor matrix of the i-th dimension, ||U (i) || 2,1is the 2,1-norm of the factor matrix for the i-th dimension, representing the sparsity constraint on the factor matrix of this dimension. By regularizing each factor matrix, the sparsity of the factor matrix during the dimensionality reduction process can be effectively controlled, avoiding overfitting or unnecessary computational complexity. α is the regularization parameter, used to balance the relationship between the nuclear norm and the sparsity constraint.

[0091] On this basis, the normalized fifth-order tensor is gradually decomposed in modes, and the factor matrices U (1) , U (2) , U (3) , U (4) , U (5) of each mode are calculated, and the core tensor is updated. The factor matrix U (i) represents the eigenvector space in mode i. Specifically, it describes the principal components of different modal data such as polarized light imaging, thermal imaging, and acoustic emission signals in each dimension. The dimension I i ×R i of each factor matrix represents the mapping relationship between mode i in the original data space and the reduced-dimensional space.

[0092] In each iteration process, the model is optimized by calculating the residual tensor . Here, × i represents the mode product operation of tensors, calculating the product of each factor matrix and the core tensor to obtain the reconstructed tensor. The residual tensor ε represents the difference between the original tensor and the reconstructed tensor. To ensure the accuracy of the dimensionality reduction result, the convergence determination needs to be carried out according to the following conditions:

[0093]

[0094] where, ||ε|| F is the Frobenius norm of the residual tensor ε, is the Frobenius norm of the original tensor , and δ is the convergence threshold, representing the maximum allowable error of the residual. When the norm of the residual tensor satisfies the convergence condition, it means that the decomposition process has been completed, and at this time, the core tensor and the factor matrices U (i) of each mode are output.

[0095] The step-by-step mode decomposition in this embodiment not only effectively extracts the main features of each mode, but also makes the decomposition result adaptively adapt to the characteristics of the data by gradually adjusting the target rank, avoiding the occurrence of overfitting. The reduced-dimensional tensor is optimized in both storage and calculation, reducing the computational complexity while retaining sufficient information for subsequent analysis.

[0096] In summary, in this embodiment, the dynamic dimensionality reduction process of the fifth-order asymmetric tensor is realized through an improved Tucker decomposition method. This method can adaptively determine the target rank of the core tensor, and extract the core features of multi-modal data through step-by-step modal decomposition, providing an efficient feature representation for subsequent defect detection, classification, and parameter inversion.

[0097] For step S4, in this embodiment, for the extracted core tensor, an inverse problem solving method of the light field state equation is adopted to establish a polarization angle-time correlation encoding model. In this process, first, based on the extracted core tensor Construct the light field state equation:

[0098]

[0099] where E represents the electric field strength, H represents the magnetic field strength, μ is the magnetic permeability, and σ(x) represents the local conductivity distribution related to the defect. The light field state equation can describe the propagation behavior of the light field in a material with defects, especially its close relationship with the conductivity distribution of the defects. In this embodiment, this equation can be effectively used to deduce the light field distribution related to the defects, providing a basis for subsequent defect feature extraction.

[0100] Next, the finite element-boundary element hybrid method is used to discretize the light field state equation, and then a numerical solution model containing the variables of polarization angle θ and time t is established. To ensure the solution accuracy, an objective function based on the least square error optimization is constructed:

[0101] min||E - E obs ||2

[0102] where E obs represents the actually measured electric field distribution. Through this objective function, the error between the simulated light field and the actual light field can be minimized, thereby improving the fitting accuracy of the model. Further, through this model, the light field distribution at different polarization angles θ and times t can be calculated, providing support for subsequent defect region feature analysis.

[0103] To control the smoothness of the model and avoid overfitting or irregular solutions, the Tikhonov regularization method is used for optimization. Its regularization form is:

[0104]

[0105] where λ is the regularization parameter, used to adjust the smoothness of the light field solution, represents the gradient of the conductivity distribution. Through this regularization method, the smoothness of the light field distribution in the solution space can be ensured, making it more in line with physical reality, and effectively avoiding the instability of the model.

[0106] Based on the calculated light field information, the spatial distribution characteristics of the defects on and inside the lens surface are further deduced. Specifically, using the polarization imaging data and the optimized light field solution, the conductivity gradient, stress concentration coefficient K t , boundary curvature κ, and defect depth h and other important physical parameters of the defect area can be calculated. These parameters provide a reliable basis for subsequent defect diagnosis and lens quality assessment.

