Method and system for rapid repositioning detection of different aspherical optical paths on air floating platform

Abstract

The application relates to the technical field of optical detection, and discloses a quick reset detection method and system for different aspheric optical paths on an air floating platform. In the first time of assembly and adjustment, the absolute mechanical position sequence and the corresponding digital interference pattern sequence of an auxiliary compensation optical path are synchronously collected, an aspheric optical path position and interference feature mapping data set is constructed, a multilayer perception machine regression network is trained based on the mapping data set to obtain a position compensation prediction model, an initial interference pattern is collected after the optical path is disassembled and reset, features are extracted, the model is input to obtain a target position compensation vector, the target position compensation vector is converted into an adjustment instruction to drive the optical path to complete precise reset, manual repeated trial and error adjustment is not needed, experience dependence is effectively eliminated, and the reset efficiency and assembly and adjustment precision of the aspheric detection optical path are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of optical detection technology, and in particular to a rapid resetting detection method and system for different aspherical optical paths on an air-bearing platform. Background Technology

[0002] The surface accuracy inspection of aspherical optical components is a core part of optical processing and quality control. Existing aspherical inspection optical paths often adopt a reset method based on the reference hole system of an air-bearing platform and the recording of the position parameters of the adjustment frame. This method can quickly reproduce the hardware installation position when inspecting similar components in the future by pre-reserving the mechanical position readings of the adjustment frame, reducing the amount of repetitive work in optical path construction. It has been widely used in various aspherical zero-position and non-zero-position inspection scenarios.

[0003] In actual batch testing of aspherical components, after completing the optical path hardware reset, the first interference pattern acquisition by the testing personnel often reveals that the interference fringes do not meet the testing accuracy requirements. This necessitates manual, iterative adjustments to the pose of the auxiliary compensation element until the fringe state meets the requirements. The root cause of this problem lies in the fact that existing reset methods only record the mechanical structural physical quantities of the adjustment frame, failing to establish a quantitative mapping relationship with the actual optical performance parameters of the optical path. Mechanical clearances from repeated disassembly and assembly of the adjustment frame, long-term wear, and nonlinear deviations resulting from the coupling of individual processing tolerances of the same model of auxiliary compensation element cannot be directly eliminated through mechanical positioning. Furthermore, existing technology has not constructed a predictive model corresponding to the pose deviation of the auxiliary compensation element and the characteristics of the interference fringes, making it impossible to directly obtain the target adjustment amount. This problem leads to deviations in the testing benchmark of the same batch of aspherical components due to the operator's assembly and adjustment experience. Some qualified components may be mistakenly judged as unqualified and discarded, and unqualified components may also flow into subsequent processing stages, increasing processing and testing costs and slowing down the overall production and delivery pace of optical components. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a rapid resetability detection method and system for different aspherical optical paths on an air-bearing platform. By constructing a quantitative mapping relationship between the absolute mechanical pose sequence of auxiliary compensation elements and the interference fringe feature set, and combining a reset deviation prediction model to automatically quantify and compensate for the pose deviation of the adjustment frame, the invention aims to eliminate the influence of mechanical clearance, wear, and individual tolerance coupling errors of components on the reset accuracy of the optical path. This enables rapid and high-precision reset of the optical path during the batch testing of aspherical components, reduces the cost of manual intervention, and improves the consistency of testing benchmarks and testing efficiency.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A rapid resettable detection method for different aspherical optical paths on an air-floating platform includes the following steps:

[0007] S1: During the initial optical path assembly and adjustment stage, the absolute mechanical pose sequence and the digital interferogram sequence are acquired. The interference fringe feature set is extracted based on the digital interferogram sequence, and the absolute mechanical pose sequence and the interference fringe feature set are fused into an aspherical optical path pose and interference feature mapping dataset.

[0008] Further, step S1 includes:

[0009] S11: During the initial detection of the optical path setup, the absolute physical coordinate parameters of each adjustment frame in the auxiliary compensation optical path assembly are read by a high-precision displacement sensor under multiple fine-tuning states to obtain the absolute mechanical pose sequence.

[0010] S12: In each corresponding fine-tuning state, the laser interferometer is synchronously driven to collect the optical path interference spot data at the corresponding moment, and generate a digital interference pattern sequence with stable power and determined shape.

[0011] S13: Perform image filtering and noise reduction and fringe extraction feature construction on the digital interferometric pattern sequence to extract aspherical interference fringe information and stitch them together to form an interference fringe feature set.

[0012] Specifically, the method for extracting interference fringe feature sets based on digital interferometric pattern sequences is as follows:

[0013] S131: The Gaussian filtering algorithm is used to perform convolution operation on the digital interferogram sequence to remove ambient light leakage noise and high-frequency noise from laser speckle, resulting in a smoothed interferogram sequence after background smoothing.

[0014] S132: Extract the alternating light and dark interference contour edges from the smooth interference pattern sequence, calculate the fringe line density of the separated interference contour edges and the curvature deformation parameters of the main fringes, and merge them to form an interference fringe feature set.

[0015] S14: Based on the simultaneous data labeling pairing criterion, the absolute physical coordinate parameters in the absolute mechanical pose sequence are mapped and combined with the corresponding fringe features in the interference fringe feature set to generate an aspherical optical path pose and interference feature mapping dataset with digital mapping scale labels.

[0016] S2: Train a multilayer perceptron regression initial network based on the aspherical optical path pose and interference feature mapping dataset, and construct and output a pose compensation prediction model that is applicable to both non-zero and zero-position detection states.

[0017] Further, step S2 includes:

[0018] S21: Extract the data row content of the aspherical optical path pose and interference feature mapping dataset according to the preset distribution ratio threshold, and divide it into feature input data and mechanical pose ground truth label data.

[0019] S22: Construct a neural network with nested logic of neuron hierarchy, specifically including introducing a neuron processing input layer, a fully connected feedforward hidden layer, and a multidimensional coordinate calibration output layer to obtain a multilayer perceptron regression initial network.

[0020] S23: Input feature input data into the multilayer perceptron regression initial network to perform forward propagation calculation, generate pose prediction results obtained by network inference, calculate the Euclidean distance discrete deviation amplitude between the pose prediction results and the mechanical pose true value label data, and form the mean square error to evaluate the loss from the Euclidean distance discrete deviation amplitude.

[0021] S24: The mean squared error is backpropagated using the gradient descent algorithm to evaluate the loss and update the neuron connection weight coefficients of the initial network of the multilayer perceptron regression. After the loss is evaluated to decrease and stabilize, the fixed neuron connection weight coefficients are extracted to generate the pose compensation prediction model.

[0022] S3: After the detection optical path is physically disassembled and re-reset, the initial digital interference pattern in the reset state is acquired and the reset interference fringe features are extracted. The reset interference fringe features are then input into the pose compensation prediction model to calculate the target pose compensation vector.

[0023] Further, step S3 includes:

[0024] S31: After the auxiliary compensation optical path component is physically removed and the hardware is replaced and reassembled based on the reference hole system, a collimated beam is projected using a laser interferometer and the reflected light from the aspherical surface is received to generate the initial digital interference pattern in the reset state.

[0025] S32: Based on the initial digital interferometric pattern recognition, extract pixel-level image data attribute information specific to the detection environment to generate reset interference fringe features.

[0026] Specifically, the method for extracting the features of reset interference fringes is as follows:

[0027] S321: A two-dimensional image enhancement program is used to perform contrast enhancement and adaptive threshold segmentation pixel retention processing on the initial digital interference pattern, separating the background dark area from the effective area of ​​the interference bright and dark fringes, and obtaining a pure interference fringe pattern that eliminates environmental background interference factors.

[0028] S322: Based on the pure interference fringe pattern, perform morphological line refinement calculations to extract the horizontal and vertical central skeleton contour lines, statistically analyze the frequency domain distribution density of the horizontal and vertical central skeleton contour lines in the image spatial domain and the fringe surface deflection index, and then connect and splice them to form the reset interference fringe feature.

[0029] S33: The reset interference fringe features are transmitted to the pose compensation prediction model through the call interface to perform forward offset calculation and analysis. The pose compensation prediction model outputs the multi-dimensional parameters of the optical coordinate center displacement deviation required for the detection system to reach the optimal detection zero position standard. The target pose compensation vector is obtained by integrating the multi-dimensional parameters.

[0030] S4: Perform a three-dimensional coordinate system decoupling transformation on the target pose compensation vector to construct the adjustment frame displacement operation command. Use the adjustment frame displacement operation command to drive the auxiliary compensation optical path component to perform position fine-tuning action, and complete the interference detection of the target aspherical reflector surface shape.

[0031] Further, step S4 includes:

[0032] S41: Classify and decouple the target pose compensation vector according to the pre-configured aspherical system category to form the adjustment frame displacement operation command specific to the measured component of the corresponding category.

[0033] S42: The drive signal corresponding to the displacement operation command of the adjustment frame is used to directly control all the linkage elements in the drive auxiliary compensation optical path assembly to produce precise position fine-tuning and angle compensation. This is to supplement and correct the mechanical tolerance adjustment deviation left after long-term use and physical reassembly, so that the quality of the detection optical path formed after the control fine-tuning correction meets the optical path acceptance criteria of precision aspherical surfaces, and completes the interference detection of the surface shape of the target aspherical reflector.

[0034] This invention also provides a rapid resetability detection system for different aspherical optical paths on an air-floating platform, which is used to implement the aforementioned rapid resetability detection method for different aspherical optical paths on an air-floating platform. The system includes:

[0035] The mapping dataset construction module is used to acquire the absolute mechanical pose sequence and the digital interferogram sequence during the initial optical path assembly and adjustment stage, extract the interference fringe feature set based on the digital interferogram sequence, and fuse the absolute mechanical pose sequence and the interference fringe feature set into an aspherical optical path pose and interference feature mapping dataset.

[0036] The model training module is used to train a multilayer perceptron regression initial network based on the aspherical optical path pose and interference feature mapping dataset, and to construct and output a pose compensation prediction model that is applicable to both non-zero and zero-position detection states.

[0037] The compensation vector calculation module is used to collect the initial digital interference pattern in the reset state and extract the reset interference fringe features after the detection optical path is physically disassembled and re-reset. The reset interference fringe features are then input into the pose compensation prediction model to calculate the target pose compensation vector.

