Online monitoring method and system for light transmittance of optical lens

Abstract

The invention discloses an online monitoring method and system for the light transmittance of an optical lens, and relates to the technical field of precise optical detection. The objective of the invention is to solve the problem of large light transmittance calculation error caused by halo scattering noise interference, low optical path calibration precision of a complex curved lens, a measurement blind area of a large-aperture lens and environmental disturbance of a high-reflection / high-light-transmittance lens in the prior art. According to the scheme, a three-dimensional curved surface model is constructed based on Gray code structured light projection and ICP registration, a dynamic partition scanning instruction set is generated through an improved genetic algorithm, and multi-wavelength polarized light is driven to scan and collect multi-dimensional data; specular reflection and scattering noise are suppressed by combining an improved Stokes vector algorithm and an attention mechanism U-Net network, and the data precision of the complex curved surface is improved by adopting Bayesian optimization and Zernike polynomial weighted fusion; constructing a double-loop coupling optimization architecture to realize closed-loop adaptive optimization; according to the invention, the anti-interference capability, the complex lens adaptability and the full-aperture detection precision are obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of precision optical detection technology, and more specifically to an online monitoring method and system for the light transmittance of an optical lens. Background Art

[0002] Online monitoring of optical lens transmittance measures the intensity change of light passing through the lens in real time, calculating transmittance to assess lens quality. Existing technologies utilize instruments such as spectrophotometers, integrating spheres, and fiber optic spectrometers, combined with automated testing devices, to achieve efficient and accurate transmittance monitoring.

[0003] The core of online monitoring of optical transmittance is to illuminate the lens with a light source (such as laser or infrared light), receive the transmitted light intensity by a photodetector, and calculate the transmittance ratio (transmitted light intensity / incident light intensity). The spectrophotometer method uses a spectrophotometer to measure the transmitted light intensity of the lens at different wavelengths to calculate the transmittance; the integrating sphere method reduces errors by evenly scattered light, while the photocurrent method requires two measurements (with / without the lens) to calculate the ratio. In existing patented technologies, such as the transmittance measuring instrument mentioned in patent CN222318409U, the position of the laser light source is adjusted in three dimensions along the Z, X, and Y axes, combined with a multi-point photodetector to achieve transmittance detection in different areas of the lens. The touch screen displays data in real time, which is suitable for curved lenses and complex incident angle scenarios. Its principle is to illuminate the lens vertically with a parallel light beam, use a photoelectric sensor to receive the transmitted light intensity, and support multi-point dynamic scanning; patent CN103149181B proposes a glass transmittance detection device that achieves optical path calibration through magnet alignment and semi-transparent and semi-reflective devices to reduce tilt errors. Using time-sharing measurements using multiple light sources, including white light, infrared, and ultraviolet, combined with an aperture to filter stray light, the device can measure the transmittance and reflectance of thick glass (such as automotive insulating glass). Furthermore, the automated inspection device disclosed in patent CN118980641A utilizes a rotating and clamping mechanism to achieve multi-angle and multi-directional imaging, automatically finding the optimal inspection position and improving inspection efficiency and accuracy.

[0004] However, although the above-mentioned equipment has improved the efficiency and accuracy of online monitoring of the transmittance of optical lenses to a certain extent, there are still some problems in its implementation. For example, in the online monitoring of electrical equipment, high-reflection / high-transmittance lenses are prone to cause halos and scattered noise. Existing filtering algorithms (such as median filtering) cannot effectively distinguish between real defects and noise, resulting in transmittance calculation errors. In other application scenarios, environmental factors such as temperature and humidity may also affect the measurement of transmittance. Secondly, although the existing technology can measure the transmittance of optical lenses within a certain range, it is difficult to measure large-aperture optical lenses, especially when the incident beam angle range is wide and the scanning mode is used. There are also problems with measuring optical lenses of different apertures in the same device. In addition, for free-form surface lenses and lenses with complex high curvature structures, the optical path calibration accuracy of the existing technology is low, and the error increases significantly when the incident angle deviates from the standard setting. It is also difficult to adapt to the optical path changes and refractive index unevenness caused by the curved surface of the lens, and it is easy to cause transmittance calculation errors due to uneven light intensity distribution. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the present invention discloses an online monitoring method and system for the light transmittance of an optical lens, aiming to solve the problems in the background technology.

[0006] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] An online monitoring method for light transmittance of an optical lens, comprising:

[0008] Step 1: Based on the Gray code structured light projection and edge extraction algorithm, the ICP point cloud registration method is used to construct a 3D surface model of the lens, and the 3D topographic feature data including curvature gradient, normal vector and vertex coordinates is output;

[0009] Step 2: Input the three-dimensional topographic feature data and lens aperture into an improved genetic algorithm module, use light intensity uniformity and scanning efficiency as multi-objective functions, iteratively calculate the focal length of the zoom lens group, the polarizer angle, and the iris opening and closing parameters, and generate a dynamic partition scanning instruction set;

[0010] Step 3: driving the tunable filter to generate multi-wavelength polarized light according to the dynamic partition scanning instruction set, and synchronously collecting the transmission spectrum, polarization state light intensity signal and environmental sensor data through the integrating sphere array to generate a multi-dimensional detection data set;

[0011] Step 4: Based on the multidimensional detection dataset, an improved Stokes vector algorithm is used to decompose the specular reflection noise component, and a U-Net network with an attention mechanism is used to suppress material scattering noise. A multivariate linear regression model is used to compensate for the light intensity offset caused by temperature drift and humidity changes in real time, and basic transmittance data is output.

[0012] Step 5: Divide the lens into 5 to 10 dynamic sub-regions based on the curvature gradient, use a Bayesian optimization algorithm to load the material history weight library, perform nonlinear weighted fusion on the transmittance basic data in a Zernike polynomial basis space, and enhance the edge area data density through Monte Carlo sampling to generate a full-aperture transmittance distribution cloud map;

[0013] Step 6: Construct a dual-loop coupled optimization architecture. The inner loop monitors the transmittance standard deviation with a 10ms period and adjusts the photomultiplier tube gain using the gradient descent method. After the outer loop completes five detections, it filters historical data based on the Mahalanobis distance criterion and uses an improved Kalman filter to update the genetic algorithm population parameters.

[0014] As a further technical solution of the present invention, an online monitoring system for the light transmittance of an optical lens includes: a three-dimensional morphology modeling module, a dynamic scanning parameter optimization module, a multi-dimensional data acquisition module, a noise suppression module, a data fusion module, a closed-loop optimization module, and a visualization interaction module;

[0015] The three-dimensional shape modeling module is used to perform three-dimensional point cloud scanning on the lens surface based on Gray code structured light projection and edge extraction algorithm, and generate three-dimensional shape feature data including curvature gradient, normal vector and vertex coordinates through ICP point cloud registration method;

[0016] The dynamic scanning parameter optimization module is used to iteratively optimize the focal length of the zoom lens group, the polarizer angle, and the aperture opening and closing parameters using an improved genetic algorithm with light intensity uniformity and scanning efficiency as multi-objective functions, generate a dynamic partition scanning instruction set, and output it to the multi-dimensional data acquisition module;

[0017] The multidimensional data acquisition module is used to trigger the tunable filter to switch the polarization state based on the dynamic partition scanning instruction set, synchronously collect the transmission spectrum, polarization state spectroscopic data and environmental sensor signals through the integrating sphere array, generate a multidimensional detection data set after wavelet denoising, and transmit it to the noise suppression module;

[0018] The noise suppression module is used to receive the multidimensional dataset and three-dimensional topography data, separate the specular reflection noise through the anisotropic Fresnel reflection tensor, suppress the material scattering noise using the curvature-guided attention mechanism U-Net network, dynamically compensate for temperature and humidity drift in combination with a multivariate linear regression model, and output the denoised transmittance data to the data fusion module;

[0019] The data fusion module is used to divide the dynamic sub-regions based on curvature gradient clustering, fuse the transmittance data of each sub-region through Bayesian optimization in Zernike polynomial space, perform Monte Carlo sampling and radial basis function interpolation on the high curvature edge area, generate a full-aperture transmittance distribution cloud map, and transmit it to the closed-loop optimization module;

[0020] The closed-loop optimization module includes an inner-loop control unit and an outer-loop control unit; the inner-loop control unit calculates the transmittance standard deviation in a 10ms period and dynamically adjusts the photomultiplier tube gain using a gradient descent method; the outer-loop control unit uses Mahalanobis distance to filter historical data after every five measurements, updates the genetic algorithm parameters using an improved Kalman filter, and feeds the optimization results back to the dynamic scanning parameter optimization module and the inner-loop control unit;

[0021] The visualization interaction module is used to receive the transmittance cloud map and optimization parameters, render the three-dimensional thermal map through the WebGL engine, display the environmental parameters, noise suppression ratio and calibration accuracy in real time, generate a test report and transmit it to the user terminal.

[0022] Based on the above technical solutions, the positive and beneficial effects of the present invention are:

[0023] 1. This solution uses multi-wavelength polarized light acquisition in step 3 and the improved Stokes vector algorithm in step 4 to decompose specular reflection noise at a physical level. Combined with the attention mechanism U-Net network, deep feature extraction of spectral signals is performed, enabling accurate distinction between material scattering noise and true light transmission signals. At the same time, a multivariate linear regression model is used to compensate in real time for the effects of environmental factors such as temperature, humidity, and air pressure on light intensity, forming a multidimensional noise suppression and environmental adaptation mechanism. This fundamentally addresses the measurement errors caused by halo, scattering noise interference, and environmental parameter fluctuations in highly reflective / highly transparent lenses.

[0024] 2. By combining the 3D lens surface model constructed in step 1 with the improved genetic algorithm in step 2 to dynamically calculate optical path control instructions such as the focal length of the zoom lens group and the aperture opening and closing parameters, a multi-region partitioned scanning strategy was implemented for lenses with apertures ranging from 10 to 200 mm. This strategy automatically adjusts the beam diameter and incident angle based on the lens aperture, enabling a single device to measure lenses of varying apertures. Furthermore, by optimizing the scanning path and beam parameters, the system significantly improves the light intensity uniformity and measurement efficiency of large-aperture lenses across a wide angle range, addressing the limited adaptability of traditional fixed optical paths to large-aperture lenses.

[0025] 3. Through ICP point cloud registration in step 1, a high-precision 3D surface model is constructed to obtain features such as the lens surface curvature gradient and normal vector, providing precise geometric parameters for dynamic optical path adjustment in step 2. Step 5 further divides the dynamic sub-regions based on the curvature gradient. A Bayesian optimization algorithm is used to load the material history weight library, and the transmittance of each sub-region is nonlinearly weighted fused within the Zernike polynomial basis space. Monte Carlo sampling is used to enhance the data density of edge regions. This process achieves adaptive optical path calibration for complex curved lenses, ensuring that the incident light is perpendicular to the local surface normal, effectively compensating for optical path variations and refractive index non-uniformity, and significantly improving the uniformity of light intensity distribution. This resolves the measurement errors caused by the complex surface structure in traditional technologies.

