Micro bearing production and manufacturing detection data processing method and system based on cloud computing

Abstract

The application discloses a cloud computing-based micro bearing production, manufacturing and detection data processing method and system, relates to the technical field of data processing, and comprises the following steps: collecting a defocus intensity image sequence of a micro bearing to be measured; calculating the rotational geometric center coordinates of an in-focus image, encapsulating the rotational geometric center coordinates, the defocus intensity image sequence and preset optical parameter data into a data packet to be measured; inputting a pre-trained physical unfolding self-supervised network, solving an initial phase tensor and constructing an initial concentric ring coordinate system; generating a flexible implicit sampling grid by using a physical texture vector flow field evolution, extracting a target latent vector and folding the target latent vector into a Cartesian reconstruction phase tensor; comparing Mahalanobis distances of the target latent vector, calculating a tensor difference value to generate an abnormal thermal matrix, and extracting absolute defect coordinates; deducing the Cartesian reconstruction phase tensor into a predicted defocus image, calculating a reconstruction confidence score, and substituting the reconstruction confidence score into a preset discrete proportional integral control equation to obtain an updated micro-motion defocus step length and an industrial camera exposure time.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for processing data related to the production and testing of micro-bearings based on cloud computing. Background Technology

[0002] As a core component in precision manufacturing, the tolerance control of the surface morphology and machining texture of miniature bearings has entered the sub-micron and even nanometer scale. In traditional industrial production lines, limited by the physical diffraction limit of optical microscopes, bearing surface images acquired using conventional industrial vision systems often exhibit extremely low contrast and severe artifact interference. Furthermore, conventional image processing algorithms are prone to confusing subtle three-dimensional scratches or indentations with normal mesh-like cutting marks. This results in the detection accuracy being extremely dependent on expensive high-order three-dimensional optical interferometry equipment, severely restricting online quality control and large-scale inspection speed throughout the entire manufacturing process.

[0003] Currently, Chinese invention patent application CN120561467B discloses an oscilloscope signal data processing method and system for bearing inspection. This method involves acquiring bearing vibration data using an oscilloscope, extracting data from a first moment for frequency domain analysis, obtaining its power spectrum, and dividing it into frequencies to obtain a first frequency sequence. Similarly, a second frequency sequence is obtained from the data from a second moment. Aligning the two frequency sequences yields frequency matching pairs. Utilizing the frequency variations of neighboring frequencies in a single matching pair, the noise figure of each frequency in the matching pair is calculated, resulting in the noise figure of any frequency in the first frequency sequence. Based on these noise figures, a first mean sequence is constructed. The distribution consistency of each frequency point in the first mean sequence is obtained, and the segmentation effectiveness of each frequency point is calculated. Combined with a noise segmentation threshold, oscilloscope signal processing is completed. However, this related technology struggles to achieve zero-loss extraction through the diffraction limit of physical optics, and it is difficult to remove normal processing tolerance interference when facing extremely small surface defects, lacking perceptuality and robustness. Summary of the Invention

[0004] The technical problem solved by this invention is that related technologies are difficult to extract with zero loss through the diffraction limit of physical optics, and are difficult to remove normal processing tolerance interference when faced with extremely small surface defects, lacking perceptibility and robustness.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A cloud-based method for processing manufacturing and inspection data of miniature bearings includes the following steps: Step S1: Based on the execution instruction packet sent by the cloud server, control the micro-motion displacement platform at the edge end to acquire the defocus intensity image sequence of the micro-bearing under test; Step S2: Calculate the rotational geometric center coordinates of the in-focus image, encapsulate them together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; Step S3: Input the data packet to be tested into the pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in the two-dimensional Cartesian coordinate system based on the preset optical parameter data, and construct the initial concentric ring coordinate system. Step S4: The initial concentric ring coordinate system is subjected to spatial ordinary differential iteration through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and the target latent vector is extracted and folded into a Cartesian reconstructed phase tensor. If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. Step S5: The Cartesian reconstructed phase tensor is derived into the predicted defocus image. The reconstruction confidence score is calculated using the mean square error. The score is then substituted into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocus step size and industrial camera exposure time.

[0006] Preferably, step S1 specifically includes: Step S11: The cloud server obtains the micro-motion defocus step size and industrial camera exposure time updated in the previous batch, and uses them as the micro-motion defocus step size and industrial camera exposure time for the current batch, respectively. Step S12: Package the micro-motion defocusing step size and the industrial camera exposure time of the current batch into an execution instruction package and send it to the edge controller. Step S13: The edge controller parses and executes the instruction packet to control the micro-motion platform arranged along the detection axis to run to the displacement point, which includes a negative displacement point, a zero point, and a positive displacement point. Step S14: When the micro-motion displacement platform moves to the displacement point, the industrial camera is triggered to sequentially acquire the negative defocus image, positive focus image and positive defocus image of the micro bearing under test, and generate a defocus intensity image sequence in a two-dimensional Cartesian coordinate system.