[0107] Meanwhile, in order to improve the resolution of the defect area, an adaptive mesh refinement technique is introduced in this embodiment. When calculating the numerical solution E of the light field state equation, by calculating its gradient the preliminary distribution of the defect area is determined. Based on the gradient change rate a grid refinement criterion is set, and the adaptive mesh refinement method is used to perform refined calculations on the defect area. Specifically, the grid size should satisfy the error constraint condition:

[0108]

[0109] where represents the maximum value of the light field gradient, and ε g is the grid refinement error tolerance. Through this method, higher-precision calculations can be achieved in the defect area to ensure the accurate extraction of defect characteristics.

[0110] After the calculation of the defect area is completed, the variational level set method is used to reconstruct the defect boundary. The variational level set method optimizes the defect boundary by minimizing the following level set energy function:

[0111]

[0112] where φ is the level set function, H(φ) is the Heaviside function, and α is the regularization parameter. By iteratively optimizing this energy function, the level set function φ can be adjusted so that the defect boundary gradually converges to the optimal position. In this way, the boundary of the defect can more accurately match the actual defect area.

[0113] Furthermore, the boundary curvature κ and defect depth h of the defect area are calculated. These information play an important role in the quantitative analysis of the defect and the subsequent derivation of physical parameters. Through the optimized defect area information, the core tensor can be updated and an optimized core tensor is generated to enhance the data quality of the defect feature dimension.

[0114] In summary, in this embodiment, by combining the inverse problem solving method of the optical field state equation and the adaptive grid meshing technology, the establishment of the polarization angle-time correlation coding model is realized. This model can not only accurately calculate the optical field distribution, but also effectively deduce the physical characteristics of defects, providing data support for subsequent defect diagnosis and quality assessment.

[0115] For step S5, in this embodiment, after optimizing the core tensor, it is mapped to the Hamiltonian of the QUBO (Quadratic Unconstrained Binary Optimization) problem of the quantum annealing machine to achieve the optimal solution of defect classification and physical parameter derivation. In this process, first, according to the optimized core tensor Construct the objective function of the QUBO problem and define the Hamiltonian H as follows:

[0116]

[0117] where s i ∈{0,1} is a binary decision variable representing the state of the qubit; Q ii is the first-order weight term corresponding to the variable s i , representing the energy weight of the independent variable; Q ij is the second-order interaction coupling weight term between the variables s i and s j , reflecting the mutual relationship between the two decision variables.

[0118] Through the modal decomposition of the core tensor , the coefficient matrix Q = {Q ii , Q ij} in the objective function can be calculated. Among them, Q ii is set according to the energy distribution weight of the defect area, while Q ij is calculated according to the spatial correlation of the defect characteristics. Specifically, the energy distribution weight of the defect area reflects the influence of the defect on the propagation of the optical field, and the spatial correlation considers the interaction effect of the defect at different positions. These coefficient matrices will be mapped into the quantum annealing machine as inputs to describe the energy function of the optimization problem.

[0119] After mapping to the quantum annealing machine, the optimal solution is executed according to the segmented temperature control strategy. The segmented temperature control strategy is divided into two stages:

[0120] The first stage: Use the exponential cooling strategy for global search. Specifically, the annealing temperature changes with time t and cools down according to the following formula:

[0121] T(t) = T0e -αt

[0122] Among them, T(t) is the current annealing temperature, T0 is the initial temperature, α is the cooling coefficient, and t is the time step. Through this cooling strategy, the temperature gradually decreases from a relatively high initial value, helping the system explore the entire solution space and avoid falling into local optimal solutions.

[0123] The second stage: Adopt a periodic oscillation cooling strategy for local optimization. The specific formula is as follows:

[0124] T(t) = T0cos(β(t - τ))

[0125] Among them, T′0 is the adjusted initial temperature, β is the oscillation frequency, and τ is the stage conversion time. The oscillation strategy in this stage helps to finely adjust the optimization of the solution, focus on approaching the local optimal solution, and thus improve the solution accuracy.

[0126] During the annealing process, continuously monitor the convergence of the solution, and judge whether it enters the optimal state based on the change rate of the Hamiltonian ΔH / Δt. When the following convergence condition is met, terminate the calculation:

[0127] |ΔH / Δt| ≤ ∈ H

[0128] Among them, ΔH is the change in the Hamiltonian between two consecutive time steps, and ∈ H is the set convergence threshold. Through this convergence criterion, it can be ensured that during the optimization process of the solution, the change in the Hamiltonian tends to be stable, indicating that the system has found the optimal solution.