[0026] 4. The dual-loop coupled optimization architecture constructed in step 6 involves an inner loop monitoring the transmittance standard deviation at a 10ms cycle and adjusting the photomultiplier tube gain in real time via gradient descent, enabling rapid response to light intensity fluctuations. The outer loop filters historical data based on the Mahalanobis distance criterion and uses a modified Kalman filter to update the genetic algorithm population parameters, forming a self-optimizing mechanism for the measurement strategy. This closed-loop feedback system dynamically adjusts measurement parameters based on lens type and test data without manual intervention, significantly improving the real-time and automated nature of the monitoring system and addressing the parameter drift and poor adaptability issues inherent in traditional open-loop testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0028] Figure 1 This is a structural diagram of an online monitoring method for light transmittance of an optical lens according to the present invention;

[0029] Figure 2 This is a working principle diagram of the online monitoring system for light transmittance of optical lenses of the present invention;

[0030] Figure 3 This is a working principle diagram of the ICP point cloud registration method of the present invention;

[0031] Figure 4 This is a working architecture diagram of the improved genetic algorithm of the present invention;

[0032] Figure 5 This is a working principle diagram of step 3 of the present invention;

[0033] Figure 6This is a specific working method step diagram for step 5 of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] In an embodiment, the online monitoring method of the light transmittance of an optical lens is as follows: Figure 1 Shown, including:

[0036] Step 1: Based on the Gray code structured light projection and edge extraction algorithm, a three-dimensional surface model of the lens is constructed by the ICP point cloud registration method, and three-dimensional morphological feature data including curvature gradient, normal vector and vertex coordinates are output; Step 2: The three-dimensional morphological feature data and the lens aperture are input into the improved genetic algorithm module, and the light intensity uniformity and scanning efficiency are used as multi-objective functions to iteratively calculate the focal length of the zoom lens group, the polarizer angle and the aperture opening and closing parameters to generate a dynamic partition scanning instruction set; Step 3: According to the dynamic partition scanning instruction set, the tunable filter is driven to generate multi-wavelength polarized light, and the transmission spectrum, polarization state light intensity signal and environmental sensor data are synchronously collected through the integrating sphere array to generate a multi-dimensional detection data set; Step 4: Based on the multi-dimensional detection data set, the improved Stokes vector algorithm is used to decompose the mirror reflection noise component, and the injection molding process is combined with the injection molding process to generate a dynamic partition scanning instruction set. The U-Net network of the Yili mechanism suppresses material scattering noise, and uses a multivariate linear regression model to compensate for the light intensity offset caused by temperature drift and humidity change in real time, and outputs basic transmittance data; Step 5, divide the lens into 5 to 10 dynamic sub-areas according to the curvature gradient, use the Bayesian optimization algorithm to load the material history weight library, perform nonlinear weighted fusion on the basic transmittance data in the Zernike polynomial basis space, and enhance the data density of the edge area through Monte Carlo sampling to generate a full-aperture transmittance distribution cloud map; Step 6, construct a dual-loop coupling optimization architecture, the inner loop monitors the transmittance standard deviation with a 10ms period, and adjusts the photomultiplier tube gain by the gradient descent method; after the outer loop completes 5 detections, it filters the historical data based on the Mahalanobis distance criterion and uses the improved Kalman filter to update the genetic algorithm population parameters.

[0037] In step 1 of the above implementation, if Figure 3As shown in the figure, the working principle of the ICP point cloud registration method is as follows: based on the Gray code structured light projection module, multi-frequency phase coding stripes are projected onto the lens surface, the absolute phase distribution map is solved by the four-step phase shift interferometry method, and the phase jump noise is eliminated by combining the multi-frequency heterodyne phase unwrapping algorithm; the edge pixel coordinates of the phase image contour are extracted based on the Canny edge detection operator, and the edge position is corrected twice using the Zernike positioning algorithm to generate an edge coordinate sequence; based on the baseline calibration parameters and the epipolar correction matrix of the binocular stereo vision system, the edge coordinate sequence is converted into The method comprises the following steps: a three-dimensional discrete point cloud is decomposed into a rigid transformation matrix by SVD decomposition, an LM optimization algorithm is used to iteratively minimize the point cloud registration residual, and a RANSAC outlier removal mechanism is used to improve the registration accuracy; the registered multi-source point cloud is input into a Poisson surface reconstruction algorithm, and a continuous surface model of the lens is generated through octree space partitioning and gradient field optimization; the neighborhood covariance matrix of the surface vertices is calculated based on the principal component analysis method, the curvature gradient tensor and the normal vector direction are solved through eigenvalue decomposition, and a three-dimensional morphological feature dataset containing curvature gradient, normal vector and vertex coordinates is constructed.

[0038] Gray code structured light projection uses spatiotemporal modulation of phase-coded fringes, generating a fringe image with a unique absolute phase value through multi-frequency phase encoding. When calculating the phase distribution using four-step phase-shift interferometry, a multi-frequency heterodyne algorithm is introduced to perform cross-cycle unwrapping of the phase differences between adjacent fringes, eliminating the jump noise associated with traditional single-frequency phase unwrapping and ensuring the continuity and global consistency of the phase data. The Canny edge detection operator extracts the coarse coordinates of the phase image contour edge through multi-scale Gaussian filtering and dual-threshold hysteresis. The Zernike positioning algorithm is then combined to perform sub-pixel correction on edge pixels. The grayscale gradient distribution characteristics of the Zernike orthogonal moments are used to calculate the precise sub-pixel coordinates of edge points, improving edge positioning accuracy from the pixel level to below 0.1 pixel. Three-dimensional reconstruction in binocular stereo vision systems is based on epipolar geometric constraints and triangulation principles. Baseline calibration parameters (such as camera focal length and baseline distance) and an epipolar correction matrix are used to eliminate distortion differences between stereo image pairs. The two-dimensional edge coordinate sequence is then mapped into three-dimensional space to generate discrete point cloud data. The ICP (Iterative Closest Point) registration method uses singular value decomposition (SVD) to calculate the rigid transformation matrix (including the rotation matrix and translation vector). Combined with the LM (Levenberg-Marquardt) optimization algorithm, it iteratively minimizes the registration residual between the source and target point clouds. RANSAC (Random Sample Consensus) is used to remove outliers introduced by uneven lens surface reflectivity or environmental noise, significantly improving the robustness of point cloud registration. The Poisson surface reconstruction algorithm utilizes octree spatial partitioning to perform multi-resolution meshing on discrete point clouds. By solving the gradient field optimization problem of an implicit function, the discrete point cloud is converted into a continuous smooth surface model, overcoming the holes and jagged edges that can occur with traditional triangulation methods in complex surface topologies. Principal component analysis (PCA) calculates the eigenvalues ​​and eigenvectors of the covariance matrix of the neighborhood of surface vertices, analyzes the curvature gradient tensor (maximum curvature, minimum curvature and direction) and normal vector direction of the local surface, and constructs a multidimensional feature data set that characterizes the three-dimensional morphology of the lens, providing geometric prior knowledge for subsequent dynamic optical path calibration and transmittance calculation.

[0039] The structured light measurement system consists of a DLP projector module, a binocular global shutter camera, a high-precision rotation stage, and a black diffuse reflective background. The projector module and camera are fixed with a rigid bracket, and the baseline distance is calibrated at 200 mm to ensure that the triangulation field of view covers the φ200 mm lens. For electrical connections, the camera and projector module are connected to an industrial computer via Gigabit Ethernet, and the rotation stage receives control commands via an RS-485 interface.

[0040] During data acquisition, the rotating stage rotates at 15° intervals. At each angle, the projection module projects three-frequency Gray code stripes (with frequencies of 1 / 64, 1 / 32, and 1 / 16 cycles / pixel) and four-step phase-shifted stripes. The binocular camera simultaneously captures 60 images (3 frequencies × 4 phase shifts × 5 angles). During the phase solution stage, the three-frequency phase images are combined using the heterodyne method to generate an absolute phase distribution in the range [-π, π] with an error of less than 0.05π. During edge processing, the phase image is Gaussian filtered (σ = 1.5), and then the Canny operator (high-low threshold ratio 2:1) is used to extract edges. The edge coordinates are then corrected to sub-pixel accuracy (0.05 pixel) using the Zernike moment algorithm to generate a list of edge point coordinates. During the 3D reconstruction phase, the stereo camera is first calibrated using the Zhang calibration method (focal length error <0.1%, distortion coefficient <0.01). Epipolar correction is then performed on the image based on the calibration results. Then, for each edge point, the spatial coordinates are calculated using binocular parallax (depth accuracy ±0.1mm) to generate an initial point cloud. For multi-view point cloud registration, the first-view point cloud is used as the benchmark, and the rigid transformation parameters of subsequent point clouds are calculated through SVD decomposition. The number of LM algorithm iterations is set to 50, the number of RANSAC random sampling is set to 1000, and the culling ratio is set to 20%. Finally, the octree depth of the Poisson reconstruction algorithm is set to 8, the gradient weight is set to 0.8, and a triangular mesh surface is generated (vertex spacing ≤ 0.5mm). For surface feature calculation, the vertex neighborhood radius is set to 1mm, and PCA analysis generates the curvature tensor and normal vector. These are stored as a feature dataset containing 100,000 vertices for subsequent use by the optical path control module.

[0041] In another implementation, the projection module projected four phase-coded stripes (at frequencies of 60, 120, 240, and 480 line pairs / mm) onto the lens surface, and the camera captured the deformed stripe images using global shutter mode. When calculating the phase using the four-step phase shifting method, the phase calculation of each image was filtered using a 5×5 Gaussian window. A multi-frequency heterodyne algorithm used weighted least squares to fuse low- and high-frequency phase data to eliminate phase jumps. The gradient threshold for Canny edge detection was set to [0.1, 0.3] (normalized intensity range). Zernike localization used a 7×7 neighborhood window to calculate moment features, with sub-pixel correction error controlled within ±0.05 pixels. Epipolar correction of the binocular camera was implemented using the Bouguet algorithm. The baseline distance for triangulation transformation was calibrated to 200 mm, and the focal length was calibrated to 12 mm. During ICP registration, the LM algorithm iteration termination criterion was set to a residual change rate <1e-6, the RANSAC sampling number was 1000, and the outlier rejection threshold was 3 times the median absolute deviation. The octree depth for Poisson reconstruction was set to 10 levels, and the conjugate gradient method was used to solve the linear equations for gradient field optimization. The neighborhood radius for principal component analysis was set to 5 times the average spacing of the point cloud, and the curvature gradient was calculated based on quadratic surface fitting.

[0042] Compared with existing technologies, this method significantly improves the three-dimensional modeling accuracy and anti-interference ability of complex curved lenses by integrating multi-frequency heterodyne phase unwrapping, sub-pixel edge positioning and point cloud registration technology, and effectively solves the phase jump and edge blur problems caused by mirror reflection or environmental noise in traditional structured light projection. At the same time, the combination of Poisson surface reconstruction and principal component analysis realizes the refined analysis of the geometric characteristics of free-form surface lenses. The overall solution shows excellent adaptability and stability in industrial-grade complex optical lens detection.

[0043] During design testing, a comparative experiment was conducted to verify the accuracy and robustness advantages of the ICP point cloud registration method (Group A) proposed in this invention over the traditional ICP method (Group B) for 3D modeling of complex curved lenses. The experimental hardware included a Gray code structured light projection module (DLP4500), a high-resolution binocular industrial camera (Basler acA2440-75um), a synchronization trigger controller, a GPU computing server (NVIDIAA6000), a six-degree-of-freedom lens mounting platform, and a temperature and humidity controlled environmental isolation enclosure. The projector and camera were positioned at a 30° angle, covering the entire lens surface. The synchronization controller precisely coordinated projection and acquisition timing using an FPGA. The computing server was directly connected to the camera via a PCIe 4.0 interface to process point cloud data in real time.