[0007] Preferably, step S2 specifically includes: Step S21: Perform multi-scale logarithmic Gabor filtering on the positive focus image to extract the real part convolution response matrix and the imaginary part convolution response matrix of the positive focus image at each filter scale. Based on the real part convolution response matrix and the imaginary part convolution response matrix, the local amplitude and local phase of each pixel in the positive focus image are calculated at each filter scale; Calculate the sum of local amplitudes of each pixel in the in-focus image at all filter scales, and calculate the phase alignment energy of the local phase at all filter scales; Divide the phase alignment energy by the sum of local amplitudes to calculate the phase consistency value of each pixel in the positive focus image. Obtain the phase consistency values ​​of all pixels to form a two-dimensional phase consistency feature map. Non-maximum suppression calculation is performed on the two-dimensional phase consistency feature map to extract the connected extremum skeleton line of the outer contour of the micro bearing under test, and the rotational geometric center coordinates of the connected extremum skeleton line are calculated by the least squares ellipse fitting operator. Step S22: Read preset optical parameter data, which includes the center wavelength of the light source, the physical pixel size of the industrial camera, and the optical magnification. Step S23: Encapsulate the defocus intensity image sequence, rotation geometric center coordinates, micro-motion defocus step size of the current batch, and preset optical parameter data to obtain the test data package, and send it to the cloud server.

[0008] Preferably, step S3 specifically includes: Step S31: The cloud server decompresses the data packet to be tested, inputs the defocus intensity image sequence and the micro-motion defocus step size of the current batch into the pre-trained physical unfolding self-supervised network, and calculates the physical space sampling equivalent through the physical computing layer of the pre-trained physical unfolding self-supervised network. The pre-trained physics unfolding self-supervised network includes a physics computation layer, a differential homeomorphism evolution layer, an implicit sampling encoder layer, a decoder layer, and a differentiable forward inference layer. Step S32: Using the center difference algorithm, calculate the light intensity derivative matrix of the defocus intensity image sequence in the physical calculation layer along the direction perpendicular to the two-dimensional Cartesian coordinate system; Step S33: Based on the center wavelength of the light source and the light intensity derivative matrix, perform phase calculation to obtain the initial phase tensor; Step S34: The differential homeomorphic evolution layer obtains the initial phase tensor and the coordinates of the rotational geometric center. Using the coordinates of the rotational geometric center as the pole, an initial concentric ring sampling coordinate system is constructed in a two-dimensional Cartesian coordinate system.

[0009] Preferably, step S33 specifically includes: The physical wave number of the light wave is calculated based on the center wavelength of the light source. The physical wave number is then multiplied by the light intensity derivative matrix. The result is then negative to obtain the source term matrix. By constructing the Poisson equation using the source term matrix, the initial phase tensor can be solved, and its calculation expression is as follows: ; in, For the initial phase tensor, For a two-dimensional Laplace operator, The source term matrix, The coordinates are in a two-dimensional Cartesian coordinate system. The optical intensity matrix of the positive focus image. For regularization parameters, For divergence operators, This is the spatial gradient operator.

[0010] Preferably, step S4 specifically includes: Step S41: Perform first-order partial derivative operation on the initial phase tensor to obtain the spatial gradient matrix, extract the tangential and normal components of the spatial gradient matrix, and construct the physical texture vector flow field on the two-dimensional Cartesian coordinate system. A flexible implicit sampling mesh is obtained by iterating the initial concentric annular coordinate system using the physical texture vector flow field through spatial ordinary differential equations. Step S42: The implicit sampling encoder layer obtains a flexible implicit sampling grid, uses the coordinates of the target grid points in the flexible implicit sampling grid as an index to obtain the target grid point coordinate index relationship, and extracts the pixel values ​​at the positions that coincide with the coordinates of the target grid points in the initial phase tensor and concatenates them to generate a one-dimensional sequence feature. The target latent vector is extracted by linear projection and attention weight calculation of one-dimensional sequence features; In step S43, the decoder layer receives the target latent vector and the flexible implicit sampling grid. Based on the coordinate index relationship of the target grid points in the flexible implicit sampling grid, the feature element values ​​in the target latent vector are filled into the corresponding coordinate positions in the two-dimensional Cartesian coordinate system one by one to obtain the Cartesian reconstructed phase tensor.