[0038] The reset detection execution module is used to perform a three-dimensional coordinate system decoupling transformation on the target pose compensation vector to construct an adjustment frame displacement operation command. The adjustment frame displacement operation command is used to drive the auxiliary compensation optical path component to perform a position fine-tuning action, thereby completing the interference detection of the target aspherical reflector surface shape.

[0039] Compared to existing technologies, the advantages of this invention are as follows: This invention simultaneously acquires the absolute mechanical pose data and corresponding digital interferometry patterns of the auxiliary compensation optical path components during the initial optical path assembly and adjustment stage, constructing a dataset mapping the pose and interference features of the aspherical optical path. This transforms the implicit experience from manual assembly and adjustment into structured, reusable data. Furthermore, by using absolute physical coordinate parameters as the pose recording benchmark, it is unaffected by zero-position drift after the adjustment frame is disassembled and reassembled, possessing comparability and reproducibility across assembly and disassembly cycles. The pose compensation prediction model trained based on the mapping dataset can generalize the discrete mapping relationship between mechanical pose and interference features into a continuous nonlinear prediction capability, eliminating the subjectivity and uncertainty caused by the reliance on experience-based judgment in traditional manual assembly and adjustment. The network structure employing two fully connected hidden layers ensures fitting ability while avoiding overfitting, resulting in excellent generalization performance. After the optical path is disassembled and reset, only a single initial interference pattern needs to be collected to quickly calculate the target pose compensation vector. This reduces the traditional manual multi-round trial and error reset process to a single orientation adjustment. It is suitable for the testing needs of various aspherical mirrors such as parabolic surfaces, coaxial hyperboloids, and freeform surfaces. With the help of a closed-loop verification mechanism, the reset accuracy is guaranteed, significantly reducing the optical path reset time and effectively improving the assembly and adjustment efficiency and testing accuracy of optical components. It is applicable to various optical performance testing scenarios and promotes the automation and standardization of optical testing operations. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the rapid resettable detection method for different aspherical optical paths on the air-floating platform in this invention;

[0042] Figure 2 This is a schematic diagram illustrating the deployment of a high-precision displacement sensor for a parabolic reflector in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram illustrating the deployment of a high-precision displacement sensor for a coaxial hyperboloid mirror in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram illustrating the deployment of a high-precision displacement sensor for a freeform surface mirror in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of the pose data point data structure in an embodiment of the present invention;

[0046] Figure 6 This is a schematic diagram illustrating synchronous acquisition and data alignment in an embodiment of the present invention;

[0047] Figure 7 This is a schematic diagram of the Gaussian filter convolution kernel sliding in an embodiment of the present invention;

[0048] Figure 8 This is a schematic diagram of interference contour extraction using Canny edge detection in an embodiment of the present invention;

[0049] Figure 9 This is a schematic diagram of the multilayer perceptron regression network structure in an embodiment of the present invention;

[0050] Figure 10 This is a schematic diagram illustrating the effects of two-dimensional image enhancement and region segmentation in an embodiment of the present invention;

[0051] Figure 11 This is a functional block diagram of the rapid resetability detection system for different aspherical optical paths on the air-float platform in this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1:

[0054] Please see Figure 1 As shown, this embodiment provides a rapid resettable detection method for different aspherical optical paths on an air-floating platform, including:

[0055] S1: In the initial optical path assembly and adjustment stage, the absolute mechanical pose sequence and the digital interferogram sequence are obtained. The interference fringe feature set is extracted based on the digital interferogram sequence, and the absolute mechanical pose sequence and the interference fringe feature set are fused into an aspherical optical path pose and interference feature mapping dataset.

[0056] Further, step S1 includes:

[0057] S11: During the initial detection of the optical path setup, the absolute physical coordinate parameters of each adjustment frame in the auxiliary compensation optical path assembly are read through a high-precision displacement sensor under multiple fine-tuning states to obtain the absolute mechanical pose sequence;

[0058] Further, step S11 includes:

[0059] S111: After the initial installation of the interferometer and auxiliary compensation optical path assembly on the air-bearing vibration isolation platform is completed, high-precision displacement sensors are installed at the motion joints of each adjustment frame in the auxiliary compensation optical path assembly. The high-precision displacement sensors include a linear encoder deployed on the lifting shaft of the optical height adjustment frame, an angle encoder deployed on the pitch and yaw joints of the optical three-dimensional precision adjustment frame, and a linear grating ruler deployed on the push rod of the precision translation adjustment frame.

[0060] See Figure 2 , Figure 3 and Figure 4 This is a schematic diagram of the deployment of a high-precision displacement sensor provided in an embodiment of this application. Figure 2 This is a testing configuration for parabolic reflectors, comprising a plane reflector. The plane reflector is mounted on an optical height adjustment frame and an optical three-dimensional precision adjustment frame, connected sequentially from bottom to top. The optical height adjustment frame is used to adjust the center height of the plane reflector to be equal to the center of the interferometer's output beam. The optical three-dimensional precision adjustment frame is used to fine-tune the pitch and yaw angles of the plane reflector, ensuring its normal is parallel to the optical axis. Since the plane mirror only needs to provide collimated light during parabolic surface testing and does not need to move along the optical axis, this configuration meets the testing requirements. Figure 3 The testing configuration for a coaxial hyperboloid mirror includes a spherical compensating mirror. This spherical compensating mirror is mounted on an optical height adjustment frame, an optical three-dimensional precision adjustment frame, and a precision translation adjustment frame, connected sequentially from bottom to top. The precision translation adjustment frame is positioned along the optical axis and is used to precisely adjust the longitudinal distance between the spherical compensating mirror and the interferometer focal point to meet the null position conditions for hyperboloid testing. Figure 4 This is a testing configuration for freeform surface mirrors, including a computer-generated hologram (CGH) compensating mirror. The CGH compensating mirror is mounted on a five-dimensional optical adjustment frame. This five-dimensional optical adjustment frame enables translation along the X, Y, and Z axes, as well as pitch and yaw adjustments around the X and Y axes, to meet the complex spatial pose requirements of the CGH compensating mirror in freeform surface testing.

[0061] S112: During the initial precision assembly and adjustment of the optical path by the testing personnel, whenever the testing personnel complete a fine-tuning operation on any adjustment frame in the auxiliary compensation optical path assembly and hold it stationary, all high-precision displacement sensors are simultaneously triggered to acquire a pose reading. A fine-tuning operation refers to a single adjustment action by which the testing personnel manually rotate the fine-tuning knob on the adjustment frame, causing a change in angle or displacement of the auxiliary compensation element. The result of each pose reading acquisition is a set of pose data points containing the absolute physical quantities of all adjustment frame degrees of freedom. The dimension of these pose data points is determined by the total number of degrees of freedom of the adjustment frames in the auxiliary compensation optical path assembly. When the device being tested is a parabolic reflector, the auxiliary compensation optical path assembly includes an optical height adjustment frame and an optical three-dimensional precision adjustment frame. The pose data points contain absolute physical coordinate parameters for three degrees of freedom: vertical lift displacement, pitch angle, and yaw angle. When the device being tested is a coaxial hyperboloid mirror, the auxiliary compensation optical path assembly includes an optical height adjustment frame, an optical three-dimensional precision adjustment frame, and a precision translation adjustment frame. The pose data points include absolute physical coordinate parameters with four degrees of freedom: vertical lift displacement, pitch angle, yaw angle, and longitudinal translation distance along the optical axis. When the device being tested is a freeform surface mirror, the auxiliary compensation optical path assembly includes a five-dimensional optical adjustment frame. The pose data points include absolute physical coordinate parameters with five degrees of freedom: translation distance along the X-axis, translation distance along the Y-axis, translation distance along the Z-axis, pitch angle around the X-axis, and yaw angle around the Y-axis. See also... Figure 5 This is a schematic diagram of the pose data point data structure provided in an embodiment of this application. For example... Figure 5 As shown, the diagram visually illustrates the differences in the dimensional composition of pose data points when detecting different types of aspherical mirrors. The top general structure shows that the data points consist of multiple degrees of freedom parameters. The branches below clearly distinguish three application scenarios: the left "parabolic mirror" corresponds to 3 degrees of freedom (vertical lift, pitch, and yaw); the middle "coaxial hyperboloid mirror" adds 4 degrees of freedom for longitudinal translation along the optical axis; and the right "freeform mirror" includes 5 degrees of freedom: XYZ-axis translation and rotation around the XY axis. Different fill textures (such as diagonal lines, grids, and solid colors) represent different types of physical quantities (angle or displacement). This structured definition allows the data acquisition program to dynamically adjust the sensor reading channels and the number of columns in the data matrix according to the type of device under test, ensuring that the subsequently constructed "absolute mechanical pose sequence" strictly matches the specific optical path configuration in terms of dimensions, laying a data foundation for the generalized software architecture of multi-type aspherical detection systems.

[0062] S113: Arrange all pose data points acquired sequentially during the initial optical path precision assembly process according to the acquisition time order to form an absolute mechanical pose sequence. See also Figure 6 This is a schematic diagram of the absolute mechanical pose sequence matrix provided in an embodiment of this application. Figure 6As shown, this illustrates the temporal data matrix structure formed as the number of fine-tuning operations increases during the initial optical path assembly. The rows of the matrix represent the temporal evolution order, from the top light gray first row (initial fine-tuning state) to the bottom dark gray last row (final completed state). Each row records the absolute physical coordinates of all degrees of freedom of the adjustment frame after one fine-tuning operation. The columns correspond to the parameters of each degree of freedom. The time arrows on the left emphasize the temporal accumulation characteristics of the data. This matrix structure not only records the final perfect pose but also completely preserves the intermediate state data during the process of gradually approaching the zero position from the deviation state. This full-process data recording method enables the subsequently trained machine learning model to learn the nonlinear dynamic law between pose deviation and interference fringe changes, rather than simply memorizing a static endpoint, thereby improving the model's generalization ability and convergence speed when dealing with different initial reset errors.

[0063] S12: In each corresponding fine-tuning state, the laser interferometer is synchronously driven to collect optical path interference spot data at the corresponding moment, and generate a digital interference pattern sequence with stable power and determined shape.