[0129] When the convergence condition is met, extract the optimal solution in the quantum annealing solution process. According to the decoding strategy of binary variables, convert the optimal solution into a defect classification result, and further restore the geometric and mechanical parameters of the defect. These parameters include but are not limited to the shape, size, distribution, stress concentration coefficient, and defect depth of the defect, etc.

[0130] Finally, by completing the optimization solution process of the QUBO Hamiltonian, the defect classification results and their physical characteristics are obtained. These results provide sufficient information support for subsequent defect diagnosis and quality assessment, and can effectively help improve the quality control and detection accuracy in the production process.

[0131] In summary, in this embodiment, by mapping the optimized core tensor to the QUBO Hamiltonian of the quantum annealer and combining the segmented temperature control strategy for optimization solution, the efficiency and accuracy of defect classification and physical parameter derivation are effectively improved. Through the introduction of quantum annealing technology, the optimization process can quickly converge in a more complex high-dimensional space to obtain a reliable optimization solution, thereby providing more accurate numerical support for defect analysis.

[0132] For step S6, in this embodiment, after the calculation of the optimized core tensor , combined with the characteristic information of polarized light, thermal imaging, and acoustic emission signals, defect classification labels are generated through a multi-modal confidence fusion method, and the geometric and mechanical parameters of the defects are inverted. The specific process includes the following steps:

[0133] First, extract the defect characteristic information in the optimized core tensor . Through the modal decomposition of the core tensor, the polarized light image feature matrix F p , the thermal imaging feature matrix F t , and the acoustic emission signal feature matrix F s are obtained. These feature matrices respectively contain the defect-related characteristic information in the polarized light imaging, thermal imaging, and acoustic emission signal modalities, serving as the basis for subsequent defect classification and inversion.

[0134] After obtaining the feature matrices of each modality, calculate the confidence score of each modality data. The confidence score is used to measure the contribution degree of each modality data to defect classification. Specifically, for the confidence scores C p , Ct, and Cs of each modality, the calculation formulas are as follows:

[0135]

[0136] where |F(i,j)| represents the absolute value of the eigenvalue of the i,j-th pixel point. These calculation formulas reflect the contribution ratio of different modalities to the characteristics of the defect area. Specifically, C p represents the contribution of the polarized light modality relative to other modalities in the entire three modalities. Similarly, Ct and Cs reflect the roles of thermal imaging and acoustic emission signals in classification.

[0137] Next, based on the calculated confidence scores C p , Ct, and Cs, construct a fusion model and calculate the multi-modal comprehensive confidence C fusion . This fusion confidence is calculated through the following weighted formula:

[0138] C fusion = w p C p + w t C t + w s C s

[0139] where w p , w t , and w sThey represent the weighting coefficients of polarized light, thermal imaging, and acoustic emission signals respectively. These weighting coefficients reflect the importance of each modality in defect classification and can be dynamically adjusted according to experimental data to improve the classification accuracy.

[0140] Based on the calculated comprehensive confidence level C fusion , different classification thresholds are set to determine the defect classification labels. The specific classification rules are as follows:

[0141] If C fusion ≥τ1, it is marked as a surface crack;

[0142] If τ2≤C fusion <τ1, it is marked as a deep defect;

[0143] If C fusion <τ2, it is marked as a defect-free area.

[0144] Among them, τ1 and τ2 are classification thresholds, which are determined by experimental data and used to distinguish different types of defects.

[0145] After completing the defect classification, the geometric parameters of the defect area are inversely calculated according to the classification results. This process calculates the depth h, boundary curvature κ, and stress concentration factor K of the defect through reverse derivation t . In addition, the physical mechanics model is combined to further calculate the material properties of the defect. These inversion results provide important geometric and mechanical parameters for defect analysis, helping to further evaluate the nature of the defect and its impact on the overall structure.

[0146] Finally, by fusing the multi-modal data of polarized light, thermal imaging, and acoustic emission signals, and combining the confidence score and the inversion process, this embodiment can achieve efficient and accurate defect classification and parameter inversion. This method not only improves the accuracy of defect recognition but also provides support for in-depth defect analysis through quantified geometric and mechanical features.