[0011] Preferably, the spatial ordinary differential iterative evolution of the initial concentric annular coordinate system through the physical texture vector flow field specifically includes: Extract the coordinates of each grid point in the initial concentric annular coordinate system and construct the initial coordinate matrix; The physical texture vector flow field is used as a two-dimensional evolution velocity field. A preset Gaussian smoothing kernel is used to perform spatial convolution smoothing on the two-dimensional evolution velocity field to generate a smooth evolution velocity field. Extract the velocity vector of each grid point's coordinate in the smoothed evolving velocity field, and combine it with the grid point coordinates of the current iteration to calculate the updated grid point coordinates. The calculation expression is as follows: ; in, For the first The updated grid point coordinates obtained from the next iteration. For the first The coordinates of the current grid point at the next iteration, when hour, The initial grid point coordinates, To preset the evolution time step, To smooth the evolution of the velocity field, The coordinates of the current grid point The velocity vector at the corresponding position in the smooth evolution velocity field; Perform numerical truncation of the edges of the updated grid point coordinates in a two-dimensional Cartesian coordinate system; Determine if the current iteration count has reached the maximum iteration count: If the target is not reached, the updated grid point coordinates will be used as the input for the next iteration, and spatial convolution smoothing will be returned for the next iteration. If the target is reached, stop the iteration, output the final grid point coordinate matrix, and use the final grid point coordinate matrix as the flexible implicit sampling grid.

[0012] Preferably, step S4 further includes: Step S44: Obtain the set of extreme values ​​for the normal state. This set includes the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the Mahalanobis distance sequence. The Mahalanobis distance scalar of the micro-bearing under test is calculated using the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the target latent vector. The calculation expression is as follows: ; in, The Mahalanobis distance scalar of the miniature bearing under test. Let be the target latent vector. The mean of the high-dimensional latent vector. For normal covariance matrix, superscript For transpose operation, superscript To perform the inverse operation; Step S45: Extract the Mahalanobis distance sequence from the set of extreme values ​​of the normal state, and truncate the Mahalanobis distance sequence that exceeds the set baseline to obtain the tail distribution sequence. Using the maximum likelihood estimation algorithm, the distribution parameters of the generalized Pareto distribution are solved based on the tail distribution sequence to obtain the target extreme value distribution function. Combined with the preset confidence level constant, the distribution quantile value corresponding to the target extreme value distribution function is solved, and the distribution quantile value is used as the extreme value threshold. If the Mahalanobis distance scalar is greater than the extreme value threshold, subtract the Cartesian reconstructed phase tensor from the initial phase tensor in the two-dimensional Cartesian coordinate system and take the absolute value to generate the Cartesian anomalous thermodynamic matrix. Step S46: Extract the row and column index subscripts of the abnormal points in the Cartesian abnormal thermal matrix that are greater than the set value in the two-dimensional Cartesian coordinate system, and perform coordinate transformation by combining the physical space sampling equivalent and the coordinates of the rotational geometric center to obtain the absolute defect coordinates of the surface of the micro bearing to be tested.

[0013] Preferably, step S5 specifically includes: Step S51: The differentiable forward derivation layer receives the Cartesian reconstructed phase tensor, combines the physical space sampling equivalent and the center wavelength of the light source, and uses the Fresnel diffraction integral formula to calculate the forward and backward light wave transfer functions in the two-dimensional Cartesian coordinate system to generate the predicted negative defocus image and the predicted positive defocus image. Step S52: Calculate the sum of the mean square errors between the negative defocus image and the positive defocus image and the predicted negative defocus image and the predicted positive defocus image, respectively. Step S53: Calculate the original reconstruction confidence score using mean square error, normalize the original reconstruction confidence score to obtain the reconstruction confidence score, and calculate the control deviation between the preset confidence threshold and the reconstruction confidence score. Step S54: Substitute the control deviation into the preset discrete proportional-integral control equation to calculate the micro-motion defocusing step length compensation and exposure time compensation. Add the micro-motion defocus step size compensation and the exposure time compensation to the current batch's micro-motion defocus step size and the current batch's industrial camera exposure time, respectively, to obtain the updated micro-motion defocus step size and industrial camera exposure time.

[0014] The cloud computing-based micro-bearing manufacturing and inspection data processing system includes an image acquisition module, a data encapsulation module, a phase calculation module, a defect location module, and a closed-loop control module. The image acquisition module is used to control the micro-displacement platform at the edge end based on the execution instruction package sent by the cloud server, and to acquire the defocus intensity image sequence of the micro bearing under test; The data encapsulation module is used to calculate the rotational geometric center coordinates of the in-focus image, encapsulate it together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; The phase calculation module is used to input the data packet to be tested into a pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in a two-dimensional Cartesian coordinate system based on preset optical parameter data, and construct an initial concentric ring coordinate system. The defect localization module is used to perform spatial ordinary differential iteration on the initial concentric ring coordinate system through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and extract the target latent vector and fold it into a Cartesian reconstructed phase tensor. If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. The closed-loop control module is used to deduce the Cartesian reconstructed phase tensor into a predicted defocused image, calculate the reconstruction confidence score using the mean square error, and substitute it into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocusing step size and industrial camera exposure time.