[0064] Further, step S12 includes:

[0065] S121: At the same moment that each pose reading acquisition is triggered in step S112, the laser interferometer is synchronously driven to perform an interferometric pattern acquisition through the data acquisition control interface of the laser interferometer. The collimated laser beam emitted by the laser interferometer reaches the surface of the aspherical mirror under test through the auxiliary compensation optical path component and returns along the same path. The returned beam interferes with the reference beam inside the interferometer, and the area array detector of the interferometer converts the interference light intensity distribution into a two-dimensional pixel grayscale matrix. The two-dimensional pixel grayscale matrix is ​​a digital interferometric pattern. The number of rows and columns of the matrix is ​​determined by the pixel resolution of the area array detector. For example, the pixel resolution of the area array detector can be set to 1024 by 1024 pixels, then each digital interferometric pattern is a grayscale matrix of 1024 rows and 1024 columns, and the value of each element in the matrix is ​​an integer grayscale value from 0 to 255.

[0066] S122: To ensure the power stability and shape certainty of each digital interferometer pattern, before each interferometer pattern acquisition, the power monitoring module built into the laser interferometer verifies whether the laser output power is within a preset power stability range. The power stability range is determined as follows: using the nominal output power of the laser interferometer as the center, the percentage fluctuation above and below the nominal output power is taken as the upper and lower limits of the power stability range. For example, the fluctuation percentage can be set to ±2%. When the laser output power is within the power stability range, interferometer pattern acquisition is performed; when the laser output power exceeds the power stability range, acquisition is performed only after the laser output power returns to the power stability range, thus eliminating the influence of laser power fluctuations on the contrast of the interference fringes.

[0067] S123: Arrange all digital interferometry patterns corresponding one-to-one with the acquisition times of all pose readings in step S112 according to the acquisition time order to form a digital interferometry pattern sequence. The data structure of the digital interferometry pattern sequence is a three-dimensional array. The length of the first dimension is equal to the total number of fine-tuning operations, and the lengths of the second and third dimensions are equal to the number of row pixels and column pixels of the area array detector, respectively. Each digital interferometry pattern in the digital interferometry pattern sequence is strictly aligned with the pose data points in the same row of the absolute mechanical pose sequence in terms of acquisition time. Strict alignment means that the difference between the acquisition timestamps of the two is less than one-tenth of the single-frame exposure time of the laser interferometer. For example, when the single-frame exposure time is 10 milliseconds, the timestamp difference is less than 1 millisecond.

[0068] S13: Perform image filtering and noise reduction and fringe extraction feature construction on the digital interferometric pattern sequence to extract aspherical interference fringe information and stitch them together to form an interference fringe feature set;

[0069] Specifically, the method for extracting interference fringe feature sets based on digital interferometric pattern sequences is as follows:

[0070] S131: A Gaussian filtering algorithm is used to perform a convolution operation on each digital interferogram in the digital interferogram sequence to remove ambient light leakage noise and high-frequency laser speckle noise, resulting in a smoothed interferogram sequence with a smoothed background. See also... Figure 7 This is a schematic diagram of the sliding Gaussian filter convolution kernel provided in an embodiment of this application. Figure 7The diagram illustrates the principle of smoothing a local grayscale matrix using a two-dimensional Gaussian convolution kernel. The left side shows a 5×5 Gaussian convolution kernel with the largest weight at the center, decreasing towards the edges, conforming to a normal distribution. The right side shows the local pixel grid of the interference pattern. The convolution kernel slides pixel-by-pixel across the image, replacing the center pixel value through weighted summation. Dark areas in the image represent interference fringes or noise points. After convolution, isolated high-frequency speckle noise (such as the isolated black dot in the upper right corner) is eliminated by averaging with surrounding pixels, while continuous interference fringe edges are preserved and smoothed. This preprocessing effectively removes high-frequency interference caused by ambient light leakage and laser speckle, improving the signal-to-noise ratio and providing high-quality input data for the subsequent Canny edge detection algorithm to extract clear and continuous interference contour edges, avoiding false edge extraction errors caused by noise.

[0071] ;

[0072] Where x and y are the horizontal and vertical pixel offsets of the current element in the convolution kernel relative to the center of the kernel, respectively. The Gaussian standard deviation parameter is determined based on the relationship between the Gaussian kernel window side length parameter and the coverage coefficient. The Gaussian standard deviation parameter is obtained by subtracting 1 from the Gaussian kernel window side length parameter and dividing by 6. For example, when the Gaussian kernel window side length parameter is 5, the Gaussian standard deviation parameter is approximately 0.67 pixels. The two-dimensional Gaussian convolution kernel matrix is ​​slid pixel-by-pixel across the grayscale matrix of each digital interferogram. For each pixel position, a weighted sum of all grayscale values ​​within the convolution kernel coverage area and the corresponding weights of the convolution kernel is calculated. The weighted sum is used to replace the original grayscale value of the current pixel, resulting in a smooth interferogram. After performing the above Gaussian filtering operation on all digital interferograms in the digital interferogram sequence, a smooth interferogram sequence with the same number and arrangement order as the digital interferogram sequence is obtained. For example, suppose the original gray value of a certain pixel position is 180. Speckle noise around the pixel position causes the gray value of neighboring pixels to jump. After Gaussian filtering, the gray value of the pixel position is corrected to 172, which is consistent with the brightness trend of the surrounding interference fringes. The gray value jump corresponding to the high-frequency speckle noise is smoothed out.

[0073] S132: Extract the alternating bright and dark interference contour edges from each smooth interference pattern in the smooth interference pattern sequence, calculate the fringe line density of the separated interference contour edges and the curvature deformation parameters of the main fringes, and merge them to form an interference fringe feature set.

[0074] Further, step S132 includes:

[0075] S1321: Perform the Canny edge detection algorithm on each smooth interferometer pattern in the smooth interferometer pattern sequence to extract the edges of the interference contour. See also Figure 8 This is a schematic diagram of interference contour extraction using Canny edge detection provided in an embodiment of this application. Figure 8 The diagram illustrates a three-stage processing flow from smoothed interferometric patterns to the final interference contour edges. The left side shows the smoothed interferometric pattern after Gaussian filtering, with wide and blurred fringes. The middle stage uses non-maximum suppression to refine the wide fringes into single-pixel-wide skeleton pixel sequences, initially locating the edges. The right side stage applies a dual-threshold method for filtering, where solid lines represent "strong edges" with gradient magnitudes higher than the high threshold, and dashed lines represent "weak edges" with values ​​between the high and low thresholds but connected to the strong edges. Isolated low-gradient noise points are completely eliminated. This processing strategy ensures the integrity of the main interference fringes (preserving weak edges through connectivity) while effectively suppressing background noise (eliminating isolated points). The extracted precise interference contour edges are the basis for calculating fringe line density and curvature deformation parameters, directly determining the accuracy of subsequent pose feature extraction.

[0076] S1322: Calculate fringe line density based on interference contour edge. The calculation method for fringe line density is as follows: Multiple equally spaced scan lines are set along the horizontal and vertical directions in the current smoothed interference pattern. The interval value of the equal spacing is determined based on one-tenth of the pixel resolution of the area array detector. For example, when the pixel resolution of the area array detector is 1024 pixels, the interval value is 102 pixels, that is, 10 scan lines are set along the horizontal direction and 10 scan lines are set along the vertical direction. The number of intersections between each scan line and the interference contour edge is counted. The number of intersections refers to the total number of times the scan line crosses the interference contour edge as it moves from one end of the pattern to the other. The number of intersections of each scan line is divided by the pixel length of the scan line to obtain the line density value of a single scan line. The arithmetic mean of the line density values ​​of all scan lines is taken as the fringe line density of the current smoothed interference pattern. The physical meaning of fringe line density is the number of interference fringes per unit pixel length. A larger fringe line density indicates a larger deviation between the auxiliary compensation element and the optimal pose, while a smaller fringe line density approaching zero indicates that the optical path is closer to the zero-position detection state.

[0077] S1323: Calculate the curvature deformation parameters of the main fringes based on the interference profile edge. The main fringes are defined as the three longest connected edge curves in the interference profile edge. For each main fringe, perform the following curvature calculation: Along the pixel coordinate sequence of the main fringe, take a sampling point at preset sampling steps, where the sampling step is determined based on one-twentieth of the total pixel length of the main fringe. For each sampling point, take three adjacent sampling points: the sampling point itself and one sampling point before and after it. Fit an arc using these three adjacent sampling points. The arc fitting method is the three-point circle method, i.e., the radius of the arc is obtained by solving the circle equation satisfied by the pixel coordinates of the three sampling points. The reciprocal of the arc radius is taken as the local curvature value at the current sampling point. The standard deviation of the local curvature values ​​of all sampling points on the main fringe is taken as the curvature deformation parameter of the current main fringe. The arithmetic mean of the curvature deformation parameters of the three main fringes is taken as the curvature deformation parameter of the current smoothed interferogram. The physical meaning of the curvature deformation parameter is the discrete fluctuation amplitude of the degree of curvature of the interference fringes along their extension direction. The larger the curvature deformation parameter, the more the angular deviation of the auxiliary compensation element causes the interference fringes to bend unevenly. The smaller the curvature deformation parameter, the closer the interference fringes are to the ideal straight line or concentric circle shape.

[0078] S1324: Concatenate the fringe line density and curvature deformation parameters of the current smoothed interferogram into a 1-row, 2-column fringe feature vector. Process steps S1321 to S1323 are performed sequentially on all smoothed interferograms in the smoothed interferogram sequence. All resulting fringe feature vectors are arranged according to their corresponding acquisition time order to form an interference fringe feature set. The data structure of the interference fringe feature set is a two-dimensional matrix, where the number of rows equals the total number of fine-tuning operations, and the number of columns is 2. The first column represents the fringe line density, and the second column represents the curvature deformation parameters.

[0079] S14: Based on the simultaneous data labeling pairing criterion, the absolute physical coordinate parameters in the absolute mechanical pose sequence are mapped and combined with the corresponding fringe features in the interference fringe feature set to generate an aspherical optical path pose and interference feature mapping dataset with digital mapping scale labels.

[0080] Further, step S14 includes:

[0081] S141: Establish a simultaneous data marker pairing criterion. The simultaneous data marker pairing criterion means that the pose data point in the nth row of the absolute mechanical pose sequence and the fringe feature vector in the nth row of the interference fringe feature set have the same acquisition time stamp, where n is the sequence number of the fine-tuning operation, and the value of n ranges from 1 to the total number of fine-tuning operations. The same acquisition time stamp comes from the timestamp recorded during the synchronous trigger acquisition in steps S112 and S121. Pose data points with the same timestamp and fringe feature vectors are determined to be paired data.