[0147] Generally speaking, the present invention asynchronously collects the surface and internal defect information of the lens through multi-modal sensors of polarized light imaging, thermal imaging, and acoustic emission signals, constructs a fifth-order asymmetric tensor containing polarization angle, spectrum, time, space, and defect features; uses improved Tucker decomposition to dynamically reduce the dimension of the tensor to extract core features, combines the inverse problem solution of the light field state equation to establish a polarization angle-time correlation coding model, deduces the defect distribution characteristics and optimizes the core tensor; maps the optimized tensor to the QUBO Hamiltonian of the quantum annealing machine, and realizes the optimization solution of defect classification through a segmented temperature control strategy; finally, fuses the multi-modal confidence to generate classification labels, and inversely calculates the defect depth, stress concentration factor, and boundary curvature parameters to achieve high-precision and high-efficiency lens defect detection and quantitative analysis.

[0148] The PC lens defect classification and recognition system described below can be correspondingly referred to the PC lens defect classification and recognition method described above.

[0149] Please refer to the appendix Figure 2 , the present invention also provides a PC lens defect classification and recognition system, including:

[0150] A data acquisition module 10, configured to obtain the polarized light image, thermal imaging data, and acoustic emission signal of the object to be measured, and convert them into a multi-modal data matrix;

[0151] A tensor construction module 20, configured to construct a fifth-order asymmetric tensor based on the multi-modal data, and perform normalization processing on the tensor;

[0152] An improved Tucker decomposition module 30, configured to perform dynamic dimensionality reduction on the fifth-order asymmetric tensor, and extract an optimized core tensor;

[0153] A polarization angle-time correlation encoding module 40, configured to establish a polarization angle-time correlation encoding model based on the inverse problem solving method of the light field state equation, deduce the distribution characteristics of defects, calculate the physical parameters of the defect area, and optimize the core tensor;

[0154] A quantum optimization solving module 50, configured to map the optimized core tensor to the QUBO Hamiltonian of the quantum annealing machine, and adopt a segmented temperature control strategy to optimize the calculation path;

[0155] A multi-modal fusion analysis module 60, configured to combine the optimization results, fuse the multi-modal confidence of polarized light, thermal imaging, and acoustic emission signals, generate defect classification labels, and invert the geometric and mechanical parameters of the defects;

[0156] A display and storage module 70, configured to display the defect detection results, store the calculation data, and output the optimized defect classification and parameter information.

[0157] The system of this embodiment can be used to execute the above method embodiment, and its principle and technical effect are similar, which will not be elaborated here.

[0158] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A PC lens defect classification and identification method, characterized in that: The following steps are involved: Asynchronously collect multi-source heterogeneous data of the lens, including polarized light imaging data, thermal imaging data and acoustic emission signals; Fuse multi-source heterogeneous data to construct a fifth-order asymmetric tensor of polarization angle, spectrum, time, space and defect characteristics; Using improved Tucker decomposition to dynamically reduce the dimension of the fifth-order asymmetric tensor, extracting the core tensor and the corresponding factor matrix; Combined with the extracted core tensor, based on the inverse problem solving method of the light field state equation, a polarization angle-time correlation coding model is established to deduce the distribution characteristics of surface and internal defects of the lens, and the physical parameters of the defect area are calculated to generate the optimized core tensor; The optimized core tensor is mapped to the QUBO Hamiltonian of the quantum annealer, and a segmented temperature control strategy is used to optimize the solution process and adjust the calculation path; Combining the obtained optimization results, the multimodal confidence of polarized light, thermal imaging and acoustic emission signals is fused to generate defect classification labels and invert the geometric and mechanical parameters of the defects.

2. The PC lens defect classification and identification method according to claim 1, characterized in that: The polarized light imaging data, thermal imaging data and acoustic emission signals meet the time synchronization constraint condition: Among them, Δt k is the maximum time deviation of each modal data at the kth sampling, τ comp (θ) is the time compensation function corresponding to the polarization angle θ, is the derivative of the time compensation function with respect to the polarization angle, T s is the minimum sampling period, and ε is the synchronization deviation control threshold.