[0015] The beneficial effects of this invention are as follows: This invention utilizes a flexible implicit sampling mesh generation architecture based on spatial ordinary differential iterative evolution of physical texture vector flow field. This architecture enables the implicit sampling encoder and decoder layers to complete feature extraction and folding without altering the two-dimensional Cartesian coordinate system, achieving zero interpolation information loss extraction for submicron defects. Phase calculation is performed using the Poisson equation and light intensity derivative matrix built into the physical calculation layer, and closed-loop verification is performed by substituting the Fresnel diffraction integral formula into the differentiable forward derivation layer, achieving three-dimensional phase reconstruction. The extreme value threshold is accurately generated by fitting the generalized Pareto distribution model using extreme value theory, and the compensation amount for micro-motion defocusing step size and exposure time is calculated by combining the discrete proportional integral control equation. Without introducing a large computational overhead, this invention endows the production line inspection system with adaptive capability without manual intervention and possesses industrial robustness. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a cloud computing-based method for processing testing data in the production of micro-bearings, as provided in one embodiment of the present invention. Figure 2 This is a schematic diagram of the basic process of a cloud computing-based micro-bearing manufacturing and testing data processing system provided in one embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1 This paper provides a cloud-based method for processing manufacturing and inspection data of micro-bearings, including the following steps: Step S1: Based on the execution instruction packet sent by the cloud server, control the micro-motion displacement platform at the edge end to acquire the defocus intensity image sequence of the micro-bearing under test; Step S2: Calculate the rotational geometric center coordinates of the in-focus image, encapsulate them together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; Step S3: Input the data packet to be tested into the pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in the two-dimensional Cartesian coordinate system based on the preset optical parameter data, and construct the initial concentric ring coordinate system. Step S4: The initial concentric ring coordinate system is subjected to spatial ordinary differential iteration through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and the target latent vector is extracted and folded into a Cartesian reconstructed phase tensor. If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. Step S5: The Cartesian reconstructed phase tensor is derived into the predicted defocus image. The reconstruction confidence score is calculated using the mean square error. The score is then substituted into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocus step size and industrial camera exposure time.

[0019] This invention utilizes a flexible implicit sampling mesh generation architecture based on spatial ordinary differential iterative evolution of physical texture vector flow fields. This architecture enables the implicit sampling encoder and decoder layers to complete feature extraction and folding without altering the two-dimensional Cartesian coordinate system, achieving zero interpolation information loss extraction for submicron defects. Phase calculation is performed using the Poisson equation and light intensity derivative matrix built into the physical computation layer, and closed-loop verification is performed by substituting the Fresnel diffraction integral formula into the differentiable forward derivation layer, achieving three-dimensional phase reconstruction. The extreme value threshold is accurately generated by fitting the generalized Pareto distribution model using extreme value theory, and the compensation amount for micro-motion defocusing step size and exposure time is calculated by combining the discrete proportional integral control equation. Without introducing huge computational overhead, this invention endows the production line inspection system with adaptive capability without human intervention and possesses industrial robustness.

[0020] In a specific embodiment, step S1 specifically includes: Step S11: The cloud server obtains the micro-motion defocus step size and industrial camera exposure time updated in the previous batch, and uses them as the micro-motion defocus step size and industrial camera exposure time for the current batch, respectively. Step S12: Package the micro-motion defocusing step size and the industrial camera exposure time of the current batch into an execution instruction package and send it to the edge controller. Step S13: The edge controller parses and executes the instruction packet to control the micro-motion platform arranged along the detection axis to run to the displacement point, which includes the negative displacement point, the zero point and the positive displacement point. Step S14: When the micro-motion displacement platform moves to the displacement point, the industrial camera is triggered to sequentially acquire the negative defocus image, positive focus image and positive defocus image of the micro bearing under test, and generate a defocus intensity image sequence in a two-dimensional Cartesian coordinate system.

[0021] Specifically, packaging the micro-motion defocus step size and industrial camera exposure time into an execution instruction package includes: The micro-motion defocus step size and industrial camera exposure time are converted into hexadecimal data format. According to the industrial communication protocol standard, the hexadecimal data is combined with the communication frame header, device physical address code and cyclic redundancy check code into a byte stream and serialized to generate a continuous binary data payload with execution timing, thus obtaining the execution instruction packet.