[0082] S142: For each pair of paired data, the fringe feature vector in the nth row of the interference fringe feature set is used as the feature input field, and the pose data point in the nth row of the absolute mechanical pose sequence is used as the mechanical pose label field. The feature input field and the mechanical pose label field are horizontally concatenated along the column direction to form a mapping data record. The number of columns in the mapping data record is equal to the sum of the number of columns of the fringe feature vector and the number of columns of the pose data points. For example, when the detection application device is a coaxial hyperboloid mirror, the fringe feature vector is 1 row and 2 columns, the pose data points are 1 row and 4 columns, and the mapping data record is 1 row and 6 columns. The first 2 columns are the fringe line density and curvature deformation parameters, and the last 4 columns are the vertical lift displacement, pitch angle, yaw angle, and longitudinal translation distance along the optical axis.

[0083] S143: Stack all mapping data records row by row in chronological order of acquisition time to form an aspherical optical path pose and interferometric feature mapping dataset. The data structure of this dataset is a two-dimensional matrix, where the number of rows equals the total number of fine-tuning operations, and the number of columns equals the sum of the fringe feature dimension and the pose degree of freedom dimension. Each row of data in the dataset carries a digital mapping scale label, consisting of the fine-tuning operation number and the acquisition timestamp, used to uniquely identify the temporal position and source of each data record during subsequent model training. See also... Figure 6 This is a schematic diagram of synchronous acquisition and data alignment provided in an embodiment of this application. For example... Figure 6As shown, this diagram illustrates the strict pairing mechanism between the digital interferometric pattern sequence and the absolute mechanical pose sequence along the time axis. The upper time axis represents the digital interferometric patterns acquired by the laser interferometer, while the lower time axis represents the mechanical pose data read by the high-precision displacement sensor. The vertical dashed line and the "synchronous trigger" label indicate that at each fine-tuning stillness, the system simultaneously triggers camera exposure and sensor readings. This millisecond-level timestamp alignment (strict alignment) eliminates the risk of data misalignment caused by manual operation delays or device response lags. By ensuring that each interferometric pattern precisely corresponds to a unique mechanical pose state, causally defined training sample pairs are constructed. This is crucial for training a high-precision pose compensation prediction model, as any temporal asynchrony introduces noise, causing deviations in the model's learned mapping relationships and thus affecting the prediction accuracy during actual reset.

[0084] Specifically, step S1 establishes a quantitative mapping relationship between the mechanical pose space and the optical interference feature space by simultaneously acquiring the mechanical pose information of the auxiliary compensation optical path component and the interference pattern information of the laser interferometer during the initial optical path assembly and adjustment. In traditional aspherical inspection optical path assembly and adjustment practices, inspectors rely on visual observation of the interference fringe morphology to determine the direction and magnitude of the pose deviation of the auxiliary compensation element. This judgment process depends on personal experience and cannot be quantified and recorded. Step S1 transforms the implicit experience knowledge accumulated by inspectors during the assembly and adjustment process into a structured dataset, enabling the interference fringe changes caused by each fine adjustment of the mechanical pose to be accurately recorded and quantified.

[0085] In step S11, a high-precision displacement sensor is used to directly read the absolute physical coordinate parameters of the adjustment frame instead of the relative adjustment amount. The physical significance of this is that the absolute physical coordinate parameters describe the absolute spatial pose of the auxiliary compensation element in the coordinate system of the air-bearing vibration isolation platform, and are unaffected by zero-position drift after the adjustment frame is disassembled and reassembled. If only the relative adjustment amount is recorded, the starting point for the accumulation of the relative adjustment amount changes after the adjustment frame undergoes disassembly and reassembly, causing the recorded pose information to lose its reference benchmark. In contrast, the absolute physical coordinate parameters always use the reference hole system on the air-bearing vibration isolation platform as a unified reference coordinate system, possessing comparability and reproducibility across disassembly and reassembly cycles.

[0086] Step S13 selects fringe line density and curvature deformation parameters as characteristic parameters of interference fringes based on the physical mechanism of aspherical interferometry. In the aspherical null detection optical path, when the auxiliary compensation element is in an ideal pose, the surface shape error of the measured aspherical mirror surface is presented as sparse and approximately uniform interference fringes. When the auxiliary compensation element deviates from the ideal pose, additional wavefront aberrations are superimposed on the surface shape error, resulting in increased interference fringe density and fringe morphology curvature distortion. Fringe line density directly reflects the magnitude of the low-order defocus and tilt components of the additional aberrations introduced by the pose deviation of the auxiliary compensation element, while the curvature deformation parameter reflects the degree of distortion of the fringe morphology by the high-order astigmatism and coma components of the additional aberrations. The two parameters characterize the influence of the auxiliary compensation element pose deviation on the interference pattern from complementary dimensions, enabling the subsequent model to calculate the direction and magnitude of the pose deviation from these two features. If only the fringe line density is used without the curvature deformation parameter, the model cannot distinguish the difference between the increase in fringe density caused by pure defocusing deviation and that caused by angular deviation; if only the curvature deformation parameter is used without the fringe line density, the model cannot perceive the uniform fringe density change caused by the translational deviation of the auxiliary compensation element along the optical axis. The synergistic use of these two features enables the aspherical optical path pose and interferometric feature mapping dataset to fully cover the influence modes of all degrees of freedom deviations of the auxiliary compensation element on the interferogram.

[0087] In step S14, the mechanical pose data and interference fringe feature data are precisely paired and fused into a unified dataset based on the simultaneous data labeling pairing criterion. This provides training samples with a strict physical correspondence for training the pose compensation prediction model in the subsequent step S2. The construction method of the aspherical optical path pose and interference feature mapping dataset ensures that there is a causal relationship between the input features and labels of each training sample. That is, a specific mechanical pose state uniquely determines a specific interference fringe pattern. This causal relationship is the data foundation for the subsequent multilayer perceptron regression model to learn the nonlinear mapping law between pose and fringes.

[0088] S2: Train a multilayer perceptron regression initial network based on the aspherical optical path pose and interference feature mapping dataset, and construct and output a pose compensation prediction model that is applicable to both non-zero and zero position detection states;

[0089] Further, step S2 includes:

[0090] S21: Extract the data row content of the aspherical optical path pose and interference feature mapping dataset according to the preset distribution ratio threshold, and divide it into feature input data and mechanical pose true value label data;

[0091] Further, step S21 includes:

[0092] S211: From the aspherical optical path pose and interference feature mapping dataset, separate the first two columns of data as the original feature matrix. The first column of the original feature matrix is ​​the fringe line density, and the second column is the curvature deformation parameter. Separate the subsequent columns of data as the original label matrix. The number of columns of the original label matrix is ​​equal to the total number of degrees of freedom of the adjustment frame.

[0093] S212: Perform min-maximum linear normalization on the original feature matrix, mapping the values ​​of the stripe line density column and the curvature deformation parameter column to the interval [0,1] respectively. The min-maximum linear normalization calculation method is as follows: for any element value in any column of the original feature matrix, subtract the minimum value of the column from the element value, and then divide by the difference between the maximum and minimum values ​​of the column to obtain the normalized element value. Perform the same min-maximum linear normalization on the original label matrix, mapping the absolute physical coordinate parameters of each degree of freedom to the interval [0,1] respectively. The normalized original feature matrix is ​​used as the feature input data, and the normalized original label matrix is ​​used as the mechanical pose true value label data. During the normalization process, record and save the minimum and maximum values ​​of each column for subsequent inverse normalization restoration of the pose compensation prediction model output results in step S3.

[0094] S213: Divide the feature input data and the ground truth label data of the mechanical pose into a training subset and a validation subset according to a preset distribution ratio threshold. The preset distribution ratio threshold is determined as follows: while ensuring that the sample size of the validation subset is not less than one-tenth of the total sample size, allocate as many samples as possible to the training subset to enhance the model's fitting ability. For example, the preset distribution ratio threshold can be set to 80% of the total sample size for the training subset and 20% for the validation subset. The division method is to take the first 80% of the data rows in the aspherical optical path pose and interference feature mapping dataset according to their temporal arrangement, and take the last 20% of the data rows to form the training subset and the validation subset.

[0095] S22: Construct a neural network with nested hierarchical logic of neurons, specifically including introducing a neuron-processing input layer, a fully connected feedforward hidden layer, and a multidimensional coordinate calibration output layer to obtain the initial network of a multilayer perceptron regression; see also Figure 9 This is a schematic diagram of the multilayer perceptron regression network structure provided in an embodiment of this application. Figure 9The diagram illustrates the internal neural network architecture of the pose compensation prediction model. The input layer receives 2D normalized fringe features (fringe line density, curvature deformation parameters); subsequently, it enters a first fully connected feedforward hidden layer (64 neurons) and a second fully connected feedforward hidden layer (32 neurons). These two layers introduce nonlinear transformation capabilities through the ReLU activation function to fit the complex physical mapping relationship between pose and fringes. Finally, a multi-dimensional coordinate calibration output layer (without an activation function) outputs a pose prediction vector consistent with the number of degrees of freedom of the adjustment frame. This structure employs a two-layer hidden layer design, ensuring universal approximation capability while avoiding the overfitting risk of excessively deep networks, making it particularly suitable for training with small sample sizes of optical path adjustment data. The model continuously optimizes the weights through the backpropagation algorithm, ultimately achieving end-to-end accurate prediction from interferometric image features to mechanical adjustment amounts.

[0096] Further, step S22 includes:

[0097] S221: Construct a neuron-based input layer. This neuron-based input layer receives the normalized stripe feature vector of a single sample from the feature input data. The normalized stripe feature vector has a dimension of 2, corresponding to the normalized stripe line density and the normalized curvature deformation parameter. This neuron-based input layer does not perform any mathematical operations; it simply passes the received normalized stripe feature vector to the first fully connected feedforward hidden layer as is.