3. The PC lens defect classification and identification method according to claim 1, characterized in that: The step of fusing multi-source heterogeneous data to construct a fifth-order asymmetric tensor of polarization angle, spectrum, time, space and defect characteristics includes: Normalize the collected polarization imaging data, thermal imaging data and acoustic emission signals; According to the polarization angle information, the polarized light imaging data is mapped to the first dimension, and the main features are extracted by spectral decomposition, and the spectral information is mapped to the second dimension; According to the sampling timestamp, the data of each modality are aligned, and the time dimension is constructed in chronological order to form the third dimension of the tensor; According to the spatial coordinates, the pixel position of the imaging data is matched with the spatial source positioning information of the acoustic emission signal to form the fourth dimension of the tensor; Combined with the characteristic parameters of the defect area, including shape, size, and boundary gradient change rate, the defect feature dimension is established to form the fifth dimension of the tensor, completing the construction of the fifth-order asymmetric tensor.

4. The PC lens defect classification and identification method according to claim 1, characterized in that: The step of dynamically reducing the dimension of the fifth-order asymmetric tensor by using the improved Tucker decomposition to extract the core tensor and the corresponding factor matrix includes: The fifth-order asymmetric tensor constructed Normalization preprocessing is performed, where I1, I2, I3, I4, and I5 represent the dimensions of the tensor in polarization angle, spectrum, time, space, and defect feature dimensions, respectively; The core tensor is determined by using a stepwise rank adaptive selection method based on nuclear norm constraints. The target rank R1, R2, R3, R4, R5 satisfies: in, The core tensor The nuclear norm of is the factor matrix of the i-th dimension, α is the regularization parameter, which represents the control of the sparse constraint on the factor matrix; Perform step-by-step modal decomposition on the normalized fifth-order tensor and calculate the factor matrix U (1) ,U (2) ,U (3) ,U (4) ,U (5) And update the core tensor Among them U (i) is the factor matrix of the i-th dimension, representing the eigenvector space on mode i; Compute the residual tensor at each iteration And adjust the target rank to meet the residual convergence condition in represents the Frobenius norm, δ is the convergence threshold, represents the allowable error of the residual, is the original fifth-order asymmetric tensor, and ε is the calculated residual tensor; After the convergence conditions are met, the core tensor is output And the factor matrix U of each mode (i) , completing the dynamic dimensionality reduction process.

5. The PC lens defect classification and identification method according to claim 1, characterized in that: The step of combining the extracted core tensor and establishing a polarization angle-time correlation coding model based on an inverse problem solving method of a light field state equation comprises: Based on the extracted core tensor Constructing the light field state equation Where E is the electric field strength, is the magnetic field intensity, μ is the magnetic permeability, and σ(x) is the local conductivity distribution associated with the defect; The finite element-boundary element hybrid method is used to discretize the light field state equation, a numerical solution model including polarization angle θ and time t variables is established, and the objective function min||EE based on least squares error optimization is constructed. obs ||2, where E obs is the actual measured electric field distribution; Using Tikhonov regularization method Optimize, λ is the regularization parameter, which is used to control the smoothness of polarization angle-time correlation coding, and calculate the light field distribution corresponding to the polarization angle θ at different times t; Based on the calculated light field information, the spatial distribution characteristics of the surface and internal defects of the lens are derived, and the physical parameters of the defect area are calculated, including the conductivity gradient and stress concentration factor K of the defect area. t , boundary curvature κ and defect depth h; Combine the calculation results and update the core tensor And generate the optimized core tensor to complete the establishment of the polarization angle-time correlation coding model.

6. The PC lens defect classification and identification method according to claim 5, characterized in that: In the process of solving the inverse problem of the light field state equation, the adaptive meshing technology is used to perform refinement calculation on the defect area, and the defect boundary is reconstructed in combination with the variational level set method. The specific steps include: Calculate the numerical solution E of the light field state equation and obtain its gradient Determine the preliminary distribution of defect areas; According to the light field gradient change rate Set the mesh refinement standard and use the adaptive meshing method to refine the defect area so that the mesh size meets the error constraint conditions. The variational level set method is used to reconstruct the defect boundary and construct the level set energy function: Among them, φ is the level set function, α is the regularization parameter, and H(φ) is the Heaviside function; The level set energy function is solved by iterative optimization, the level set function φ is adjusted to make the defect boundary gradually converge to the optimal position, and the boundary curvature κ and depth h of the defect area are calculated; Feedback the optimized defect area information to the core tensor Update the data of defect feature dimension.