[0022] In a specific embodiment, step S2 specifically includes: Step S21: Perform multi-scale logarithmic Gabor filtering on the positive focus image to extract the real part convolution response matrix and the imaginary part convolution response matrix of the positive focus image at each filter scale. Based on the real part convolution response matrix and the imaginary part convolution response matrix, calculate the local amplitude and local phase of each pixel in the positive focus image at various filter scales; Calculate the sum of local amplitudes of each pixel in the in-focus image at all filter scales, and calculate the phase alignment energy of the local phase at all filter scales; Divide the phase alignment energy by the sum of local amplitudes to calculate the phase consistency value of the pixels in the positive focus image. Obtain the phase consistency values ​​of all pixels to form a two-dimensional phase consistency feature map. Non-maximum suppression calculation is performed on the two-dimensional phase consistency feature map to extract the connected extremum skeleton line of the outer contour of the micro bearing under test, and the rotational geometric center coordinates of the connected extremum skeleton line are calculated by ellipse fitting least squares method. Specifically, multi-scale log-Gapper filtering is performed on the in-focus image to extract the real and imaginary convolutional response matrices of the in-focus image at each filter scale. This includes: The positive-focus image in two-dimensional Cartesian coordinates is transformed from the spatial domain to the frequency domain using Fast Fourier Transform (FFT) to obtain the frequency domain image. Logarithmic Gabor filter transfer functions with different center frequencies and directional angles are constructed in the frequency domain. The frequency domain image is then multiplied by the transfer function at each filter scale using complex dot multiplication. The inverse Fast Fourier Transform is performed on each product result to restore it to the spatial domain. The complex matrix output by the inverse transform is then separated into real and imaginary components. The real and imaginary convolution response matrices of the positive-focus image at each filter scale are then extracted.

[0023] Calculating the local amplitude and local phase of each pixel in a focused image at various filter scales specifically includes: Extract the real and imaginary convolution response matrix values ​​corresponding to each pixel coordinate position in the two-dimensional Cartesian coordinate system at a specific filter scale. Add the squares of the real and imaginary convolution response matrix values ​​and perform a square root algebra operation to obtain the local amplitude of the pixel. Call the bivariate arctangent function to calculate the ratio angle of the real and imaginary convolution response matrix values ​​to obtain the local phase angle of the pixel in the polar coordinate domain.

[0024] Calculating the phase alignment energy of the local phase across all filter scales specifically includes: The local phase angle of a single pixel is extracted at all filter scales. The local amplitude value of the pixel at each corresponding filter scale is used as the calculation weight. The cosine value of the difference between the local phase angle at each scale and the weighted average phase angle of the pixel is calculated. The cosine value of the difference is subtracted from the preset noise compensation threshold and the negative interference is eliminated by using a non-negative truncation function. The truncated result is multiplied by the corresponding local amplitude and summed in all filter scale dimensions to obtain the phase alignment energy of the pixel. The preset noise compensation threshold is preferably twice the statistical value of the ultra-high frequency noise amplitude of the system background.

[0025] Non-maximum suppression calculations are performed on the two-dimensional phase coherence feature map to extract the connected extremum skeleton lines of the outer contour of the micro-bearing under test. Specifically, this includes: Each pixel in the two-dimensional phase consistency feature map is traversed and its spatial gradient direction is calculated. The consistency value of the pixel is compared with that of its neighboring pixels along the spatial gradient direction and its opposite direction. If the value of the pixel is not a local maximum, it is forcibly set to zero. After retaining the local maximum value, a dual-threshold hysteresis tracking algorithm is introduced. Pixels with a phase consistency value greater than or equal to a preset high threshold are defined as strong edge points that absolutely represent the physical true contour. Pixels with values ​​between the preset low threshold and high threshold and which have direct or indirect paths to any strong edge point in the eight-neighbor spatial topology are defined as connected low-threshold weak edge points. Strong edge points that meet the high threshold and connected low-threshold weak edge points are topologically concatenated to remove isolated noise points. A closed contour line with a single pixel width is output as the connected extreme value skeleton line of the outer contour of the micro-bearing under test. The high threshold is preferably set to a floating-point constant of 0.40, and the low threshold is preferably set to one-third of the high threshold value.

[0026] The calculation of the rotational geometric center coordinates of the connected extremum skeleton line using the ellipse fitting least squares method specifically includes: Two-dimensional Cartesian coordinates of all discrete pixels on the connected extremum skeleton line are extracted. These two-dimensional Cartesian coordinates are substituted into the general conic section algebraic equations to construct an overdetermined linear system of equations. The objective functional is to minimize the sum of squared algebraic distances from all discrete coordinate points to the boundary of the fitted ellipse. The extrema of this objective functional are solved using the singular value decomposition algorithm. The geometric parameter matrix of the optimal fitted ellipse is derived, and the coordinates of the absolute center point of the ellipse are analytically extracted from the geometric parameter matrix as the coordinates of the rotational geometric center.

[0027] Step S22: Read the preset optical parameter data, which includes the center wavelength of the light source, the physical pixel size of the industrial camera, and the optical magnification. Step S23: Encapsulate the defocus intensity image sequence, rotation geometric center coordinates, micro-motion defocus step size of the current batch, and preset optical parameter data to obtain the test data package, and send it to the cloud server.