[0098] S222: Construct a fully connected feedforward hidden layer. The fully connected feedforward hidden layer consists of a first fully connected feedforward hidden layer and a second fully connected feedforward hidden layer connected in series. The processing procedure of the first fully connected feedforward hidden layer is as follows: The normalized fringe feature vector transmitted from the neuron's input layer is multiplied by a first weight matrix; the result of the matrix multiplication is added element-wise with a first bias vector; the result of the element-wise addition is input into a modified linear unit activation function for nonlinear transformation, and the first hidden feature vector is output. The first weight matrix has a dimension of 2 rows and 64 columns, the first bias vector has a dimension of 1 row and 64 columns, and the first hidden feature vector has a dimension of 1 row and 64 columns. The mathematical expression of the modified linear unit activation function is:

[0099] ;

[0100] Where z is any element value after element-wise addition, the modified linear unit activation function remains unchanged for element values ​​greater than 0 and sets element values ​​less than or equal to 0 to 0, thereby introducing non-linear expressive power. The processing procedure of the second fully connected feedforward hidden layer is as follows: perform matrix multiplication on the first hidden feature vector and the second weight matrix, perform element-wise addition on the result of the matrix multiplication with the second bias vector, input the result of the element-wise addition into the modified linear unit activation function for non-linear transformation, and output the second hidden feature vector. The second weight matrix has a dimension of 64 rows and 32 columns, the second bias vector has a dimension of 1 row and 32 columns, and the second hidden feature vector has a dimension of 1 row and 32 columns. The first weight matrix, the first bias vector, the second weight matrix, and the second bias vector are all learnable parameters learned from training data, and are given random initial values ​​using the Xavier uniform distribution initialization method before training begins. The basis for setting the number of neurons in the first fully connected feedforward hidden layer to 64 is as follows: the input feature dimension is 2, the output label dimension is the total number of degrees of freedom of the adjustment frame, the number of neurons in the hidden layer should be much larger than the sum of the input and output dimensions to provide sufficient nonlinear fitting capacity, while it should not be too large to avoid overfitting. For example, a value in the range of 8 to 12 times the sum of the input and output dimensions is taken.

[0101] S223: Construct a multidimensional coordinate calibration output layer. This layer performs a linear transformation on the second hidden feature vector, performs matrix multiplication on the second hidden feature vector and the output weight matrix, and performs element-wise addition on the result of the matrix multiplication with the output bias vector to output a pose prediction vector. The output weight matrix has a dimension of 32 rows multiplied by the total number of degrees of freedom of the frame, the output bias vector has a dimension of 1 row multiplied by the total number of degrees of freedom of the frame, and the pose prediction vector has a dimension of 1 row multiplied by the total number of degrees of freedom of the frame. The multidimensional coordinate calibration output layer does not use an activation function and directly outputs the linear transformation result because pose prediction is a regression task, and the output value needs to cover a continuous range of real numbers. Using an activation function would truncate or compress the output value range, leading to inaccurate prediction of extreme pose values. For example, when the detection application device type is a coaxial hyperboloid mirror, the total number of degrees of freedom of the adjustment frame is 4, then the dimension of the output weight matrix is ​​32 rows and 4 columns, the dimension of the output bias vector is 1 row and 4 columns, and the dimension of the pose prediction vector is 1 row and 4 columns.

[0102] S23: Input feature input data into the multilayer perceptron regression initial network to perform forward propagation processing calculation, generate pose prediction results obtained by network inference, calculate the Euclidean distance discrete deviation amplitude between the pose prediction results and the mechanical pose true value label data, and form the mean square error to evaluate the loss from the Euclidean distance discrete deviation amplitude.

[0103] Further, step S23 includes:

[0104] S231: The feature input data from the training subset is input sample by sample into the initial regression network of the multilayer perceptron. Each sample is processed sequentially through the neuron processing input layer, the first fully connected feedforward hidden layer, the second fully connected feedforward hidden layer, and the multidimensional coordinate calibration output layer, outputting a pose prediction vector. The pose prediction vectors of all samples in the training subset are aggregated to form the pose prediction result. The data structure of the pose prediction result is a two-dimensional matrix, where the number of rows in the matrix equals the number of samples in the training subset, and the number of columns in the matrix equals the total number of degrees of freedom of the adjustment frame.

[0105] S232: Calculate the Euclidean distance discrete deviation magnitude between the pose prediction result and the corresponding part of the training subset in the mechanical pose ground truth label data. The method for calculating the Euclidean distance discrete deviation magnitude is as follows: For the nth pose prediction vector in the pose prediction result and the nth ground truth label vector in the mechanical pose ground truth label data, calculate the sum of the squares of the differences between the corresponding elements of the two vectors, and then take the square root to obtain the Euclidean distance value of the nth sample. The Euclidean distance values ​​of all samples in the training subset are then aggregated to form a sequence of Euclidean distance discrete deviation magnitudes.

[0106] S233: The mean square error assessment loss is formed from the discrete deviation magnitude sequence of Euclidean distance. The formula for calculating the mean square error assessment loss is:

[0107] ;

[0108] in To assess the loss for mean square error, The number of samples in the training subset. Let be the pose prediction vector for the nth sample. Let n be the truth label vector of the nth sample. This is the sum of squared differences between the elements of the vector. For example, suppose the training subset contains 48 samples, the pose prediction vector of the first sample is [0.52, 0.31, 0.48, 0.65], and the ground truth label vector is [0.50, 0.30, 0.50, 0.63]. Then the squared Euclidean distance of the first sample is the square of (0.52-0.50) + the square of (0.31-0.30) + the square of (0.48-0.50) + the square of (0.65-0.63), which equals 0.0004 + 0.0001 + 0.0004 + 0.0004 = 0.0013. Summing the squared Euclidean distances of the 48 samples and dividing by 48 gives the mean squared error evaluation loss.

[0109] S24: The mean squared error is backpropagated using the gradient descent algorithm to evaluate the loss and update the neuron connection weight coefficients of the initial network of the multilayer perceptron regression. After the loss is evaluated to decrease and stabilize, the fixed neuron connection weight coefficients are extracted to generate the pose compensation prediction model.

[0110] Further, step S24 includes:

[0111] S241: The Adam adaptive moment estimation optimization algorithm is used as the gradient descent algorithm. The Adam adaptive moment estimation optimization algorithm calculates the partial derivatives of the mean squared error evaluation loss with respect to all learnable parameters in the initial regression network of the multilayer perceptron. All learnable parameters include a first weight matrix, a first bias vector, a second weight matrix, a second bias vector, an output weight matrix, and an output bias vector. The Adam adaptive moment estimation optimization algorithm adaptively adjusts the update step size of each learnable parameter based on the first-order and second-order momentum estimates of the partial derivatives, replacing the original learnable parameters with the updated ones to complete one round of parameter updates. The learning rate parameter of the Adam adaptive moment estimation optimization algorithm is determined based on the sample size of the training subset; for example, the learning rate parameter can be set to 0.001.

[0112] S242: Repeat steps S231 to S241 to form multiple rounds of iterative training. After each round of iterative training, the feature input data from the validation subset is input into the multilayer perceptron regression initial network of the current round, and the mean squared error evaluation loss on the validation subset is calculated as the validation loss value. The trend of the validation loss value with the iteration rounds is monitored. When the validation loss value does not decrease within a preset number of patience rounds, it is determined that the mean squared error evaluation loss has stabilized. The preset number of patience rounds is determined based on the ratio of the number of training subset samples to the network complexity. For example, the preset number of patience rounds can be set to 20 rounds. The validation loss value not decreasing means that the validation loss value of the current round is greater than or equal to the validation loss value of the previous round.

[0113] S243: When the mean square error assessment loss is determined to be stable, extract the values ​​of all learnable parameters in the current round of the multilayer perceptron regression initial network and perform a solidification operation. The solidification operation refers to storing the values ​​of all learnable parameters as persistent files in the control computer of the air-bearing vibration isolation platform, so that they can be directly loaded without retraining during subsequent calls. The solidified multilayer perceptron regression initial network is the pose compensation prediction model. The input of the pose compensation prediction model is the normalized fringe feature vector, and the output is the normalized pose prediction vector. The pose compensation prediction model is applicable to both non-zero and zero detection states because the physical mapping relationship between the pose deviation of the auxiliary compensation element and the change of interference fringes in both non-zero and zero detection states can be uniformly approximated by the nonlinear fitting capability of the multilayer perceptron within a local range. The difference between the two states lies only in the different numerical distribution range of the interference fringe features. Furthermore, the multilayer perceptron regression initial network has already covered the continuous pose change range from far from zero to near zero through initial data adjustment during training.

[0114] Specifically, step S2, by constructing and training a multilayer perceptron regression model, generalizes the discrete mapping relationship between the mechanical pose and interference fringe features established in step S1 into a continuous nonlinear prediction capability. In the traditional detection optical path reset process, inspectors need to rely on personal experience to observe the interference fringe morphology and determine the adjustment direction and amount. This judgment process is subjective and uncertain; different inspectors may have significantly different judgments on the same interference pattern. The pose compensation prediction model replaces this subjective judgment process with objective mathematical reasoning and calculation. Inputting the actual interference fringe features acquired after reset directly outputs the target pose to which the auxiliary compensation element should be adjusted, eliminating the uncertainty introduced by human judgment.

[0115] In step S22, a network structure with two fully connected feedforward hidden layers is used instead of a deeper network structure. This design is based on the fact that the mapping relationship between the pose deviation of the auxiliary compensation element and the interference fringe features in the aspherical detection optical path is physically constrained by the geometrical optics and wave optics theories of wavefront propagation. This is a low-dimensional nonlinear mapping with definite physical laws. A multilayer perceptron with two hidden layers already possesses universal approximation capabilities, able to approximate any continuous function with arbitrary precision. Using an excessively deep network structure can easily lead to overfitting when the number of training samples is limited, resulting in good model performance on training data but decreased generalization ability in actual reset scenarios. The two-hidden-layer design achieves a suitable balance between model fitting ability and generalization ability. In step S23, mean squared error is used as the loss function instead of a classification loss function because the pose compensation prediction task is a continuous value regression task. Mean squared error can measure the distance deviation between the predicted pose and the true pose in Euclidean space, directly corresponding to the physical accuracy target of the auxiliary compensation element's pose adjustment. The early stopping strategy in step S24 terminates training before the model begins to overfit by monitoring and verifying the changing trend of the loss value, ensuring that the pose compensation prediction model has reliable predictive ability for unseen reset state interference fringe features.