[0044] The experiment used free-form lenses (radius of curvature range 10-200mm, anti-reflection coating), temperature 25±0.5°C, humidity 50%±5%, and identical hardware configuration. Group B used the basic ICP algorithm without RANSAC outlier removal or LM optimization, and only calculated the rigid transformation matrix through SVD decomposition. The experimental data is shown in Table 1:

[0045] Table 1 Experimental record of ICP point cloud registration method

[0046]

[0047]

[0048] Experiments show that the ICP algorithm of the present invention in Group A significantly reduces registration residuals (mean 0.011mm vs. 0.088mm in Group B) through RANSAC outlier removal and LM optimization. The outlier ratio drops to 1.1% (16.5% in Group B), the surface reconstruction error is reduced to 3.7μm (29.6μm in Group B), the normal vector deviation is stabilized at 0.24° (2.3° in Group B), and the registration speed is increased to 2.8 times that of Group B. Experiments demonstrate that the improved ICP offers higher accuracy, interference resistance, and efficiency in modeling complex curved lenses, providing a reliable geometric data foundation for subsequent transmittance calculations.

[0049] In step 2 of the above implementation, if Figure 4 As shown, the improved genetic algorithm module includes a parameter encoding unit, an adaptive crossover mutation unit, a multi-objective fitness evaluation unit, a constraint processing unit, an elite retention unit and a dynamic partition optimization unit; the parameter encoding unit is used to map the focal length of the zoom lens group, the polarizer angle, and the aperture opening and closing parameters into chromosome gene sequences through a real number encoding mechanism, and output an initial population matrix; the adaptive crossover mutation unit is used to dynamically adjust the crossover probability and the mutation probability through a Sigmoid function, combine the roulette wheel selection strategy to screen the parent generation individuals, trigger the two-point crossover operation to generate the offspring chromosome, and use the Gaussian perturbation mechanism to perform local perturbation optimization on the mutant gene; the multi-objective fitness evaluation unit is used to construct a dual-objective genetic algorithm based on the uniformity of light intensity and the scanning efficiency. The objective function generates a Pareto rank and a fitness value through non-dominated sorting and crowding calculation using the NSGA-II algorithm; the constraint processing unit is used to process the focal length range and polarization angle limit through a penalty function method, and embed the focal length range and polarization angle limit into the fitness function through the Lagrange multiplier method to generate a feasible solution set; the elite retention unit is used to adopt a steady-state selection strategy, retain elite individuals in the current population through a tournament mechanism, and output an optimized offspring population; the dynamic partition optimization unit is used to divide the scanning area based on the lens curvature gradient distribution data through the K-means clustering algorithm, and generate a dynamic partition scanning instruction set containing a time series of focal length, polarization angle, and aperture parameters in combination with the scanning path ant colony optimization algorithm.

[0050] The parameter encoding unit uses a real-number encoding mechanism to map continuous hardware parameters (zoom lens focal length f∈[5,150]mm, polarizer angle θ∈[0,90]°, and aperture d∈[1,50]mm) into chromosomal gene sequences. Each parameter corresponds to a gene position on the chromosome, avoiding the precision loss and decoding complexity of binary encoding. The adaptive crossover and mutation unit constructs a dynamic probability adjustment model using a sigmoid function. The crossover probability Pc=Pc_min+(Pc_max-Pc_min) / (1+ek(t-t0)) varies nonlinearly with the number of iterations t. High initial probability promotes population diversity, while low probability later focuses on local search. The mutation probability uses a Gaussian perturbation mechanism, adding Gaussian noise with a mean of zero and a standard deviation that decays with iterations to the selected genes, balancing global exploration and local exploitation. The roulette wheel selection strategy is combined with proportional sampling of individual fitness to ensure that high-quality individuals have a higher probability of participating in reproduction. The two-point crossover operation randomly selects two crossover points on the parent chromosome, exchanges the intermediate gene segments to generate offspring, and introduces a local optimization operator to enhance the quality of the solution.

[0051] The multi-objective fitness evaluation unit, based on the NSGA-II algorithm, constructs a dual objective function: an intensity uniformity objective of f1 = 1 - σ(I) / μ(I) (σ is the standard deviation of the intensity, μ is the mean), and a scanning efficiency objective of f2 = 1 / (Ts + Tp) (Ts is the scanning time, Tp is the optical path adjustment time). The population is divided into different Pareto levels using non-dominated sorting, and the density of solutions, represented by the crowding distance, is calculated to ensure a uniform distribution of solutions along the Pareto front. The constraint processing unit employs a penalty function to impose fitness penalties on individuals outside the focal length range (f < 5 or f > 150°) and the polarization angle limit (θ < 0 or θ > 90°). Inequality constraints are embedded in the objective function using the Lagrange multiplier method to generate a feasible solution space. The elite retention unit uses tournament selection (randomly selecting five individuals each time and retaining the best) to maintain a steady-state population update, preventing the loss of high-quality solutions during evolution. The dynamic partition optimization unit uses the K-means algorithm to divide the lens surface into 5-10 homogeneous areas (curvature standard deviation <10%) based on curvature gradient data. The ant colony algorithm is used to optimize the scanning path based on the geometric characteristics of each area (such as curvature radius and normal direction), minimize the optical path adjustment time between adjacent areas, and generate a timing control instruction set containing the focal length, polarization angle, and aperture parameters unique to each area.

[0052] During implementation, the improved genetic algorithm module runs on an industrial control computer, connected via a PCI-E interface to a motion control card (e.g., GALIL DMC-2143). This drives a motorized zoom lens assembly (Optotune EL-10-40, with a focal length adjustment accuracy of ±0.1 mm), a polarizer rotation stage (Newport 805G, with an angular resolution of ±0.01°), and a programmable aperture (Edmund Optics 85772, with an aperture adjustment step of 0.1 mm). Ethernet is used for electrical communication with the hardware devices, and the control command update frequency is 100 Hz, ensuring real-time adjustment of optical path parameters.

[0053] During the operation, in the initialization phase, the 3D morphological feature data is obtained from step 1, the lens aperture D and the curvature gradient distribution matrix G are extracted, and the parameter encoding unit generates an initial population of 50 individuals. The gene sequence of each individual is a 3D vector of [f, θ, d], with f∈[5,150] uniformly sampled, θ∈[0,90] randomly generated, and d∈[1,50] ranging from 10% to 80% of the aperture D. In the evolutionary computation phase, each generation performs the following operations:

[0054] ① According to the current number of iterations t (total iterations 100 generations), Pc (initial 0.9, final 0.6) and Pm (initial 0.1, final 0.01) are calculated using the Sigmoid function, and two-point crossover (probability Pc) and Gaussian mutation (standard deviation 0.1f, 0.5θ, 0.5d) are performed on the selected parent generation;

[0055] ② Call the optical simulation module (Zemax Optic Studio) to calculate the light intensity uniformity and scanning time of each individual, generate the Pareto rank through the NSGA-II algorithm, and use the crowding distance normalized as the fitness;

[0056] ③ Imposing a 10% fitness penalty on individuals that violate hardware restrictions to ensure that the proportion of feasible solutions is >95%;

[0057] ④ A tournament selection process retains 20% of high-quality individuals for the next generation. During the dynamic partition optimization phase, K-means clustering (number of clusters K = 5 + int(D / 40)) is performed on the curvature gradient matrix G to generate the geometric center and characteristic parameters of each region. For each region, an ant colony algorithm initializes 10 ants and optimizes the scanning sequence with the goal of minimizing optical path adjustment time. This generates an instruction set containing the parameters [f_i, θ_i, d_i] for each region and the scanning timing. The instruction accuracy is f ± 0.1 mm, θ ± 0.1°, d ± 0.1 mm, and the timing error is < 50 ms.

[0058] Compared to existing technologies, the multi-objective optimization and dynamic partitioning strategy of the improved genetic algorithm in this paper effectively resolves the conflict between scanning efficiency and light intensity uniformity caused by traditional single-objective optimization, significantly improving the parameter adaptation accuracy and scanning path planning efficiency of complex curved lenses. Furthermore, the combination of an adaptive constraint processing mechanism and an elite retention strategy ensures that the algorithm rapidly converges to the global optimal solution while meeting hardware physical constraints. This overcomes the shortcomings of traditional genetic algorithms, which are prone to local optimality and weak anti-interference capabilities, and provides a more robust dynamic control solution for the online monitoring of high-reflectivity / high-transmittance lenses.

[0059] Similarly, when designing the test, to verify the multi-objective solving capability and efficiency advantages of the improved genetic algorithm module (Group A) compared to the traditional genetic algorithm (Group B) in dynamic parameter optimization, a free-form surface lens (curvature radius 5-150mm, aperture 50mm, anti-reflection coating) was selected. A comparative experiment was conducted at a temperature of 25±0.5°C, a humidity of 50%±5%, and the same hardware configuration (zoom lens assembly, polarizer, aperture, FPGA controller, spectrometer, integrating sphere). Group B used a single-objective genetic algorithm (optimizing only light intensity uniformity), fixed crossover probability (P_c=0.8) and mutation probability (P_m=0.05), no dynamic partitioning strategy, and a simple penalty function for constraint processing. The comparative data is shown in Table 2:

[0060] Table 2 Comparison data of improved genetic algorithm modules

[0061]

[0062]

[0063] Experiments show that the improved algorithm significantly improves scanning efficiency while ensuring the uniformity of light intensity distribution, and has stronger global convergence ability and stability, solving the defects of traditional single-objective algorithms such as low efficiency and easy violation of hardware constraints.

[0064] In the above implementation step 3, if Figure 5 As shown in Figure 3, the specific working principle of step 3 is as follows: first, the tunable filter is triggered to switch the target wavelength through the dynamic partition scanning instruction set, a sinusoidal voltage signal is applied to the electro-optical modulator to generate linearly polarized light, and the elliptical polarization state is dynamically adjusted through the delayed phase matrix of the birefringent crystal; six groups of sub-spheres in the integrating sphere array are synchronously triggered to capture the transmission spectrum in a time-sharing manner, where each group of sub-spheres has a built-in four-channel detector to collect Stokes parameters in the polarization directions of 0°, 45°, 90°, and 135°, respectively, and the spectral signal is transmitted to the high-speed ADC module through the fiber coupler; at the same time, the temperature, humidity, and air pressure data from the environmental sensor are read, and the Kalman filter algorithm is used to eliminate the wavelength offset error caused by temperature drift. Based on the IEEE1588 clock protocol, the polarization state light intensity, spectral intensity, and environmental parameters are aligned according to the timestamp to construct a three-dimensional transmittance tensor containing wavelength, polarization state, and environmental factors; during the acquisition process, if the light intensity of any sub-sphere is detected to exceed the preset threshold, the missing signal is reconstructed based on the data of the adjacent sub-spheres using the radial basis function interpolation algorithm.

[0065] Among them, in the acousto-optic effect of the tunable optical filter (AOTF), its wavelength selectivity is derived from the phase matching condition of Bragg diffraction: when the ultrasonic wave driven by the RF signal frequency ff forms a periodic refractive index modulation in the TeO2 crystal, the momentum conservation of the incident light must satisfy Δk = kacoustic, that is, 2πn / λsinθ = 2πf / vacoustic, where vacoustic is the crystal sound velocity. The transmission wavelength can be precisely controlled by adjusting f, but the nonlinear dispersion effect will cause interference of high-order diffracted light, and the sidelobe noise needs to be eliminated by iterative correction of the grating equation λ = vacousticf·Δn. The polarization conversion of the electro-optic modulator is based on the linear electro-optic tensor characteristics of the lithium niobate crystal, and its refractive index change satisfies Δnij = γ ijk Ek, when an alternating electric field is applied along the Z axis of the crystal, the refractive index difference in the X / Y direction forms a birefringent phase delay Where ne and no are the extraordinary and ordinary refractive indices, thereby achieving dynamic modulation of the linearly polarized light output. The polarization state control of the birefringent crystal is modeled using the Jones matrix: the output of the incident polarized light after passing through the crystal is jout = R(-θ)·Γ(δ)·R(θ)·Jin, where R(θ) is the rotation matrix and Γ(δ) = [e-iδ / 200eiδ / 2] is the phase delay matrix. By adjusting the crystal tilt angle θ and thickness δ, the major axis azimuth and ellipticity of the elliptical polarization state can be precisely controlled.