[0028] Specifically, the preset optical parameter data includes the center wavelength of the illumination source, the physical pixel size of the industrial camera's image sensor chip, and the optical magnification of the industrial lens. During the initialization phase, the edge controller reads the factory hardware configuration information from the firmware memory of the camera and lens through a standard application programming interface.

[0029] The encapsulation of the defocus intensity image sequence, rotation geometric center coordinates, current batch micro-defocus step size, and optical parameter data specifically includes: The lossless image compression algorithm is called to perform spatial redundancy reduction compression on the two-dimensional intensity matrices of the negative defocus image, positive focus image, and positive defocus image to obtain the compressed image byte stream. The rotation geometric center coordinates, micro-motion defocus step size, and optical parameter data are serialized into key-value pair metadata text. The compressed image byte stream and the key-value pair metadata text are concatenated into the same data structure container at a fixed offset address, and a cloud transmission routing protocol header is attached to generate the data packet to be tested.

[0030] In a specific embodiment, step S3 specifically includes: Step S31: The cloud server decompresses the data packet to be tested, inputs the defocus intensity image sequence and the micro-motion defocus step size of the current batch into the pre-trained physical unfolding self-supervised network, and calculates the physical space sampling equivalent through the physical computing layer of the pre-trained physical unfolding self-supervised network. The pre-trained physics-expanded self-supervised network includes a physics computation layer, a differential homeomorphism evolution layer, an implicit sampling encoder layer, a decoder layer, and a differentiable forward inference layer. Step S32: Using the center difference algorithm, calculate the light intensity derivative matrix of the defocus intensity image sequence in the physical calculation layer along the direction perpendicular to the two-dimensional Cartesian coordinate system; Step S33: Based on the center wavelength of the light source and the light intensity derivative matrix, perform phase calculation to obtain the initial phase tensor; Step S34: The differential homeomorphic evolution layer obtains the initial phase tensor and the coordinates of the rotational geometric center. Using the coordinates of the rotational geometric center as the pole, an initial concentric ring sampling coordinate system is constructed in a two-dimensional Cartesian coordinate system.

[0031] Specifically, the calculation of the physical space sampling equivalent through the physical computation layer of the pre-trained physical unfolding self-supervised network includes: The physical computation layer of the self-supervised network reads the physical pixel size and optical magnification values ​​of the industrial camera from the decompressed test data packet. It divides the physical pixel size value by the optical magnification value to obtain the physical micro-size represented by each pixel on the physical surface of the micro-bearing under test, and outputs it as the physical space sampling equivalent.

[0032] Using the central difference algorithm, the calculation of the light intensity derivative matrix of the defocus intensity image sequence along the direction perpendicular to the two-dimensional Cartesian coordinate system in the physical computing layer specifically includes: While maintaining strict alignment of rows and columns of pixels in a two-dimensional Cartesian coordinate system, the intensity matrix of the positive defocus image is subtracted from the intensity matrix of the negative defocus image using a matrix subtraction operator to obtain the optical intensity difference matrix. The micro-motion defocus step size is extracted and multiplied by a constant two to obtain the optical longitudinal span. Each element in the optical intensity difference matrix is ​​divided by the optical longitudinal span to complete the discrete differentiation operation in the depth direction and output the light intensity derivative matrix.

[0033] Constructing an initial concentric annular sampling coordinate system in a two-dimensional Cartesian coordinate system with the rotational geometric center coordinates as the poles specifically includes: In a two-dimensional Cartesian coordinate system, the extracted rotational geometric center coordinates are set as the origin of polar coordinates. A multidimensional arithmetic sequence is generated based on the preset radial discrete sampling interval and tangential angle discrete interval. The sine and cosine projection formulas in trigonometric functions are used to map and restore all grid points in the polar coordinate system to floating-point coordinate positions in the Cartesian tensor space, thus constructing an initial concentric ring sampling coordinate system composed of an absolutely circular topological trajectory. The preset radial discrete sampling interval is preferably 1.0 pixel equivalent of the corresponding physical underlying tensor grid point, and the tangential angle discrete interval is preferably the radian value corresponding to 1.0 degree.

[0034] The physical computation layer is a pre-trained physical unfolding self-supervised network responsible for the underlying optical physical mapping and metric standard unification. It receives the defocus intensity image sequence uploaded from the edge, the physical pixel size of the industrial camera, and the optical magnification. It accurately obtains the physical space sampling equivalent by performing algebraic division. It calculates the light intensity derivative matrix of the defocus intensity image sequence along the direction perpendicular to the two-dimensional Cartesian coordinate system using the central difference algorithm. It calculates the physical wavenumber of the light wave based on the center wavelength of the light source. It multiplies the physical wavenumber with the light intensity derivative matrix and takes the negative value to obtain the source term matrix. Based on the source term matrix and under the constraints of the two-dimensional Laplacian operator and divergence operator with regularization parameters, it constructs the Poisson equation, thereby directly mathematically solving the initial phase tensor in the two-dimensional Cartesian coordinate system.