[0116] S3: After the detection optical path is physically disassembled and re-reset, the initial digital interference pattern in the reset state is collected and the reset interference fringe features are extracted. The reset interference fringe features are input into the pose compensation prediction model to calculate the target pose compensation vector.

[0117] Further, step S3 includes:

[0118] S31: After the auxiliary compensation optical path component is physically removed and hardware replaced and reassembled based on the reference hole system, a collimated beam is projected using a laser interferometer and the reflected light from the surface of the aspherical mirror is received to generate the initial digital interference pattern in the reset state.

[0119] Further, the specific process of step S31 is as follows: The testing personnel re-fix the adjustment frame base in the auxiliary compensation optical path assembly to the preset reference hole position on the air-bearing vibration isolation platform using hexagonal screws, and restore the coarse adjustment knobs of each adjustment frame to approximately the original reading position according to the final position parameters recorded in step S1. After completing the hardware assembly and reset, the power supply of the laser interferometer is turned on and the laser output power is waited for to enter the power stability range defined in step S122. The laser interferometer is driven to perform one interference pattern acquisition. The laser interferometer emits a collimated laser beam, which passes through the auxiliary compensation optical path assembly and reaches the surface of the aspherical mirror under test. The surface of the aspherical mirror under test reflects the collimated laser beam back to the auxiliary compensation optical path assembly and returns to the laser interferometer along the same path. The returned beam interferes with the reference beam inside the laser interferometer. The area array detector of the laser interferometer converts the interference light intensity distribution into a two-dimensional pixel grayscale matrix, which is the initial digital interference pattern. The initial digital interferometry pattern reflects the additional wavefront aberration introduced by the deviation between the actual pose and the optimal detection pose of the auxiliary compensation optical path component after physical disassembly and reassembly.

[0120] S32: Based on the initial digital interferometry pattern recognition, extract pixel-level image data attribute information specific to the detection environment and generate reset interference fringe features;

[0121] Specifically, the method for extracting the features of reset interference fringes is as follows:

[0122] S321: A two-dimensional image enhancement program is used to perform contrast enhancement and adaptive threshold segmentation pixel retention processing on the initial digital interferogram, separating the dark background area from the effective area of ​​the interference fringes, resulting in a pure interference fringe pattern free from environmental background interference. See also Figure 10 This is a schematic diagram illustrating the two-dimensional image enhancement and region segmentation effects provided in an embodiment of this application. Figure 10 As shown, the initial digital interference pattern (left) acquired after the physical disassembly and reassembly of the optical path is compared with the processed clean interference fringe pattern (right). The left image, due to a large reassembly deviation, exhibits obvious uneven background grayscale, stray light spots, and low contrast, resulting in blurred fringes. After histogram equalization to enhance contrast and combined with Otsu's adaptive threshold segmentation, the right image successfully eliminated background noise and non-interference regions, retaining only high-sharp binary interference fringes. This enhancement significantly improves the usability of low-quality images in the reassembly state, ensuring that even under non-ideal lighting or large pose deviations, reliable frequency domain distribution density and fringe surface deflection indices can be stably extracted, providing accurate input features for the pose compensation prediction model and guaranteeing a high first-time success rate in the reassembly process.

[0123] S322: Based on the pure interference fringe pattern, perform morphological line thinning calculations to extract the horizontal and vertical central skeleton contour lines. Calculate the frequency domain distribution density of the horizontal and vertical central skeleton contour lines in the image spatial domain, as well as the fringe surface deflection index, and then concatenate and combine them to form the reset interference fringe feature. The specific process of the morphological line thinning calculation is as follows: Perform the Zhang-Suen thinning algorithm on the effective region of the interference bright and dark fringes in the pure interference fringe pattern. The Zhang-Suen thinning algorithm peels away edge pixels layer by layer through alternating execution of two sub-iteration steps until all fringe regions are reduced to skeleton lines of single-pixel width. These skeleton lines are the horizontal and vertical central skeleton contour lines. The calculation method for the frequency domain distribution density is the same as the fringe line density calculation method in step S1322: Set multiple equally spaced scan lines along the horizontal and vertical directions in the pure interference fringe pattern, count the number of intersections between each scan line and the horizontal and vertical central skeleton contour lines, divide the number of intersections of each scan line by the scan line pixel length to obtain the density value of a single scan line, and take the arithmetic mean of all scan line density values ​​as the frequency domain distribution density. The calculation method for the fringe surface deflection index is the same as the calculation method for the curvature deformation parameter in step S1323: The three longest connected skeleton lines from the horizontal and vertical central skeleton contour lines are selected as the main skeleton lines. Each main skeleton line is sampled at equal intervals along the pixel coordinate sequence. The local curvature value of each sampling point is calculated using the three-point circle method. The standard deviation of the local curvature values ​​of all sampling points is taken as the surface deflection value of a single main skeleton line. The arithmetic mean of the surface deflection values ​​of the three main skeleton lines is taken as the fringe surface deflection index. The frequency domain distribution density and the fringe surface deflection index are concatenated into a 1x2 vector, which is the reset interference fringe feature.

[0124] S33: The reset interference fringe features are transmitted to the pose compensation prediction model through the call interface to perform forward offset calculation and analysis. The pose compensation prediction model outputs the multi-dimensional parameters of the optical coordinate center displacement deviation required for the detection system to reach the optimal detection zero position standard. The target pose compensation vector is obtained by integrating the multi-dimensional parameters.

[0125] Further, step S33 includes:

[0126] S331: The frequency domain distribution density and fringe surface deflection index in the reset interference fringe features are subjected to min-max linear normalization using the minimum and maximum values ​​of the feature columns saved in step S212 to obtain the normalized reset fringe feature vector. The minimum and maximum values ​​used in the normalization process are exactly the same as those used in the training phase in step S212, ensuring that the input data in the reset phase and the input data in the training phase are in the same numerical scale space.

[0127] S332: The normalized reset stripe feature vector is transmitted to the pose compensation prediction model stored in step S243 through the model call interface of the control computer. The pose compensation prediction model sequentially passes through the neuron processing input layer, the first fully connected feedforward hidden layer, the second fully connected feedforward hidden layer and the multi-dimensional coordinate calibration output layer for forward propagation calculation, and outputs the normalized pose prediction vector.

[0128] S333: Perform denormalization restoration on the normalized pose prediction vector using the minimum and maximum values ​​of the label column saved in step S212. Multiply each element value in the normalized pose prediction vector by the difference between the maximum and minimum values ​​of the corresponding column, and add the minimum value of the corresponding column to obtain the pose prediction absolute coordinate vector with actual physical dimensions. Calculate the element-wise difference between the pose prediction absolute coordinate vector and the final absolute physical coordinate parameters recorded in step S1 when the initial setup is completed. This element-wise difference is the multidimensional parameter of the optical coordinate center displacement deviation required for the detection system to reach the optimal detection zero-position standard. Encapsulate the multidimensional parameter of the optical coordinate center displacement deviation into a one-dimensional vector, which is the target pose compensation vector. The dimension of the target pose compensation vector is equal to the total number of degrees of freedom of the adjustment frame, and the physical meaning of each element value is the amount of displacement or angle that the auxiliary compensation element needs to move from the current reset position to the optimal position in the corresponding degree of freedom direction. For example, when the detection application device is a coaxial hyperboloid mirror, the target pose compensation vector is a 4-element one-dimensional vector. Assuming that the absolute coordinate vector of the pose prediction calculated in step S333 is [85.023 mm, 1.205 arcseconds, -0.873 arcseconds, 152.671 mm], and the final absolute physical coordinate parameters after the first assembly are [85.000 mm, 1.200 arcseconds, -0.870 arcseconds, 152.500 mm], then the target pose compensation vector is [0.023 mm, 0.005 arcseconds, -0.003 arcseconds, 0.171 mm]. This means that the optical height adjustment frame needs to be finely adjusted upward by 0.023 mm, the pitch direction of the optical three-dimensional precision adjustment frame needs to be increased by 0.005 arcseconds, the yaw direction needs to be decreased by 0.003 arcseconds, and the precision translation adjustment frame needs to be moved forward by 0.171 mm along the optical axis.

[0129] Specifically, step S3 applies the pose compensation prediction model to the actual optical path reset scenario. By acquiring the first interference pattern after reset and extracting interference fringe features, the precise compensation amount required for the auxiliary compensation element to adjust from the current reset pose to the optimal detection pose is directly calculated. In traditional reset methods, after completing hardware assembly and reset, the testing personnel need to repeatedly manually fine-tune the auxiliary compensation element and visually observe the changes in interference fringes. After each fine-tuning, it is necessary to wait for the interference pattern to stabilize and re-determine the adjustment direction and amount, forming multiple rounds of trial and error iteration. Step S3 compresses this multi-round trial and error iteration into a single calculation, requiring only one initial digital interference pattern to obtain the target pose compensation vector, fundamentally eliminating the manual trial and error link in the reset process. In step S321, contrast enhancement and adaptive threshold segmentation are performed on the initial digital interference pattern, which can effectively address the problem of reduced interference fringe contrast and enhanced background stray light caused by large pose deviations of the auxiliary compensation element after optical path reset, ensuring that reliable fringe features can still be extracted from low-quality reset interference patterns. In step S333, the normalized predicted value output by the model is converted into an absolute coordinate value with actual physical dimensions through inverse normalization restoration processing, so that the target pose compensation vector can be directly used to drive the precision displacement mechanism of the adjustment frame, realizing a seamless connection from the data space to the physical operation space.

[0130] S4: Perform three-dimensional coordinate system decoupling transformation on the target pose compensation vector to construct the adjustment frame displacement operation command. Use the adjustment frame displacement operation command to drive the auxiliary compensation optical path component to perform position fine adjustment action, and complete the interference detection of the target aspherical reflective mirror surface shape.

[0131] Further, step S4 includes:

[0132] S41: Classify and decouple the target pose compensation vector according to the pre-configured aspherical system category to form the adjustment frame displacement operation command specific to the measured component of the corresponding category;

[0133] Specifically, the method for constructing the adjustment frame displacement operation command by performing three-dimensional coordinate system decoupling transformation on the target pose compensation vector is as follows:

[0134] S411: When the detection application device type is a parabolic reflector, obtain the vertical lift displacement coordinates, spatial pitch offset angle coordinates, and spatial yaw offset angle coordinates in the target pose compensation vector. The vertical lift displacement coordinates are the element values ​​of the corresponding optical height adjustment frame lifting axis degrees of freedom in the target pose compensation vector; the spatial pitch offset angle coordinates are the element values ​​of the corresponding pitch angle degrees of freedom in the target pose compensation vector; and the spatial yaw offset angle coordinates are the element values ​​of the corresponding yaw angle degrees of freedom in the target pose compensation vector.