[0066] Light intensity detection using an integrating sphere array is based on the integral equation of radiative transfer: after multiple diffuse reflections within the sphere, the outgoing light flux density E satisfies E = ρΦ / 4πR²(1-ρ), where ρ is the inner wall reflectivity and R is the sphere radius. Four-channel polarization detection uses light intensity measurements at four different analyzer angles (0°, 45°, 90°, and 135°) to invert the Stokes parameters. The mathematical relationship can be expressed as a linear equation system I = M·S, where M is the Mueller matrix. Solving S = (MTM)-1MTI using the least squares method yields a complete description of the polarization state. For ambient temperature and humidity compensation, the Kalman filter's state-space model models the wavelength drift Δλ as a linear combination of temperature TT and humidity HH: Δλ = kTΔT + kHΔH + w, where the process noise ww follows a Gaussian distribution. The kT and kH coefficients are iteratively optimized through time updates and measurement updates, enabling dynamic correction of the spectrometer wavelength calibration. IEEE1588 clock synchronization eliminates network delay asymmetry by measuring the round-trip time (RTT) of the master and slave clocks. The slave clock frequency deviation Δf is adjusted by the PI controller to adjust the local crystal voltage to achieve sub-microsecond time alignment. The error convergence condition is ∫0tΔfdt<∈. The radial basis interpolation algorithm constructs an implicit function in a high-dimensional space. The missing signal is approximated, where the bandwidth parameter ∈ of the kernel function φ(r) = e-∈r2 is optimized by cross-validation, and the weight wi is determined by the regularized least squares solution w = (Φ+λI)-1y to ensure the light intensity continuity of the interpolation result.

[0067] During implementation, in terms of hardware deployment, step 3 requires the integration of a liquid crystal tunable filter (Cambridge Research Instruments, Varispec, wavelength range 400-720 nm, switching time 50 ms), a lithium niobate electro-optical modulator (Thorlabs, EO-AM-NR-C4, bandwidth 10 GHz), six integrating spheres (LabSphere, 4P-GPS-040-SL, inner diameter 40 mm, four-channel InGaAs detector), a temperature and humidity sensor (Honeywell, HIH9000, accuracy ±1.5% RH), and an air pressure sensor (TE Connectivity, MS5611, resolution 0.012 mbar), which are interconnected with the FPGA controller (National Instruments, PXIe-8880) via the PXIe bus. The electro-optic modulator's sinusoidal drive voltage (frequency 1 kHz, amplitude 0-10 Vpp) is generated by a DAC module (NIPXI-6733). The integrating sphere shutter is driven by a stepper motor (Sanyo Denki, R2AA08010D, step angle 0.72°) with a switching period of 1 ms. In the software flow, the Kalman filter's state equation noise covariance Q is set to 0.01, the observation noise covariance R is set to 0.1, and the wavelength correction value is updated every 100 ms. The IEEE 1588 protocol synchronizes the device clocks via PTPv2 messages, and the timestamp offset compensation is written to the FPGA's FIFO buffer. The Stokes parameter calculation uses the four-intensity method. A Gaussian kernel function (bandwidth σ = 0.5) is used for RBF interpolation, and weights of neighboring sub-spheres are distributed inversely proportional to their distance from the sphere. The ADC module sampling data is transmitted in real time to the GPU server (NVIDIA Tesla V100) via the PCIe 3.0 interface to perform tensor construction and anomaly detection, completing all data processing within a single scan cycle (500ms).

[0068] In another implementation, the instruction parsing module reads the dynamic partition scanning instruction set generated in step 2 and parses the target wavelength (e.g., 400nm, 550nm, and 800nm ​​three-wavelength scanning), polarization angle (0°, 60°, 90°), and integration time (50-500ms adaptive adjustment) of each area. During the signal acquisition phase, the motion control card (GALIL DMC-2143) drives the filter switching at a frequency of 100Hz, synchronously triggering the integrating sphere detector to acquire 100 frames of spectral data (signal-to-noise ratio > 50dB) when each sub-area is scanned. The data synchronization module implements hardware timestamp marking through the IEEE 1588 clock chip (Microsemi AXL1001) and performs Kalman filtering on the environmental data (state transfer matrix F = 1, observation matrix H = 1, process noise covariance Q = 10 -4 , observation noise covariance R = 10-3 ) to eliminate wavelength drift caused by temperature (e.g. wavelength shift <0.05nm at 25°C).

[0069] In the signal reconstruction phase, if the light intensity value of a channel exceeds the linear range of [5%, 80%], it is automatically identified as invalid data and RBF interpolation is started: the synchronous data of the five frames before and after the channel and the adjacent sub-balls are used as input to construct the interpolation function f(x) = ∑w_iexp(-||x-x_i|| 2 / (2σ 2 ), where x is the timestamp, w_i is the weight coefficient, σ = 2ms, and the reconstruction error is <3%. The resulting multidimensional detection dataset contains the spectral matrix (wavelength × intensity), Stokes parameter vector, and environmental parameter set for each subregion. The data format is HDF5, and the single-lens data volume is approximately 200MB, which is used by the subsequent noise separation module.

[0070] Compared with the single-wavelength non-polarized light acquisition in the existing technology, this method significantly improves the anti-interference ability and data integrity of the transmittance detection of complex optical lenses through dynamic control of multi-wavelength polarization states and high-precision spatiotemporal synchronous acquisition mechanism, and effectively solves the signal distortion problem caused by environmental disturbances and mirror scattering in traditional single-wavelength or fixed polarization state measurements. At the same time, the distributed design of the integrating sphere array and the radial basis function interpolation strategy overcome the data loss problem caused by local light intensity saturation or occlusion, providing a highly robust, high-dimensional fusion data acquisition solution for industrial-grade complex lens online monitoring.

[0071] In step 4 of the above implementation, the improved Stokes vector algorithm is based on the Stokes parameter matrix S in the multidimensional detection data set. toal =[S0,S1,S2,S3] T , where S0=I0+I 90 , S1=I0-I 90 , S2=I 45 -I 135 , S3 is obtained by differential measurement of left-hand and right-hand circular polarizers; T is temperature; a polarization reflection and transmission separation model is constructed that includes lens anisotropy and multi-wavelength coupling effects. The working principle is as follows:

[0072] S1. Introduce the anisotropic Fresnel reflection tensor R(θ,λ,n) of the lens surface, establish the polarization state mapping relationship of the reflection noise, and calculate the mirror reflection Stokes vector S reflected , the calculation formula is:

[0073] S reflected =R(θ,λ,n)·S incident +Δ roug hness (1)

[0074] In formula (1), Sincident represents the Stokes vector of the incident light; θ is the incident angle, which is used to quantify the geometric relationship between the light and the lens surface; λ is the wavelength of light, which determines the spectral dependence of the reflectivity; n is the refractive index tensor, which characterizes the anisotropic optical properties of the lens material; Represents the anisotropic Fresnel reflection tensor, which is used to quantify the polarization crosstalk effect of specular reflection. Its matrix element R ss 、R sp 、R ps and R pp represents the cross-polarization reflectivity; represents the surface roughness scattering noise term, which is used to reflect the polarization state disturbance of non-specular scattered light, where σ represents the root mean square value of surface roughness, which is used to determine the scattering noise intensity; L c represents the surface topography autocorrelation length; k is the wave number, which is inversely proportional to the wavelength λ and is used to quantify the sensitivity of the scattering effect to the wavelength;

[0075] S2. Extract the transmission component S transmitted Eliminate the material body scattering noise, the transmission component is based on the coupling model of birefringence effect and scattering phase function, and define the polarization scattering correlation tensor P scatter :

[0076]

[0077] In formula (2), β is the scattering angle and φ is the azimuth angle;

[0078] S3, the mirror reflection Stokes vector S reflected From the Stokes parameter matrix S total Subtract the mid-vectors to get the transmitted Stokes vector S transmitted , the transmitted Stokes vector S is segmented by the polarization threshold algorithm transmitted The residual volume scattering noise in the image is suppressed.

[0079] The working principle of the attention mechanism U-Net network is as follows: first, the local texture features and global scattering patterns of the light intensity distribution map are extracted through the multi-scale residual convolution block on the encoder side. The multi-scale residual convolution block uses an expanded convolution kernel to capture the scattering noise characteristics of different scales; then, based on the three-dimensional morphology curvature gradient data of step 1, a curvature-guided attention gating mechanism is embedded in the jump connection layer, and the association weights of the curvature features and the light intensity features are calculated through the covariance matrix to dynamically suppress the characteristic responses of the areas with high incidence of scattering noise; on the decoder side, a polarization-wavelength dual-branch deconvolution module is used. Branch one analyzes the spatial distribution of the Stokes parameter, and branch two solves the wavelength-scattering phase function mapping relationship, and the dual-branch feature map is fused in combination with the gated recurrent unit; finally, the binary noise distribution map is output through the scattering noise mask generation layer, and the original light intensity data is differentially operated with the noise mask to obtain the de-scattered transmitted light intensity matrix.

[0080] The working principle of the multiple linear regression model is: first, based on the temperature data T, humidity data H and descattered transmitted light intensity I transmitted The mapping relationship between environmental variables and light intensity offset is solved by ridge regression regularization algorithm: ΔI=β0+β1T+β1H+β3TH, where TH is the temperature-humidity interaction term, which is used to quantify the nonlinear influence of the synergistic effect of temperature and humidity on the refractive index; β0, β1, β1 and β3 represent the regression intercept term, temperature coefficient, humidity coefficient and cross-coupling coefficient, respectively; then, the training set is updated in real time based on the sliding window mechanism, and the regression coefficient β is dynamically corrected using the recursive least squares method. i , compensate for the time-varying characteristics of the environmental drift; finally, the offset ΔI predicted by the regression is reversely corrected to the transmitted light intensity through the ambient light intensity coupling matrix C(T,H), and the compensated transmittance basic data η=(I transmitted -ΔI) / I reference , where I reference Indicates the reference light intensity value, which is used to normalize the calculation of transmittance.

[0081] Among them, when implementing the improved Stokes vector algorithm, the algorithm hardware runs on an industrial computer with GPU acceleration (NVIDIA RTX A6000), and the input data is the multidimensional detection data set generated in step 3 (including the Stokes parameter matrix S total (λ, θ), environmental parameters [T, H], and normal vector data from step 1). In the data preprocessing stage, the Stokes parameters are first wavelength aligned (resolution 1 nm) and polarization angle normalized (interval 5°), and high-frequency random noise is eliminated by median filtering (window size 3×3).