[0035] The differential homeomorphic evolution layer is a topological evolution computation layer in the network used to perceive real physical clamping deformation and generate adaptive dynamic sampling indexes. It obtains the initial phase tensor and the coordinates of the rotation geometric center, and constructs an initial concentric ring sampling coordinate system with this center as the pole in a two-dimensional Cartesian coordinate system. It calculates the first-order partial derivative of the initial phase tensor to obtain the spatial gradient matrix and extracts its tangential and normal components to construct the physical texture vector flow field. This flow field is used as a two-dimensional evolution velocity field and is spatially convolved and smoothed using a preset Gaussian smoothing kernel to generate a tear-resistant smooth evolution velocity field. Finally, it extracts the velocity vector corresponding to the coordinates of each grid point in the smooth evolution velocity field, and performs spatial ordinary differential iteration and numerical truncation operations on the edge of the two-dimensional Cartesian coordinate system in combination with a preset evolution time step until the maximum number of iterations is reached. Then, it outputs the final grid point coordinate matrix as a flexible implicit sampling grid that perfectly fits the real blade pattern deformation.

[0036] The implicit sampling encoder layer is an abstract dimensionality reduction layer in the network responsible for avoiding the loss of traditional coordinate transformation interpolation and extracting high-dimensional physical manifold features. Its specific function is to simultaneously acquire the initial phase tensor and the flexible implicit sampling grid, directly use the coordinates of the target grid points in the flexible implicit sampling grid as the spatial extraction index, accurately address and extract the low-level pixel values ​​at the positions that completely coincide with the coordinates of the target grid points in the initial phase tensor, and perform one-dimensional expansion and concatenation to generate one-dimensional sequence features. Then, the one-dimensional sequence features are input into the self-attention mechanism network architecture to perform mutually independent linear projection mapping and multi-head attention weight distribution calculation, thereby extracting the target latent vector with highly condensed physical texture fluctuation and potential defect information of the micro bearing surface in the extremely high-dimensional mathematical abstract space.

[0037] The decoder layer is a zero-interpolation spatial folding layer in the network responsible for reverse reconstruction of high-dimensional abstract features strictly according to the original physical evolution topology path. Its specific function is to synchronously receive the target latent vector from the implicit sampling encoder layer and the flexible implicit sampling grid from the differential homeomorphism evolution layer. According to the target grid point coordinate index mapping relationship in the flexible implicit sampling grid, it performs spatial reverse addressing operation, and fills the feature element values ​​contained in the target latent vector after network decoding and expansion back into the spatial matrix grid in the original two-dimensional Cartesian coordinate system that completely corresponds to the coordinate position of each target grid point. Thus, without producing any pixel-level grayscale interpolation smoothing loss, the output size is equivalent to the original input Cartesian reconstructed phase tensor.

[0038] The differentiable forward inference layer is a virtual simulation operator layer at the end of the network responsible for rigorous closed-loop verification of the optical physical macroscopic self-consistency of the reconstructed phase tensor. Its specific function is to receive the Cartesian reconstructed phase tensor and use it as the initial complex amplitude phase angle of the spatial light field. Combined with the physical space sampling equivalent with absolute metric scale and the center wavelength parameter of the light source obtained from the physical calculation layer, it strictly calls the Fresnel diffraction integral formula in the two-dimensional Cartesian space frequency domain to simulate the physical propagation evolution law of light waves in real free space. By mathematically calculating the forward and backward light wave transfer functions and extracting the modulus square operator of the final interference intensity, it accurately simulates and generates predicted negative defocus images and predicted positive defocus images for comparison with the mean square error of the real acquired images.

[0039] In a specific embodiment, step S33 specifically includes: The physical wave number of the light wave is calculated based on the center wavelength of the light source. The physical wave number is then multiplied by the light intensity derivative matrix. The result is then negative to obtain the source term matrix. By constructing the Poisson equation using the source term matrix, the initial phase tensor can be solved, and its calculation expression is as follows: ; in, For the initial phase tensor, For a two-dimensional Laplace operator, The source term matrix, The coordinates are in a two-dimensional Cartesian coordinate system. The optical intensity matrix of the positive focus image. For regularization parameters, For divergence operators, This is the spatial gradient operator.

[0040] Specifically, the regularization parameter is a non-zero minimum constant, and the spatial gradient operator is a first-order spatial partial differential operator in vector calculus. In a two-dimensional Cartesian coordinate system, its mathematical definition is a two-dimensional vector function formed by taking the first-order partial derivatives of the target scalar field along the horizontal X-axis and the vertical Y-axis.