7. The PC lens defect classification and identification method according to claim 1, characterized in that: The steps of mapping the optimized core tensor to the QUBO Hamiltonian of the quantum annealing machine, optimizing the solution process by adopting a segmented temperature control strategy, and adjusting the calculation path include: Based on the optimized core tensor Construct the objective function of the QUBO problem and define the Hamiltonian H as: Among them, s i ∈{0,1} is a binary decision variable, Q ii For variable s i The corresponding first-order weight term, Q ij For variable s i and j The second-order interaction coupling weight term between ; Through the core tensor The modal decomposition of the objective function is used to calculate the coefficient matrix Q = {Q ii ,Q ij }, where Q ii According to the energy distribution weight setting of the defect area, Q ij Calculated based on the spatial correlation of defect characteristics; The objective function is mapped to the quantum annealing machine, and the optimization solution is performed according to the segmented temperature control strategy. The exponential cooling strategy T(t) = T0e is used in the first stage. -αt A global search is performed, where T(t) is the annealing temperature, T0 is the initial temperature, α is the cooling coefficient, and t is the time step; in the second stage, a periodic oscillation cooling strategy T(t) = T′0cos(β(t-τ)) is used for local optimization, where T′0 is the adjusted initial temperature, β is the oscillation frequency, and τ is the stage conversion time; During the annealing process, the convergence of the solution is monitored, and the optimal state is determined based on the rate of change of the Hamiltonian ΔH / Δt. H The calculation is terminated when ΔH is the change of Hamiltonian in two consecutive time steps, ∈ H is the set convergence threshold; The optimal solution is extracted, and the defect classification results and the geometric and mechanical parameters of the defects are restored according to the decoding strategy of binary variables to complete the QUBO Hamiltonian optimization solution process.

8. The PC lens defect classification and identification method according to claim 1, characterized in that: The steps of combining the obtained optimization results, fusing the multimodal confidence of polarized light, thermal imaging and acoustic emission signals, generating defect classification labels, and inverting the geometric and mechanical parameters of the defects include: Extract optimized core tensors The defect feature information in the image is used to obtain the polarized light image feature matrix F p , thermal imaging feature matrix F t The characteristic matrix F of the acoustic emission signal s ; Calculate the confidence score of each modal data; Construct a fusion model based on the confidence score and calculate the multimodal comprehensive confidence C fusion : C fusion =w p C p +w t C t +w s C s Among them, C p , C t , C s Represent the confidence scores of polarized light, thermal imaging and acoustic emission signals, respectively, w p 、w t 、w s Represent the weighting coefficients of polarized light, thermal imaging and acoustic emission signals respectively; According to C fusion Determine the defect classification label, where: If C fusion ≥τ1, it is marked as a surface crack; If τ2≤C fusion <τ1, it is marked as a deep defect; If C fusion <τ2, it is marked as a defect-free area; Among them, τ1, τ2 are classification thresholds, determined by experimental data; According to the classification results, the geometric parameters of the defect area are inverted to calculate the defect depth h, boundary curvature κ and stress concentration factor K t , and combined with the physical and mechanical model to determine the material properties of the defects, completing the defect classification and parameter inversion process.

9. The PC lens defect classification and identification method according to claim 8, characterized in that: The confidence scores of the modal data are calculated as follows: Among them, |F(i,j)| is the absolute value of the eigenvalue of the i,jth pixel.

10. A PC lens defect classification and identification system, used to execute the PC lens defect classification and identification method according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to obtain polarized light images, thermal imaging data and acoustic emission signals of the object to be tested, and convert them into a multimodal data matrix; Tensor construction module, used to construct a fifth-order asymmetric tensor based on multimodal data and normalize the tensor; Improved Tucker decomposition module, used to dynamically reduce the dimension of fifth-order asymmetric tensors and extract optimized core tensors; Polarization angle-time correlation coding module, which is used to establish a polarization angle-time correlation coding model based on the inverse problem solving method of the light field state equation, derive the distribution characteristics of defects, calculate the physical parameters of the defect area, and optimize the core tensor; The quantum optimization solution module is used to map the optimized core tensor to the QUBO Hamiltonian of the quantum annealer and optimize the calculation path using a segmented temperature control strategy; Multimodal fusion analysis module, which is used to combine the optimization results, fuse the multimodal confidence of polarized light, thermal imaging and acoustic emission signals, generate defect classification labels, and invert the geometric and mechanical parameters of the defects; The display and storage module is used to display defect detection results, store calculation data, and output optimized defect classification and parameter information.

Images