[0082] The calculation of the Stokes parameter matrix is ​​based on the four-intensity method. In the Fresnel reflection tensor of the improved Stokes vector algorithm, the non-diagonal elements R sp 、Rps The polarization response curves of the lens material at different incident angles (5°-85°) and wavelengths (400-1000nm) were calibrated by an ellipsometer (JAWoollam M-2000V) and stored as a lookup table for real-time access. c The lens surface topography data was measured using a white light interferometer (Zygo NewView 9000) and fitted to a Gaussian correlation function model. The input layer of the U-Net network is a 512×512 pixel transmitted light intensity distribution map. The encoder uses the ResNet-34 backbone network, the compression ratio of the channel attention module is set to 16, the spatial attention kernel size is 7×7, and the Adam optimizer (learning rate 1e-4) and Focal Loss loss function are used for training. The dataset contains 2000 sets of annotated samples of lens scattering noise. The environmental variable weight update cycle of the multivariate linear regression model is 1 second, the forgetting factor of the recursive least squares method is set to 0.99, and the temperature drift compensation coefficient is calibrated through closed-loop feedback of the thermoelectric cooler (TEC). In the software processing flow, the FPGA synchronously triggers the polarization camera and environmental sensor data acquisition with a 10ms period, transfers the data to the GPU server (NVIDIA A100) via DMA to perform noise separation and compensation operations, and outputs the basic transmittance data in real time. The specific algorithm execution process is as follows:

[0083] Specular reflection modeling: Based on the three-dimensional morphology of the lens in step 1, the average incident angle θ is calculated for each sub-area. The anisotropic Fresnel reflection tensor Ra is generated by combining the material database (including the refractive index tensors of 10 common lens materials such as glass and plastic) to calculate the specular reflection Stokes vector S. reflected , where the surface roughness parameters σ,l are determined by historical measurement data statistics (default σ=5nm,l=50nm).

[0084] Transmission component extraction: S total Perform vector subtraction S per wavelength and per polarization angle transmitted =S total -S reflected , generate a preliminary transmission vector matrix, and then remove the mirror reflection residual signal with polarization degree P>0.1 through polarization threshold segmentation.

[0085] Volume scattering suppression: S transmitted Convert it to a grayscale image (l channel as brightness, Q / U channel as color channel), input the pre-trained attention mechanism U-Net network (the training data contains 200,000 annotated noise-signal image pairs), the network uses skip connections and channel attention weight adjustment to enhance the feature suppression of scattered noise, and outputs the denoised transmitted light intensity matrix I clean (λ).

[0086] Environmental compensation: Based on the real-time temperature and humidity data [T, H], the multivariate linear regression model (training error RMSE < 0.5%) is used to calculate the light intensity offset ΔI. clean Perform wavelength-by-wavelength compensation and finally output the basic transmittance data T(λ)=I clean (λ) / I ref (λ), where I ref It is the reference light intensity when there is no lens.

[0087] During implementation, the algorithm software is implemented in Python language, and the U-Net network is built based on the PyTorch framework. The Stokes tensor calculation part is accelerated by CUDA, and the single-lens processing time is controlled within 100ms, meeting the real-time requirements of online monitoring.

[0088] In the implementation of the attention mechanism U-Net network, the multi-scale residual convolution block uses dilated convolution technology. By setting convolution kernels with different dilation rates (such as dilation rate = 2, 4, 6), the receptive field is expanded without reducing the size of the feature map, thereby capturing cross-scale noise characteristics from microscopic textures to macroscopic scattering patterns. Its mathematical essence is to adjust the sampling interval of the convolution kernel, with an equivalent kernel size keff = k + (k-1)(d-1) (k is the original kernel size, d is the dilation rate), to achieve hierarchical modeling of multi-level scattering noise.

[0089] The curvature-guided attention gating mechanism is based on the curvature gradient data of the three-dimensional lens morphology, and calculates the association weights between the curvature feature C∈RH×W×3 and the light intensity feature I∈RH×W×4 through the covariance matrix. Specifically, the covariance matrix It quantifies the spatial coupling strength between regions of sudden curvature changes and scattering noise, and generates a spatial attention mask using a sigmoid function to dynamically suppress noise responses in regions of sudden curvature changes (such as edges and scratches). The polarization-wavelength dual-branch deconvolution module employs a physics-driven design: Branch 1 uses a four-channel convolution kernel to analyze the spatial distribution characteristics of polarization scattering (S0 to S3) based on the polarization state mapping relationship of the Stokes parameters; Branch 2 utilizes the wavelength-angle coupling characteristics of the wavelength scattering phase function P(θ,λ) to extract wavelength-dependent scattering patterns through frequency-domain convolution.

[0090] The Gated Recurrent Unit (GRU) introduces a time gating mechanism in the feature fusion stage, and its update gate z t and reset gate r t The memory retention and forgetting ratios of polarization and wavelength characteristics are controlled separately by This modeling of the temporal correlation of noise features is achieved. The scattering noise mask generation layer adopts a differential learning strategy, using a binary cross-entropy loss function to optimize the threshold segmentation boundary of the noise distribution map. Finally, the residual scattering component is eliminated by pixel-by-pixel difference between the light intensity data and the noise mask.

[0091] During implementation, the network ran on an industrial computer equipped with an NVIDIA A100 GPU, using the PyTorch framework for distributed training. The input data was the multidimensional detection data generated in step 3 (resolution 256×256 pixels, 3 channels (I / Q / U) + 1 (wavelength)). During preprocessing, the light intensity data was normalized to [0, 1], and the curvature gradient data was reduced to image resolution using Gaussian blur (σ = 2 pixels) to generate the geometric feature guidance map G(x, y). The network training process was as follows:

[0092] Dataset construction: 200,000 annotated synthetic images were generated by optical simulation software (such as TracePro) to simulate different curvatures (0.01-0.1mm -1 ), surface roughness (σ = 2-10nm) and scattering type (Mie scattering / Rayleigh scattering) noise scenes, and pixel-level masks are used to mark the noise areas.

[0093] Encoder feature extraction: The input light intensity map is processed by 3 layers of multi-scale residual blocks. Each layer contains parallel convolutions with dilation rates of 1 / 2 / 4, and outputs feature maps with 8 / 16 / 32 channels to capture scattering patterns from edges to the global level.

[0094] Curvature attention modulation: At the 2nd / 3rd layer jump connection, the curvature guidance map G is spliced ​​with the light intensity feature map, and the attention weight is generated through 1×1 convolution to suppress the noise features in the high curvature area (weight attenuation coefficient 0.3-0.7).

[0095] Dual-branch decoding and fusion: The polarization branch processes the Stokes parameters, the spectral branch processes the wavelength dimension features (the wavelength axis is converted to the spatial dimension through channel reorganization), and the GRU unit processes the dual-branch features in scan line order and outputs the fused noise probability map.

[0096] Mask generation and signal stripping: After the probability map is activated by Sigmoid, a binary mask is generated by the Otsu threshold method, which is subtracted pixel by pixel from the original light intensity data to obtain the de-scattered transmitted light intensity matrix I clean (λ,x,y).

[0097] During the inference phase, the processing time for a single image is controlled within 20ms (batch size 16), and the network parameters are compressed to 128MB through model quantization technology (FP16), supporting real-time collaborative operations with the Stokes vector algorithm.

[0098] This method significantly improves the ability to suppress scattering noise of complex curved lenses through a curvature-guided attention mechanism and a multimodal feature fusion strategy, effectively solving the defect of traditional U-Net networks in misjudging noise in geometrically distorted areas. At the same time, the dual-branch decoding structure takes into account the feature decoupling requirements of polarization and wavelength domains, overcoming the problem of insufficient adaptability of a single feature extraction model to multi-physical field noise.

[0099] In the multivariate linear regression model, the ridge regression regularization algorithm constrains the model coefficients by introducing the L2 norm penalty term (λ∑i=1nwi2). Its mathematical essence is the covariance matrix X of the design matrix X. T X is diagonalized (X T X+λI), thereby overcoming the multicollinearity between temperature, humidity, and their interaction terms and suppressing parameter estimation bias caused by ill-conditioned matrices. The temperature-humidity interaction term (T·H) expands the model's expressive dimensionality by introducing a nonlinear coupling term. Its physical mechanism stems from the synergistic effect of Fick's diffusion law and the thermodynamic expansion effect. It quantifies the cross-effect of temperature gradient and humidity penetration on the lens' refractive index as a quadratic coefficient γ, thereby accurately describing the nonlinear modulation of the environmental variables on the optical path difference.

[0100] The sliding window mechanism is based on the principle of dynamic truncation of time series. It reconstructs the training set through a data subset within a fixed time span (such as the last N samples). The technical key lies in the matching of window length and system response time. It uses an exponential decay function to weight historical data (such as ρn-k, ρ<1), thereby enhancing the ability to track sudden environmental changes while retaining long-term trends.

[0101] The core of the recursive least squares method (RLS) lies in the recursive update of the covariance matrix. Its iterative formula w n =w n -1+K n e n In the equation, Kalman gain K n Through the covariance matrix P n The inverse operation of Adaptive parameter adjustment is achieved, where the forgetting factor λ controls the decay rate of historical data, allowing the model coefficients to converge to the optimal solution in real time as the environment drifts. The construction of the ambient light intensity coupling matrix is ​​based on the generalized inverse matrix theory. The regression prediction offset is mapped to the light intensity correction space through the Moore-Penrose pseudo-inverse (J+ = (JTJ)-1JTJ+ = (JTJ)-1JT). Its normalized reference value I0 is set according to the Lambert-Beer law, using the transmitted light intensity of a reference light source at standard temperature and humidity as the benchmark. Linear superposition compensation of multi-physics field coupling errors is achieved through matrix multiplication. By integrating regularization constraints, nonlinear coupling modeling, dynamic parameter updates, and multivariable collaborative correction, this model breaks through the generalization bottleneck of traditional static regression models in time-varying environments and achieves the maintenance of submicron transmittance detection accuracy.

[0102] During the design and testing of step 4, to verify the comprehensive performance advantages of this step in noise suppression, environmental compensation, and transmittance accuracy, a comparative experiment was conducted with the existing method of static temperature and humidity compensation using a single-wavelength (550nm) reflectance model, a fixed threshold segmentation algorithm, and linear regression. The experimental hardware components included: a multispectral polarized light source (covering 400-1100nm, including the 850 / 940nm infrared bands), a high-resolution spectrometer (0.1nm resolution), a 3D topography scanner (laser confocal probe, 10nm longitudinal resolution), an infrared temperature measurement module (±0.2°C accuracy), a temperature and humidity sensor (SHT85, 100Hz sampling rate), an FPGA board (Xilinx Zynq UltraScale+), and an industrial computer (NVIDIA Jetson AGX Xavier). The layout was as follows: the polarized light source and spectrometer were coaxially aligned with the center of the lens, the 3D scanner was positioned at a 45° angle to the lens plane, and the temperature and humidity sensors were mounted around the lens. The FPGA board was connected to the industrial computer via a PCIe interface for real-time data processing.

[0103] In the experiment, 5 groups of coated lenses (IR ink aperture 1mm) were selected, and the ambient temperature and humidity fluctuation range was set to 20-40℃ / 30-70%RH to simulate industrial field conditions. Experimental Group A adopted the following step 4 scheme: Improved Stokes vector algorithm: Decomposition of noise through mirror reflection tensor, combined with curvature attention gated U-Net to suppress volume scattering noise (network input 512×512×4 Stokes matrix, ResNet-34 backbone); Multivariate linear regression compensation: Dynamic update of temperature and humidity weight matrix based on Kalman filter (initial coefficient 0.32 / -0.18); Experimental Group B: Use single wavelength (550nm) reflection model + fixed threshold segmentation algorithm, statically compensate temperature and humidity through linear regression. The experimental data are shown in Table 3:

[0104] Table 3 Comparison data record table

[0105]

[0106] Experiments show that by improving the Stokes algorithm and attention U-Net, Group A reduced the RMS specular reflection noise to 12.2dB (45.9dB in Group B) and increased the volume scattering suppression rate to 98.7% (71.3% in Group B), demonstrating significant multi-physics field coupling noise separation. Temperature and humidity compensation errors were controlled at 0.17°C / 1.2%RH (2.0°C / 6.0%RH in Group B), respectively, and the dynamic regression model effectively offset environmental disturbances. Although Group A experienced a slightly higher single-frame processing latency (15.4ms vs. 7.8ms), its overall accuracy met the requirements of industrial online monitoring, particularly in complex curved surfaces and high-humidity scenarios.