[0135] The vertical lifting displacement coordinates are divided by the single-step displacement resolution of the optical height adjustment frame's lifting stepper motor to convert them into the operating pulse value of the optical height adjustment frame's lifting stepper motor. The single-step displacement resolution value is provided in the manufacturer's technical manual for the optical height adjustment frame; for example, the single-step displacement resolution can be set to 0.001 mm per pulse. The spatial pitch offset angle coordinates and spatial yaw offset angle coordinates are divided by the single-step angle resolution of the stepper motor to convert them into the operating pulse value of the stepper motor required to control the pitch and yaw of the optical three-dimensional precision adjustment frame. The single-step angle resolution value of the stepper motor is provided in the manufacturer's technical manual for the optical three-dimensional precision adjustment frame; for example, the single-step angle resolution of the stepper motor can be set to 0.01 arcseconds per pulse. The stepper motor operating pulse values ​​from the above three channels are jointly encoded into a digital pulse sequence signal, which is the adjustment frame displacement operation command. For example, if the vertical lifting displacement coordinate is 0.023 mm and the single-step displacement resolution is 0.001 mm per pulse, then the number of pulses for the stepper motor in the lifting direction is 23 pulses; if the spatial pitch offset angle coordinate is 0.005 arcseconds and the single-step angle resolution of the stepper motor is 0.01 arcseconds per pulse, then the number of pulses for the stepper motor in the pitch direction is rounded to 1 pulse.

[0136] S412: When the detection application device type is a coaxial hyperboloid mirror, obtain the vertical lift displacement coordinates, pitch and yaw degree of freedom components, and longitudinal deviation coordinate components along the optical axis in the target pose compensation vector.

[0137] Divide the vertical lifting displacement coordinate by the single-step displacement resolution of the optical height adjustment frame lifting stepper motor to convert it into the operating pulse value of the optical height adjustment frame lifting stepper motor. The conversion method is the same as in S411. Multiply the longitudinal deviation coordinate component along the optical axis by the adjustment structure ratio to convert it into the precision translation adjustment frame push rod stroke feed displacement adjustment amount. The value of the adjustment structure ratio is determined by the transmission mechanism type of the precision translation adjustment frame. For example, when the precision translation adjustment frame uses a differential screw transmission mechanism, the adjustment structure ratio can be set to 0.01 mm per revolution. Convert the push rod stroke feed displacement adjustment amount into the number of rotations of the drive motor, then convert it into the drive pulse value, and encode it into a digital pulse sequence signal. Obtain the pitch and yaw degree of freedom components in the target pose compensation vector, and convert them into the stepper motor operating pulse value of the optical three-dimensional precision adjustment frame using the same method as in step S411. The values ​​of the lifting drive pulse of the optical height adjustment frame, the longitudinal drive pulse of the precision translation adjustment frame, and the pitch and yaw drive pulse of the optical three-dimensional precision adjustment frame are combined and encoded to form an adjustment frame displacement operation command with 4 channels.

[0138] S413: When the detection application device type is a freeform surface reflector, obtain the three-dimensional spatial coordinate translation distance component and the three-dimensional spatial attitude rotation angle in the target pose compensation vector. The three-dimensional spatial coordinate translation distance component is the element value of the three degrees of freedom corresponding to the X-axis translation distance, Y-axis translation distance, and Z-axis translation distance in the target pose compensation vector. The three-dimensional spatial attitude rotation angle is the element value of the two degrees of freedom corresponding to the pitch angle around the X-axis and the yaw angle around the Y-axis in the target pose compensation vector. Combine the three-dimensional spatial coordinate translation distance component and the three-dimensional spatial attitude rotation angle, divide the compensation amount of each degree of freedom by the single-step displacement resolution or single-step angle resolution of the corresponding mechanism axis of the five-dimensional optical adjustment frame to obtain the driving pulse value of each of the five mechanism axes. The single-step resolution of each mechanism axis of the five-dimensional optical adjustment frame is provided in the manufacturer's technical manual of the five-dimensional optical adjustment frame. The driving pulse values ​​of the five mechanism axes are jointly encoded into a digital pulse sequence signal containing five channels. The digital pulse sequence signal constitutes the adjustment frame displacement operation command.

[0139] S42: The drive signal corresponding to the displacement operation command of the adjustment frame is used to directly control all the linkage elements in the drive auxiliary compensation optical path assembly to produce precise position fine-tuning and angle compensation. This is to supplement and correct the mechanical tolerance adjustment deviation left after long-term use and physical reassembly, so that the quality of the detection optical path formed after the control fine-tuning correction meets the optical path acceptance criteria of precision aspherical surfaces, and completes the interference detection of the surface shape of the target aspherical reflector.

[0140] Further, step S42 includes:

[0141] S421: The adjustment frame displacement operation command generated in step S41 is sent to the stepper motor driver of each adjustment frame in the auxiliary compensation optical path assembly through the digital output interface of the control computer. After receiving the digital pulse sequence signal in the adjustment frame displacement operation command, the stepper motor driver drives the stepper motor on each adjustment frame to perform precise rotational motion according to the direction and number of pulses specified by the pulse sequence signal. The rotational motion of the stepper motor is converted into linear displacement or angular deflection of the auxiliary compensation element through the transmission mechanism inside the adjustment frame. All stepper motors of all adjustment frames synchronously receive the adjustment frame displacement operation command and synchronously execute precise rotational motion, so that the auxiliary compensation element synchronously completes pose fine adjustment in all degrees of freedom directions.

[0142] S422: After all the stepper motors of the adjustment frames have completed the movement amount specified by the adjustment frame displacement operation command, drive the laser interferometer to re-acquire a verification digital interferogram. Extract the verification fringe line density and verification curvature deformation parameters from the verification digital interferogram using the same method as in steps S321 and S322. Determine whether the verification fringe line density is less than a preset line density acceptance threshold and whether the verification curvature deformation parameter is less than a preset curvature acceptance threshold. The line density acceptance threshold is determined based on the maximum permissible residual fringe density corresponding to the surface shape detection accuracy requirements of the aspherical mirror being tested. The curvature acceptance threshold is determined based on the maximum permissible fringe curvature degree corresponding to the surface shape detection accuracy requirements of the aspherical mirror being tested. For example, the line density acceptance threshold can be set to 0.5 fringe per 100 pixels, and the curvature acceptance threshold can be set to 0.002 fringe per pixel. When the verification fringe line density is less than the line density acceptance threshold and the verification curvature deformation parameter is less than the curvature acceptance threshold, it is determined that the detection optical path quality meets the optical path acceptance criteria for precision aspherical surfaces, and the detection optical path reset is completed. When the verification fringe line density is greater than or equal to the line density acceptance threshold, or the verification curvature deformation parameter is greater than or equal to the curvature acceptance threshold, the verification digital interferogram is used as a new initial digital interferogram and returned to step S32 to re-extract and reset the interference fringe features. Steps S33 and S41 are then re-executed to generate new adjustment frame displacement operation commands, performing secondary compensation fine-tuning. The maximum allowable number of executions for the secondary compensation fine-tuning is determined based on the relationship between the single prediction accuracy of the pose compensation prediction model and the mechanical repeatability positioning accuracy of the adjustment frame. For example, the maximum allowable number of executions can be set to 3 times.

[0143] S423: After the detection optical path is reset, the aspherical mirror under test is placed at the designated position in the detection optical path. The laser interferometer is driven to collect the surface interference pattern of the aspherical mirror under test according to the standard interferometric detection procedure. The surface interference pattern is then fitted with Zernike polynomials by the surface analysis software built into the interferometer. The surface peak and valley values ​​and the root mean square value of the surface shape of the aspherical mirror under test are extracted as surface shape accuracy evaluation indicators, thus completing the interferometric detection of the surface shape of the target aspherical mirror.

[0144] Specifically, step S4 converts the target pose compensation vector calculated in step S3 into physical motion commands that the adjustment frame can directly execute. Closed-loop verification ensures that the reset detection optical path meets the accuracy requirements for precision aspherical shape detection, achieving a complete control closed loop from data calculation to physical operation. Step S41 classifies and decouples the adjustment frame's degree-of-freedom configurations for different types of aspherical detection optical paths. The physical significance is as follows: In the parabolic detection optical path, the plane mirror only needs to adjust the pitch and yaw angles to achieve accurate return of collimated light; therefore, the adjustment frame displacement command only includes drive pulses for the two angle channels. In the coaxial hyperboloid detection optical path, the spherical compensation mirror needs to adjust not only the angles but also precisely control the longitudinal position along the optical axis to meet the zero-position condition; therefore, the adjustment frame displacement command must additionally include drive pulses for the longitudinal translation channel. In the freeform surface detection optical path, the computer-generated hologram compensation mirror is sensitive to all six degrees of freedom poses; therefore, the adjustment frame displacement command must include drive pulses for all five channels. This classification and decoupling design avoids performing invalid adjustments on non-existent degrees of freedom, improving the accuracy and efficiency of the reset operation. The closed-loop verification mechanism in step S422 ensures that even if the pose compensation prediction model has certain prediction errors or the adjustment frame has small mechanical gaps, the detection optical path can still meet the acceptance criteria after no more than a preset number of iterative compensations. Compared with traditional manual multi-round trial and error, the iterative compensation in step S422 is a model-driven directional correction. Each iteration converges towards the optimal pose direction, rather than the blind trial and error relying on experience in manual assembly. Therefore, the convergence speed is significantly faster than traditional methods, reducing the total time for optical path reset from tens of minutes or even hours in traditional methods to less than a few minutes.

[0145] Example 2:

[0146] This embodiment, based on Embodiment 1, provides a rapid resettable detection system for different aspherical optical paths on an air-float platform, such as... Figure 11 As shown, it includes:

[0147] The mapping dataset construction module is used to acquire the absolute mechanical pose sequence and the digital interferogram sequence during the initial optical path assembly and adjustment stage, extract the interference fringe feature set based on the digital interferogram sequence, and fuse the absolute mechanical pose sequence and the interference fringe feature set into an aspherical optical path pose and interference feature mapping dataset.