[0107] In step 5 of the above implementation, if Figure 6 As shown, the specific working method of step 5 is:

[0108] R1. Based on the 3D topography curvature gradient data output in step 1, the lens surface is divided into 5 to 10 non-uniform dynamic sub-regions using a curvature gradient threshold segmentation algorithm. The boundary of each region is adaptively determined by the curvature mutation point.

[0109] R2, load the transmittance statistical distribution data of the same type of lenses stored in the material history weight library, and use the Bayesian optimization algorithm to construct the sub-region transmittance mapping function f in the Zernike polynomial basis space. i

[0110] (Z), the formula is:

[0111]

[0112] In formula (3), f i (Z) represents the transmittance mapping function of the i-th sub-region; w ij is the dynamic weight coefficient of the j-th Zernike basis function in the ith subregion, which is dynamically adjusted by Gaussian process regression combined with historical prior distribution; Z j (ρ,θ) is the jth Zernike polynomial basis function, which is used to describe the spatial pattern of transmittance distribution, where ρ is the normalized radial coordinate and θ is the azimuth angle;

[0113] R3, for the basic transmittance data of each sub-area, the nonlinear weighted fusion strategy is used for fusion. The formula is:

[0114]

[0115] In formula (4), η fused Represents the transmittance distribution after fusion; a irepresents the fusion weight coefficient of the i-th sub-region, which is jointly calibrated by the sub-region curvature entropy value and the material absorption characteristics; N represents the number of dynamic sub-regions;

[0116] R4. In the edge area where the curvature gradient is greater than or equal to the threshold, the Monte Carlo importance sampling mechanism is used to generate a high-density interpolation point set through the Markov chain, and the radial basis function neural network is used to reconstruct the missing data, and finally the full-aperture transmittance distribution cloud map is output.

[0117] Among them, the curvature gradient threshold segmentation algorithm is based on differential geometry theory. It identifies the curvature mutation point by calculating the second-order derivative of the lens surface curvature tensor (Hessian matrix eigenvalue difference) and divides the lens into 5-10 non-uniform dynamic sub-regions. Its mathematical essence is to construct a curvature gradient field. (k1, k2 are the principal curvatures), and an adaptive bimodal threshold segmentation (such as the improved Otsu algorithm) is used to dynamically determine the boundary, thereby capturing the heterogeneity of transmittance distribution caused by lens deformation. Bayesian optimization and Zernike basis mapping use Gaussian process regression (GPR) to model the prior distribution of the historical material weight library, and the covariance kernel function is used to Capture the weight correlation of different Zernike basis functions and dynamically adjust the weight coefficient θ using Bayesian posterior update i,j . Zernike polynomial basis function Z nm The orthogonality of (ρ,θ) ensures the multi-scale decomposition of the transmittance spatial pattern, where the normalized radial coordinate ρ eliminates the lens size dependence through polar coordinate transformation, and the azimuth angle θ encodes the anisotropic scattering characteristics. The nonlinear weighted fusion strategy introduces the curvature entropy H c =-∑p(k)log p(k) is used as a measure of information complexity, and the weight is constructed by combining the material absorption characteristics (such as extinction coefficient β) The Monte Carlo importance sampling uses the Metropolis-Hastings algorithm to generate a Markov chain, and the curvature gradient is The edge area of ​​​​the sample points is guided to gather in the low confidence area through the transition probability T(x′|x)=min(1,p(x′) / p(x))T(x′|x)=min(1,p(x′) / p(x)), combined with the radial basis function (RBF) neural network f(x)=∑ω i φ(||xc i ||) to interpolate missing data, and its kernel function The local sensitivity ensures the fidelity of edge details.

[0118] During implementation, the curvature gradient threshold segmentation, the curvature mutation threshold Δκ is set to 0.05mm -1The boundary recognition algorithm used Canny edge detection combined with morphological closing operations. The number of dynamic partitions (5-10 partitions) was adaptively determined by K-means clustering of the curvature gradient histogram. The order of the Zernike polynomials in the Bayesian optimization module was set to 36 (radial order n = 6, angular order m = 6). The Gaussian process regression kernel function was Matern5 / 2, and hyperparameters were initialized using maximum a posteriori estimation (MAP). The curvature entropy calculation window radius for nonlinear weighted fusion was 2 mm, and the material absorption coefficient was calibrated using a spectrophotometer (PerkinElmer Lambda 950). Monte Carlo sampling generated 1000 interpolation points in the edge region. The Markov chain step size was set to 0.1 mm. The number of hidden nodes in the RBF neural network was set to 256, and the activation function was a Gaussian kernel (σ = 0.2 mm). During software processing, the FPGA receives 3D topography data in real time and performs dynamic partitioning. The GPU server calls the CUDA-accelerated Bayesian Optimization Library (GPyOpt) to iterate 100 times to generate a mapping function. The weight coefficients are integrated to solve the overdetermined equations through QR decomposition. The final transmittance cloud map is output through the OpenGL rendering engine, with a single processing cycle of ≤200ms. The specific software work is divided into the following four stages:

[0119] Dynamic sub-region division: Load the curvature gradient matrix, extract the curvature mutation edge through the non-maximum suppression algorithm, and use Delaunay triangulation to generate the sub-region boundary, ensuring that the curvature change rate of the adjacent region boundary is less than 20%. Typical lens ( The aspherical surface is divided into 7 sub-areas, and the edge area (curvature gradient ≥ 0.01mm-1) accounts for about 30%.

[0120] Bayesian Optimization and Basis Function Fitting: The Scikit-optimize library was used to implement Bayesian optimization. The maximum order of the Zernike polynomials was set to n = 8 (corresponding to 15 basis functions). The Gaussian process regression kernel function was the Matérn kernel (length scale 0.5 mm). The subregion mapping function was generated through 50 iterations. The historical weight library contained over 5000 transmittance distribution samples for lenses of the same material, with a prior distribution accuracy of ±0.3%.

[0121] Nonlinear weighted fusion: Calculate the curvature entropy value of each sub-region (formula: H = -∑p i lnp i , where p i is the curvature distribution probability density), combined with the material absorption coefficient (obtained by integrating the spectral data in step 4), to generate a fusion weight matrix (weight range 0.8-1.2), and use sparse matrix operations to accelerate the fusion process. The processing time for a single lens is less than 150ms.

[0122] Data enhancement in edge areas: In areas where the curvature gradient is ≥ the threshold, Monte Carlo sampling (sample number 10,000+) is initiated. Sampling points that conform to the edge curvature distribution are generated using the Metropolis-Hastings algorithm. Transmittance is reconstructed using RBF-NN (hidden layer number 50, Gaussian kernel σ = 0.3 mm). The data density in the edge area is increased to three times that of the central area, ensuring that the edge resolution of the cloud image reaches 0.1 mm / pixel.

[0123] During system integration, the data streams from steps 1-4 interact in real time through the HDF5 format, supporting dynamic expansion of the weight library for multiple lens models, with a historical data accumulation rate of 100 groups per hour.

[0124] This method overcomes the over-smoothing defect of traditional global fitting methods on complex curved lenses through dynamic sub-region division and Bayesian-Zernike fusion model; the Monte Carlo edge enhancement strategy significantly improves the ability to restore details in high curvature areas. Compared with the fixed grid interpolation method, it is more adaptable to local deformation and material unevenness in lens manufacturing, providing a high-resolution and highly robust transmittance distribution analysis solution for online quality inspection of optical lenses.

[0125] To verify the effectiveness of the design approach in Step 5, five sets of coated lenses (50mm diameter, 80-120mm radius of curvature) were selected. The ambient temperature and humidity were controlled within a range of 25±2°C / 50±5% RH, the light source wavelength was 550nm, and the light intensity stability was ±0.5%. A comparative experiment was conducted in this environment. The control group (Group B) employed a fixed 5×5 uniform grid, a global Zernike polynomial fit (with the first 25 terms as the basis), and bilinear interpolation of the edge regions. The experimental data are recorded in Table 4:

[0126] Table 4 Step 5 comparative experimental data record table

[0127]

[0128]

[0129] Experiments show that the transmittance RMS error of group A is reduced to 0.19% (group B 0.68%) through dynamic sub-region division and Bayesian-Zernike fusion, proving that non-uniform partitioning significantly improves the accuracy of complex surface fitting; the edge data density reaches 1200 points / mm 2 (Group B 470 points / mm 2 ), Monte Carlo sampling improves the detail restoration rate of high curvature areas by 155%. Although the processing delay of Group A is slightly higher (49.2ms vs 22.2ms), its maximum curvature deviation is controlled at 0.0022μm. -1 (Group B 0.0085μm -1), which verifies the technical advantages of this method in analyzing the edge morphology of coated lenses.

[0130] In step 6 of the above implementation, the inner loop calculates the transmittance standard deviation σ through a sliding window. If the transmittance standard deviation σ exceeds the preset threshold, the gradient descent algorithm is triggered to calculate the transmittance fluctuation gradient ΔG and iteratively update the photomultiplier tube. The calculation formula is:

[0131]

[0132] In formula (5), V HV is the PMT high voltage supply voltage, which is obtained by gradient descent iteration V HV ←V HV -δΔG updates the voltage, where the learning rate δ is adaptively adjusted using a root mean square back propagation mechanism corrected by a momentum factor; after the outer loop completes every five detection cycles, it extracts a historical transmittance data set and calculates the deviation of each data point from the cluster center using the Mahalanobis distance criterion. If the deviation exceeds three times the eigenvalue of the covariance matrix, it is determined to be an outlier and removed; then a state space model is constructed based on the filtered data, and a Kalman filter algorithm is used to predict the drift of the genetic algorithm population parameters. The crossover probability, mutation probability, and fitness weight matrix of the improved genetic algorithm module in step 2 are updated through a parameter injection mechanism, triggering population reinitialization to maintain global search capability; the inner and outer loops exchange transmittance standard deviation σ, photomultiplier tube gain coefficient, and genetic algorithm parameter drift data through a shared memory pool to form a cascade feedback control link.

[0133] The dual-loop coupled optimization architecture in step 6 is based on a collaborative optimization mechanism of dynamic feedback control and data-driven optimization. Its core principle is to achieve the adaptability of the transmittance monitoring system through the rapid response of the inner loop and the global optimization of the outer loop. The inner loop gradient adjustment mechanism uses a sliding window (e.g., window length N = 50) to calculate the transmittance standard deviation σ in real time and dynamically adjusts the learning rate η through the momentum factor-corrected root mean square back propagation (RMSProp). Its mathematical essence is: where v t is the second-order moment of the historical gradient, g2]t is the current gradient, and the adaptive weighting of the gradient direction is achieved through the exponential decay coefficients β1 = 0.9 and β2 = 0.99. This mechanism can suppress the voltage mutation of the high-voltage power supply of the photomultiplier tube (PMT) and avoid the degradation of the signal-to-noise ratio caused by gain overshoot. The outer loop optimization architecture adopts the Mahalanobis distance criterion (threshold λ = 3√λ_max, λ_max is the maximum eigenvalue of the covariance matrix) to eliminate outliers. Its technical principle is based on the multivariate normal distribution assumption. By calculating the Mahalanobis distance D_M = √{(x-μ)^T∑^{-1}(x-μ)} between the data point and the cluster center, robust filtering of non-Gaussian noise is achieved. The improved Kalman filter predicts the parameter drift Δθ of the genetic algorithm through the state space model.k , its state equation integrates the process noise covariance Q = diag (0.01, 0.05) and the observation noise covariance R = 0.1, and uses the Kalman gain matrix K k Dynamically correct crossover probability P c , mutation probability P m And the fitness weight matrix W f , thereby avoiding premature convergence of the population. Data exchange between the two rings is achieved through a shared memory pool. A ring buffer structure (capacity M = 1024) and an atomic lock mechanism are used to ensure low-latency (<1μs) synchronization of transmittance standard deviation, gain coefficient, and parameter drift, forming a cascade feedback control link.

[0134] During implementation, the hardware system for this step consists of a PMT module (Hamamatsu H10721-01, high-voltage power supply range 0-1500V), an FPGA signal processing board, and an industrial control machine. The inner loop is deployed in the real-time thread of the FPGA, collecting 2048 points of transmittance data every 10ms. A sliding window standard deviation algorithm (window step size Δ = 10, forgetting factor α = 0.95) is used to calculate σ. When σ > threshold 0.015, gradient descent is triggered: the PMT anode current I is read through the AD7606 data acquisition card. a , calculate the gain G = ΔI a / ΔV hv , iteratively update the voltage in The outer loop is deployed in the optimization thread of the industrial computer, and the historical data set D = {x i ∈R^5} (including transmittance mean, variance, skewness and other features), using the JBLAS library to calculate the covariance matrix Σ and perform eigendecomposition, eliminating the Mahalanobis distance After finding the outliers, construct the state space model:

[0135] θ k +1=Aθ k +w k

[0136] y k =Hθ k +v k

[0137] Where A = 0.98I (I is the unit matrix), H is the observation matrix, and the parameter drift Δθ is estimated by Kalman filtering k And inject the genetic algorithm module to adjust the crossover probability P c ∈[0.6,0.9], mutation probability P m ∈[0.001,0.1] and fitness weight Wf =[0.3,0.5,0.2]. The shared memory pool uses the PCIe DMA channel to implement data transmission between the FPGA and the industrial computer. It is equipped with dual-port RAM (address mapping 0x8000-0x8FFF) to achieve lock-free queue access. The data encapsulation protocol uses a custom binary format (packet header 0xAA55, CRC16 checksum).

[0138] During implementation, the inner-loop gradient adjustment effectively suppresses the instantaneous noise caused by PMT gain drift, and the outer-loop Mahalanobis distance-Kalman filter joint mechanism improves data quality and algorithm convergence. Compared with the traditional single-loop PID control method, the transmittance fluctuation is reduced by more than 40% under complex working conditions; the shared memory pool and parameter injection mechanism enable the genetic algorithm to break through the local optimal solution, and the global search efficiency is significantly improved, which is suitable for online quality control scenarios of high-precision optical devices.

[0139] In the implementation of the present invention, an online monitoring system for the light transmittance of an optical lens is provided. Figure 2 As shown, it includes: a three-dimensional morphology modeling module, a dynamic scanning parameter optimization module, a multi-dimensional data acquisition module, a noise suppression module, a data fusion module, a closed-loop optimization module and a visualization interaction module; wherein, the curvature gradient and normal vector data output by the three-dimensional morphology modeling module are transmitted in parallel to the curvature feature analysis unit of the noise suppression module and the dynamic partition control unit of the data fusion module through a high-speed data bus; after the dynamic scanning parameter optimization module receives the genetic algorithm parameter update instruction from the outer loop of the closed-loop optimization module, the generated dynamic partition scanning instruction set is sent to the FPGA timing controller of the multi-dimensional data acquisition module through the EtherCAT industrial Ethernet protocol to drive the tunable filter and the integrating sphere array to perform polarization state switching and spectrum acquisition; the multi-dimensional data acquisition module transmits the original spectral data with environmental sensor compensation to the FPGA timing controller of the multi-dimensional data acquisition module through the PCIe The 3D morphology data is transmitted to the pre-processing unit of the noise suppression module through the 3.0 interface, and the synchronous trigger signal is fed back to the point cloud registration engine of the 3D morphology modeling module. After the noise suppression module fuses the 3D morphology data with the original spectral data, the transmittance basic data that removes the specular reflection noise is directly written into the Zernike basis calculation unit of the data fusion module through the DMA channel. The full-aperture transmittance cloud generated by the data fusion module Figure 1On the one hand, the WebGL rendering engine of the visualization interaction module is pushed to the HDMI 2.0 interface to generate a three-dimensional heat map. On the other hand, the transmittance standard deviation sequence is input into the inner-loop control unit of the closed-loop optimization module to trigger the PMT gain dynamic adjustment signal based on gradient descent. At the same time, the outer-loop control unit generates the genetic algorithm parameter correction value through the improved Kalman filter of the historical data set screened by Mahalanobis distance, and transmits it back to the population initialization unit of the dynamic scanning parameter optimization module through the shared memory pool, forming a full-link cascade control system of "3D modeling-dynamic scanning-noise suppression-data fusion-closed-loop feedback". Finally, the visualization interaction module synchronously transmits the detection report and calibration parameters to the MES system and user terminal through Gigabit Ethernet.

[0140] The three-dimensional morphology modeling module is used to perform three-dimensional point cloud scanning on the lens surface based on the Gray code structured light projection and edge extraction algorithm, and generate three-dimensional morphology feature data including curvature gradient, normal vector and vertex coordinates through the ICP point cloud registration method; the dynamic scanning parameter optimization module is used to use the light intensity uniformity and scanning efficiency as multi-objective functions, and adopt the improved genetic algorithm to iteratively optimize the focal length of the zoom lens group, the polarizer angle and the aperture opening and closing parameters, generate a dynamic partition scanning instruction set and output it to the multidimensional data acquisition module; the multidimensional data acquisition module is used to trigger the tunable filter to switch the polarization state based on the dynamic partition scanning instruction set, synchronously collect the transmission spectrum, polarization state spectroscopic data and environmental sensor signals through the integrating sphere array, generate a multidimensional detection data set after wavelet denoising, and transmit it to the noise suppression module; the noise suppression module is used to receive the multidimensional data set and three-dimensional morphology data, separate the mirror reflection noise through the anisotropic Fresnel reflection tensor, adopt the curvature guided attention mechanism U-Net network to suppress the material scattering noise, and combine the multivariate linear regression model to obtain the three-dimensional morphology data. The system dynamically compensates for temperature and humidity drift and outputs denoised transmittance data to the data fusion module. The data fusion module is used to divide dynamic sub-regions based on curvature gradient clustering, fuse the transmittance data of each sub-region through Bayesian optimization in Zernike polynomial space, perform Monte Carlo sampling and radial basis function interpolation on high curvature edge regions, generate a full-aperture transmittance distribution cloud map, and transmit it to the closed-loop optimization module. The closed-loop optimization module includes an inner-loop control unit and an outer-loop control unit. The inner-loop control unit calculates the transmittance standard deviation in a 10ms cycle and dynamically adjusts the photomultiplier tube gain through the gradient descent method. After every five tests, the outer-loop control unit uses Mahalanobis distance to filter historical data, updates the genetic algorithm parameters through an improved Kalman filter, and feeds the optimization results back to the dynamic scanning parameter optimization module and the inner-loop control unit. The visualization interaction module is used to receive the transmittance cloud map and optimization parameters, render a three-dimensional thermal map through the WebGL engine, display the environmental parameters, noise suppression ratio, and calibration accuracy in real time, generate a test report, and transmit it to the user terminal.

[0141] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these specific embodiments are merely illustrative, and that those skilled in the art may omit, substitute, and modify the details of the methods and systems described above without departing from the principles and spirit of the present invention. For example, combining the above method steps to perform substantially the same functions in substantially the same manner to achieve substantially the same results falls within the scope of the present invention. Accordingly, the scope of the present invention is limited solely by the appended claims.

Claims

1. A method for online monitoring of light transmittance of an optical lens; characterized by: include: Step 1: Based on the Gray code structured light projection and edge extraction algorithm, the ICP point cloud registration method is used to construct a 3D surface model of the lens, and the 3D topographic feature data including curvature gradient, normal vector and vertex coordinates is output; Step 2: Input the three-dimensional topographic feature data and lens aperture into an improved genetic algorithm module, use light intensity uniformity and scanning efficiency as multi-objective functions, iteratively calculate the focal length of the zoom lens group, the polarizer angle, and the iris opening and closing parameters, and generate a dynamic partition scanning instruction set; Step 3: driving the tunable filter to generate multi-wavelength polarized light according to the dynamic partition scanning instruction set, and synchronously collecting the transmission spectrum, polarization state light intensity signal and environmental sensor data through the integrating sphere array to generate a multi-dimensional detection data set; Step 4: Based on the multidimensional detection dataset, an improved Stokes vector algorithm is used to decompose the specular reflection noise component, and a U-Net network with an attention mechanism is used to suppress material scattering noise. A multivariate linear regression model is used to compensate for the light intensity offset caused by temperature drift and humidity changes in real time, and basic transmittance data is output. Step 5: Divide the lens into 5 to 10 dynamic sub-regions based on the curvature gradient, use a Bayesian optimization algorithm to load the material history weight library, perform nonlinear weighted fusion on the transmittance basic data in a Zernike polynomial basis space, and enhance the edge area data density through Monte Carlo sampling to generate a full-aperture transmittance distribution cloud map; Step 6: Construct a dual-loop coupled optimization architecture. The inner loop monitors the transmittance standard deviation with a 10ms period and adjusts the photomultiplier tube gain using the gradient descent method. After the outer loop completes five detections, it filters historical data based on the Mahalanobis distance criterion and uses an improved Kalman filter to update the genetic algorithm population parameters.

2. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The working principle of the ICP point cloud registration method is: A Gray code structured light projection module projects multi-frequency phase-coded stripes onto the lens surface, and the absolute phase distribution map is solved by a four-step phase shift interferometry method. Phase jump noise is eliminated by combining a multi-frequency heterodyne phase unwrapping algorithm. The pixel coordinates of the phase image contour edge are extracted based on the Canny edge detection operator, and the edge position is corrected twice using the Zernike positioning algorithm to generate an edge coordinate sequence. Based on the baseline calibration parameters and epipolar correction matrix of the binocular stereo vision system, the edge coordinate sequence is converted into a three-dimensional discrete point cloud through a triangulation conversion method; The rigid transformation matrix of the three-dimensional discrete point cloud is obtained by SVD decomposition, the LM optimization algorithm is used to iteratively minimize the point cloud registration residual, and the RANSAC outlier removal mechanism is used to improve the registration accuracy; The registered multi-source point cloud is input into the Poisson surface reconstruction algorithm, and the lens continuous surface model is generated through octree space partitioning and gradient field optimization; The neighborhood covariance matrix of surface vertices is calculated based on principal component analysis. The curvature gradient tensor and normal vector direction are solved by eigenvalue decomposition, and a three-dimensional morphological feature dataset containing curvature gradient, normal vector and vertex coordinates is constructed.

3. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The improved genetic algorithm module includes a parameter encoding unit, an adaptive crossover mutation unit, a multi-objective fitness evaluation unit, a constraint processing unit, an elite retention unit and a dynamic partition optimization unit; the parameter encoding unit is used to map the focal length of the zoom lens group, the polarizer angle, and the aperture opening and closing parameters into chromosome gene sequences through a real number encoding mechanism, and output an initial population matrix; the adaptive crossover mutation unit is used to dynamically adjust the crossover probability and the mutation probability through a Sigmoid function, combine the roulette wheel selection strategy to screen the parent generation individuals, trigger a two-point crossover operation to generate the offspring chromosome, and use the Gaussian perturbation mechanism to perform local perturbation optimization on the mutant genes; the multi-objective fitness evaluation unit is used to construct a binocular based on light intensity uniformity and scanning efficiency. The objective function is to generate a Pareto rank and a fitness value through the non-dominated sorting and crowding calculation of the NSGA-II algorithm; the constraint processing unit is used to process the focal length range and the polarization angle limit through the penalty function method, and embed the focal length range and the polarization angle limit into the fitness function through the Lagrange multiplier method to generate a feasible solution set; the elite retention unit is used to adopt a steady-state selection strategy, retain the elite individuals in the current population through the tournament mechanism, and output the optimized offspring population; the dynamic partition optimization unit is used to divide the scanning area based on the lens curvature gradient distribution data through the K-means clustering algorithm, and generate a dynamic partition scanning instruction set containing the focal length, polarization angle, and aperture parameter time series sequence in combination with the scanning path ant colony optimization algorithm.