[0148] The model training module is used to train a multilayer perceptron regression initial network based on the aspherical optical path pose and interference feature mapping dataset, and to construct and output a pose compensation prediction model that is applicable to both non-zero and zero-position detection states.

[0149] The compensation vector calculation module is used to collect the initial digital interference pattern in the reset state and extract the reset interference fringe features after the detection optical path is physically disassembled and re-reset, and input the reset interference fringe features into the pose compensation prediction model to calculate the target pose compensation vector.

[0150] The reset detection execution module is used to perform a three-dimensional coordinate system decoupling transformation on the target pose compensation vector to construct an adjustment frame displacement operation command. The adjustment frame displacement operation command is used to drive the auxiliary compensation optical path component to perform a position fine-tuning action, thereby completing the interference detection of the target aspherical reflector surface shape.

Claims

1. A rapid resettable detection method for different aspherical optical paths on an air-floating platform, characterized in that, The method includes: In the initial optical path assembly and adjustment phase, an absolute mechanical pose sequence and a digital interferometric pattern sequence are acquired. Specifically, during the initial optical path setup, a high-precision displacement sensor reads the absolute physical coordinate parameters of each adjustment frame in the auxiliary compensation optical path assembly under multiple fine-tuning states, obtaining the absolute mechanical pose sequence. These absolute physical coordinate parameters always use the reference hole system on the air-bearing vibration isolation platform as a unified reference coordinate system. In each corresponding fine-tuning state, a laser interferometer is synchronously driven to acquire optical path interference spot data at the corresponding moment, generating the digital interferometric pattern sequence. Each digital interferometric pattern in the digital interferometric pattern sequence is strictly aligned with the pose data points in the same row of the absolute mechanical pose sequence at the acquisition time. Strict alignment means that the difference in their acquisition timestamps is less than one-tenth of the single-frame exposure time of the laser interferometer. An interference fringe feature set is extracted based on the digital interferometric pattern sequence, and the absolute mechanical pose sequence and the interference fringe feature set are fused into an aspherical optical path pose and interference feature mapping dataset. Based on the aforementioned aspherical optical path pose and interference feature mapping dataset, a multilayer perceptron regression initial network is trained to construct and output a pose compensation prediction model that is applicable to both non-zero and zero-position detection states. After the auxiliary compensation optical path component is physically dismantled and hardware replaced and reassembled based on the reference hole system, an initial digital interference pattern in the reset state is generated using a laser interferometer; the reset interference fringe features are extracted based on the initial digital interference pattern; the reset interference fringe features are input into the pose compensation prediction model to perform positive offset calculation and analysis to obtain the target pose compensation vector. The target pose compensation vector is decoupled and transformed in three-dimensional coordinate system to construct the adjustment frame displacement operation command. The adjustment frame displacement operation command is used to drive the auxiliary compensation optical path component to perform position fine adjustment action, thereby completing the interference detection of the target aspherical reflective mirror surface shape.

2. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 1, characterized in that, The dataset of aspherical optical path pose and interference feature mapping includes: The digital interference pattern sequence is subjected to image filtering and noise reduction and fringe extraction feature construction processing to extract aspherical interference fringe information and stitch them together to form the interference fringe feature set; Based on the simultaneous data labeling pairing criterion, the absolute physical coordinate parameters in the absolute mechanical pose sequence are mapped and combined with the corresponding fringe features in the interference fringe feature set to generate the aspherical optical path pose and interference feature mapping dataset with digital mapping scale labels.

3. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 2, characterized in that, Obtaining the absolute mechanical pose sequence includes: After the initial installation of the interferometer and auxiliary compensation optical path assembly is completed on the air-bearing vibration isolation platform, high-precision displacement sensors are installed at the moving joints of each adjustment frame in the auxiliary compensation optical path assembly. During the initial precision assembly and adjustment of the optical path, whenever a fine-tuning operation is completed on any adjustment frame in the auxiliary compensation optical path assembly and it remains stationary, all high-precision displacement sensors are simultaneously triggered to acquire a pose reading. All pose data points collected sequentially during the initial optical path precision assembly process are arranged in chronological order of collection time to form the absolute mechanical pose sequence.

4. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 2, characterized in that, Generating the digital interferometric pattern sequence includes: At the same moment that each pose reading acquisition is triggered, the laser interferometer is synchronously driven to perform an interferometric pattern acquisition through the data acquisition control interface of the laser interferometer; Before each interferogram acquisition, verify whether the laser output power is within the preset power stability range. If so, proceed with the interferogram acquisition. All digital interferograms corresponding to the acquisition times of all pose readings are arranged in chronological order to form the digital interferogram sequence.

5. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 2, characterized in that, The formation of the interference fringe feature set includes: The Gaussian filtering algorithm is used to perform a convolution operation on each digital interferogram in the digital interferogram sequence to obtain a smoothed interferogram sequence with background smoothing. Extract the alternating light and dark interference contour edges from each smooth interference pattern in the smooth interference pattern sequence, calculate the fringe line density of the separated interference contour edges and the curvature deformation parameters of the main fringes, and combine them to form the interference fringe feature set.

6. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 1, characterized in that, The steps for constructing the pose compensation prediction model include: According to the preset distribution ratio threshold, the data row content of the aspherical optical path pose and interference feature mapping dataset is extracted and divided into feature input data and mechanical pose true value label data. Construct a multilayer perceptron to revert to the initial network; The feature input data is input into the multilayer perceptron regression initial network to perform forward propagation inference calculation, generate pose prediction results, and calculate the mean square error of the pose prediction results and the mechanical pose true value label data to evaluate the loss. The mean squared error is backpropagated using the gradient descent algorithm to evaluate the loss and update the neuron connection weight coefficients of the initial regression network of the multilayer perceptron. After the evaluation loss stabilizes, the neuron connection weight coefficients are fixed to generate the pose compensation prediction model.

7. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 6, characterized in that, The multilayer perceptron regression initialization network includes a neuron processing input layer, a fully connected feedforward hidden layer, and a multidimensional coordinate calibration output layer connected in sequence. The neuron processing input layer is used to receive the normalized stripe feature vector; The fully connected feedforward hidden layer consists of two cascaded fully connected feedforward hidden layers, used to perform nonlinear transformations on the input features; The multidimensional coordinate calibration output layer is used to output the pose prediction vector.

8. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 1, characterized in that, The extracted and reset interference fringe features include: The initial digital interference pattern is subjected to contrast enhancement and adaptive threshold segmentation pixel retention processing to obtain a pure interference fringe pattern that eliminates environmental background interference; Based on the pure interference fringe pattern, morphological line refinement calculations are performed to extract the horizontal and vertical central skeleton contour lines, statistically analyze the frequency domain distribution density of the horizontal and vertical central skeleton contour lines and the fringe surface deflection index, and then splice them together to form the reset interference fringe feature.

9. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 1, characterized in that, The interferometric detection of the target aspherical reflective mirror surface shape includes: The target pose compensation vector is classified and decoupled according to the pre-configured aspherical system category to form the adjustment frame displacement operation command of the corresponding category of the measured element; The displacement operation command of the adjustment frame drives the linkage element in the auxiliary compensation optical path assembly to generate precise position fine adjustment and angle compensation. After the quality of the optical path to be tested meets the acceptance criteria for precision aspherical optical path, the interference detection of the surface shape of the target aspherical reflector is completed.

10. The rapid resettable detection method for different aspherical optical paths on an air-floating platform according to claim 9, characterized in that, The classification processing decoupling transformation includes: When the detection application device is a parabolic reflector, the decoupling yields the driving pulses for three degrees of freedom: vertical lift, pitch, and yaw. When the detection application device is a coaxial hyperboloid mirror, decoupling is used to obtain four degrees of freedom of driving pulses: vertical lifting, pitching, yaw, and longitudinal translation along the optical axis. When the detection application device is a freeform surface reflector, decoupling yields driving pulses with five degrees of freedom.

11. A rapid resetability detection system for different aspherical optical paths on an air-floating platform, used to implement the rapid resetability detection method for different aspherical optical paths on an air-floating platform as described in any one of claims 1-10, characterized in that, The system includes: A mapping dataset construction module is used to acquire the absolute mechanical pose sequence and digital interferometric pattern sequence during the initial optical path assembly and adjustment phase. Specifically, during the initial detection of the optical path setup, a high-precision displacement sensor reads the absolute physical coordinate parameters of each adjustment frame in the auxiliary compensation optical path component under multiple fine-tuning states, obtaining the absolute mechanical pose sequence. The absolute physical coordinate parameters always use the reference aperture system on the air-bearing vibration isolation platform as a unified reference coordinate system. In each corresponding fine-tuning state, a laser interferometer is synchronously driven to acquire optical path interference spot data at the corresponding time, generating the digital interferometric pattern sequence. Each digital interferometric pattern in the digital interferometric pattern sequence is strictly aligned with the pose data points in the same row of the absolute mechanical pose sequence at the acquisition time. Strict alignment means that the difference in their acquisition timestamps is less than one-tenth of the single-frame exposure time of the laser interferometer. An interference fringe feature set is extracted based on the digital interferometric pattern sequence, and the absolute mechanical pose sequence and the interference fringe feature set are fused into an aspherical optical path pose and interference feature mapping dataset. The model training module is used to train a multilayer perceptron regression initial network based on the aspherical optical path pose and interference feature mapping dataset, and to construct and output a pose compensation prediction model that is applicable to both non-zero and zero-position detection states. The compensation vector calculation module is used to generate an initial digital interference pattern in the reset state using a laser interferometer after the auxiliary compensation optical path component has been physically dismantled and hardware replaced and reassembled based on the reference aperture system. Based on the initial digital interference pattern, the module extracts the reset interference fringe features and inputs the reset interference fringe features into the pose compensation prediction model to perform forward offset calculation and analysis to obtain the target pose compensation vector. The reset detection execution module is used to perform a three-dimensional coordinate system decoupling transformation on the target pose compensation vector to construct an adjustment frame displacement operation command. The adjustment frame displacement operation command is used to drive the auxiliary compensation optical path component to perform a position fine-tuning action, thereby completing the interference detection of the target aspherical reflector surface shape.

Images