4. The method for online monitoring of light transmittance of an optical lens according to claim 1, wherein: The working principle of step 3 is as follows: first, a tunable filter is triggered to switch the target wavelength through a dynamic partition scanning instruction set, a sinusoidal voltage signal is loaded on the electro-optical modulator to generate linearly polarized light, and the elliptical polarization state is dynamically adjusted through the delayed phase matrix of the birefringent crystal; six groups of sub-spheres in the integrating sphere array are synchronously triggered to capture the transmission spectrum in a time-sharing manner, wherein each group of sub-spheres has a built-in four-channel detector, which respectively collects Stokes parameters in polarization directions of 0°, 45°, 90°, and 135°, and transmits the spectral signal to a high-speed ADC module through an optical fiber coupler; simultaneously, temperature, humidity, and air pressure data from the environmental sensor are read, a Kalman filter algorithm is used to eliminate wavelength offset errors caused by temperature drift, and the polarization state light intensity, spectral intensity, and environmental parameters are aligned according to timestamps based on the IEEE 1588 clock protocol to construct a three-dimensional transmittance tensor containing wavelength, polarization state, and environmental factors; During the acquisition process, if the light intensity of any sub-sphere is detected to exceed the preset threshold, the missing signal is reconstructed based on the data of the adjacent sub-spheres through the radial basis function interpolation algorithm.

5. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The improved Stokes vector algorithm is based on the Stokes parameter matrix S in the multidimensional detection data set. total =[S0,S1,S2,S3] T , a polarization reflection-transmission separation model including lens anisotropy and multi-wavelength coupling effects is constructed. The working principle of the polarization reflection-transmission separation model is as follows: S1. Introduce the anisotropic Fresnel reflection tensor R(θ,λ,n) of the lens surface, establish the polarization state mapping relationship of the reflection noise, and calculate the mirror reflection Stokes vector S reflected , the calculation formula is: S reflected =R(θ,λ,n)·S incident +D roughness (1) In formula (1), S incident represents the Stokes vector of the incident light; θ is the incident angle, which is used to quantify the geometric relationship between the light and the lens surface; λ is the wavelength of light, which determines the spectral dependence of the reflectivity; n is the refractive index tensor, which characterizes the anisotropic optical properties of the lens material; Represents the anisotropic Fresnel reflection tensor, which is used to quantify the polarization crosstalk effect of specular reflection. Its matrix element R ss 、R sp 、R ps and R pp represents the cross-polarization reflectivity; represents the surface roughness scattering noise term, which is used to reflect the polarization state disturbance of non-specular scattered light, where σ represents the root mean square value of surface roughness, which is used to determine the scattering noise intensity; L c represents the surface topography autocorrelation length; k is the wave number, which is inversely proportional to the wavelength λ and is used to quantify the sensitivity of the scattering effect to the wavelength; S2. Extract the transmission component S transmitted Eliminate the material body scattering noise, the transmission component is based on the coupling model of birefringence effect and scattering phase function, and define the polarization scattering correlation tensor P scatter : In formula (2), β is the scattering angle and φ is the azimuth angle; S3, the mirror reflection Stokes vector S reflected From the Stokes parameter matrix S total Subtract the mid-vectors to get the transmitted Stokes vector S transmitted , the transmitted Stokes vector S is segmented by the polarization threshold algorithm transmitted The residual volume scattering noise in the image is suppressed.

6. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The working principle of the attention mechanism U-Net network is as follows: first, the local texture features and global scattering patterns of the light intensity distribution map are extracted through the multi-scale residual convolution block on the encoder side. The multi-scale residual convolution block uses an expanded convolution kernel to capture the scattering noise characteristics of different scales; then, based on the three-dimensional morphology curvature gradient data of step 1, a curvature-guided attention gating mechanism is embedded in the jump connection layer, and the association weights of the curvature features and the light intensity features are calculated through the covariance matrix to dynamically suppress the characteristic responses of the areas with high incidence of scattering noise; on the decoder side, a polarization-wavelength dual-branch deconvolution module is used. Branch one analyzes the spatial distribution of the Stokes parameter, and branch two solves the wavelength-scattering phase function mapping relationship. The dual-branch feature map is fused in combination with the gated recurrent unit; finally, the binary noise distribution map is output through the scattering noise mask generation layer, and the original light intensity data is differentially operated with the noise mask to obtain the de-scattered transmitted light intensity matrix.

7. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The working principle of the multivariate linear regression model is as follows: First, based on the temperature data T, humidity data H and descattered transmitted light intensity I transmitted The mapping relationship between environmental variables and light intensity offset is solved by ridge regression regularization algorithm: ΔI=β0+β1T+β1H+β3TH, where TH is the temperature-humidity interaction term, which is used to quantify the nonlinear influence of the synergistic effect of temperature and humidity on the refractive index; β0, β1, β1 and β3 represent the regression intercept term, temperature coefficient, humidity coefficient and cross-coupling coefficient, respectively; then, the training set is updated in real time based on the sliding window mechanism, and the regression coefficient β is dynamically corrected using the recursive least squares method. i , compensate for the time-varying characteristics of the environmental drift; finally, the offset ΔI predicted by the regression is reversely corrected to the transmitted light intensity through the ambient light intensity coupling matrix C(T,H), and the compensated transmittance basic data η=(I transmitted -ΔI) / I reference , where I reference Indicates the reference light intensity value, which is used to normalize the calculation of transmittance.

8. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The working method of step 5 is: R1. Based on the 3D topography curvature gradient data output in step 1, the lens surface is divided into 5 to 10 non-uniform dynamic sub-regions using a curvature gradient threshold segmentation algorithm. The boundary of each region is adaptively determined by the curvature mutation point. R2, load the transmittance statistical distribution data of the same type of lenses stored in the material history weight library, and use the Bayesian optimization algorithm to construct the sub-region transmittance mapping function f in the Zernike polynomial basis space. i (Z), the formula is: In formula (3), f i (Z) represents the transmittance mapping function of the i-th sub-region; w ij is the dynamic weight coefficient of the j-th Zernike basis function in the ith subregion, which is dynamically adjusted by Gaussian process regression combined with historical prior distribution; Z j (ρ,θ) is the J-th Zernike polynomial basis function, which is used to describe the spatial pattern of transmittance distribution, where ρ is the normalized radial coordinate and θ is the azimuth angle; R3, for the basic transmittance data of each sub-area, the nonlinear weighted fusion strategy is used for fusion. The formula is: In formula (4), η fused Represents the transmittance distribution after fusion; a i represents the fusion weight coefficient of the i-th sub-region, which is jointly calibrated by the sub-region curvature entropy value and the material absorption characteristics; N represents the number of dynamic sub-regions; R4. In the edge area where the curvature gradient is greater than or equal to the threshold, the Monte Carlo importance sampling mechanism is used to generate a high-density interpolation point set through the Markov chain, and the radial basis function neural network is used to reconstruct the missing data, and finally the full-aperture transmittance distribution cloud map is output.

9. The method for online monitoring of light transmittance of an optical lens according to claim 1, characterized in that: The inner loop calculates the transmittance standard deviation σ through a sliding window. If the transmittance standard deviation σ exceeds a preset threshold, the gradient descent algorithm is triggered to calculate the transmittance fluctuation gradient ΔG and iteratively update the photomultiplier tube. The calculation formula is: In formula (5), V HV is the PMT high voltage supply voltage, which is obtained by gradient descent iteration V HV ←V HV -δΔG updates the voltage, where the learning rate δ is adaptively adjusted using a momentum-corrected RMS backpropagation mechanism; After the outer loop completes five detection cycles, it extracts the historical transmittance data set and calculates the deviation of each data point from the cluster center using the Mahalanobis distance criterion. If the deviation exceeds three times the eigenvalue of the covariance matrix, it is determined to be an outlier and removed. Then, a state space model is constructed based on the filtered data, and the Kalman filter algorithm is used to predict the drift of the genetic algorithm population parameters. The crossover probability, mutation probability, and fitness weight matrix of the improved genetic algorithm module in step 2 are updated through the parameter injection mechanism, triggering population reinitialization to maintain global search capabilities. The inner and outer loops exchange transmittance standard deviation σ, photomultiplier tube gain coefficient, and genetic algorithm parameter drift data through a shared memory pool to form a cascade feedback control link.

10. An online monitoring system for light transmittance of an optical lens, characterized by: An online monitoring method for light transmittance of an optical lens as claimed in any one of claims 1 to 9, comprising: a three-dimensional topography modeling module, a dynamic scanning parameter optimization module, a multi-dimensional data acquisition module, a noise suppression module, a data fusion module, a closed-loop optimization module, and a visualization interaction module; The three-dimensional shape modeling module is used to perform three-dimensional point cloud scanning on the lens surface based on Gray code structured light projection and edge extraction algorithm, and generate three-dimensional shape feature data including curvature gradient, normal vector and vertex coordinates through ICP point cloud registration method; The dynamic scanning parameter optimization module is used to iteratively optimize the focal length of the zoom lens group, the polarizer angle, and the aperture opening and closing parameters using an improved genetic algorithm with light intensity uniformity and scanning efficiency as multi-objective functions, generate a dynamic partition scanning instruction set, and output it to the multi-dimensional data acquisition module; The multidimensional data acquisition module is used to trigger the tunable filter to switch the polarization state based on the dynamic partition scanning instruction set, synchronously collect the transmission spectrum, polarization state spectroscopic data and environmental sensor signals through the integrating sphere array, generate a multidimensional detection data set after wavelet denoising, and transmit it to the noise suppression module; The noise suppression module is used to receive the multidimensional dataset and three-dimensional topography data, separate the specular reflection noise through the anisotropic Fresnel reflection tensor, suppress the material scattering noise using the curvature-guided attention mechanism U-Net network, dynamically compensate for temperature and humidity drift in combination with a multivariate linear regression model, and output the denoised transmittance data to the data fusion module; The data fusion module is used to divide the dynamic sub-regions based on curvature gradient clustering, fuse the transmittance data of each sub-region through Bayesian optimization in Zernike polynomial space, perform Monte Carlo sampling and radial basis function interpolation on the high curvature edge area, generate a full-aperture transmittance distribution cloud map, and transmit it to the closed-loop optimization module; The closed-loop optimization module includes an inner-loop control unit and an outer-loop control unit; the inner-loop control unit calculates the transmittance standard deviation in a 10ms period and dynamically adjusts the photomultiplier tube gain using a gradient descent method; the outer-loop control unit uses Mahalanobis distance to filter historical data after every five measurements, updates the genetic algorithm parameters using an improved Kalman filter, and feeds the optimization results back to the dynamic scanning parameter optimization module and the inner-loop control unit; The visualization interaction module is used to receive the transmittance cloud map and optimization parameters, render the three-dimensional thermal map through the WebGL engine, display the environmental parameters, noise suppression ratio and calibration accuracy in real time, generate a test report and transmit it to the user terminal.

Images