[0041] It should be noted that in traditional technologies, the phase distribution of the three-dimensional microscopic height features on the surface of micro-bearings is usually obtained by using confocal microscopy and black-box neural networks to obtain three-dimensional point clouds. This method has the problems of extremely high detection hardware costs and high confidence level misjudgment due to interference from ambient light. In contrast, this invention fuses the physical wavenumber of the real light wave and the light intensity derivative matrix along the depth direction by multiplying the source term matrix based on the physical laws of optical intensity transmission. The resulting source term matrix is ​​then used to construct a Poisson equation and solve for the initial phase tensor. Traditional methods for constructing the Poisson equation and solving for the initial phase tensor usually update the value by directly dividing the matrix gradient by the intensity matrix of the positive focus image. This method suffers from the technical problem that when there are extremely dark pixel areas in the image, the denominator approaches zero, leading to rapid numerical divergence and low-frequency artifacts. However, the gradient division anti-divergence update method adopted in this invention, which introduces a non-zero minimum constant as a regularization parameter constraint, can force smoothing and isolate extreme numerical abrupt changes in low-brightness noise areas during two-dimensional fast Fourier inversion operations. This effectively solves the problem of global matrix solution failure caused by local contrast collapse during optical computation inversion.

[0042] In a specific embodiment, step S4 specifically includes: Step S41: Perform first-order partial derivative operation on the initial phase tensor to obtain the spatial gradient matrix, extract the tangential and normal components of the spatial gradient matrix, and construct the physical texture vector flow field on the two-dimensional Cartesian coordinate system. A flexible implicit sampling mesh is obtained by iterating the initial concentric annular coordinate system using the physical texture vector flow field through spatial ordinary differential equations. Step S42: The implicit sampling encoder layer obtains a flexible implicit sampling grid, uses the coordinates of the target grid points in the flexible implicit sampling grid as an index to obtain the target grid point coordinate index relationship, and extracts the pixel values ​​at the positions that coincide with the coordinates of the target grid points in the initial phase tensor and concatenates them to generate a one-dimensional sequence feature. The target latent vector is extracted by linear projection and attention weight calculation of one-dimensional sequence features. In step S43, the decoder layer receives the target latent vector and the flexible implicit sampling grid. Based on the coordinate index relationship of the target grid points in the flexible implicit sampling grid, the feature element values ​​in the target latent vector are filled into the corresponding coordinate positions in the two-dimensional Cartesian coordinate system one by one to obtain the Cartesian reconstructed phase tensor.

[0043] Specifically, performing first-order partial derivative operations on the initial phase tensor includes: For the initial phase tensor in a two-dimensional Cartesian coordinate system, the physical unfolding self-supervised network calls the Sober operator to perform two-dimensional spatial sliding window convolution operations with the matrix data of the initial phase tensor along the horizontal X-axis and vertical Y-axis, respectively. The local phase space change rate in the horizontal direction and the local phase space change rate in the vertical direction of each tensor coordinate point are calculated respectively. The numerical matrices of the change rates in these two orthogonal directions are combined by channel merging at corresponding positions to output the spatial gradient matrix.

[0044] Extracting the tangential and normal components of the spatial gradient matrix specifically includes: The horizontal and vertical partial derivatives of each coordinate point in the spatial gradient matrix are extracted. The normal angle of the point is calculated using the bivariate arctangent function. At the same time, the normal gradient angle is derived by performing a clockwise 90-degree orthogonal deflection in the two-dimensional geometric plane. The tangential angle parallel to the direction of the micro-machining tool marks is calculated. Combined with the local gradient magnitude calculated by the square root of the sum of squares, the normal and tangential angles are mathematically mapped back into two-dimensional vectors with direction and magnitude attributes. In this way, the tangential and normal components are accurately separated and output.

[0045] Constructing a physical texture vector flow field in a two-dimensional Cartesian coordinate system specifically includes: The scalar values ​​of the tangential and normal components of the spatial gradient matrix at each discrete coordinate point in a two-dimensional Cartesian coordinate system are extracted. The scalar value of the tangential component is taken as the magnitude of the local geometric flow velocity vector in the horizontal X-axis direction of the coordinate point, and the scalar value of the normal component is taken as the magnitude of the local geometric flow velocity vector in the vertical Y-axis direction of the coordinate point. The two mutually orthogonal unidirectional velocity vectors are merged into a comprehensive two-dimensional fluid velocity vector by the principle of vector synthesis. Under the premise of keeping the original spatial mapping topology of the two-dimensional Cartesian coordinate system absolutely unchanged, the comprehensive two-dimensional fluid velocity vectors corresponding to all pixel coordinate points are arranged in a global mathematical array to construct the physical texture vector flow field.

[0046] The linear projection and attention weight calculation of one-dimensional sequence features specifically include: One-dimensional sequence features extracted by the flexible implicit sampling grid are input into the attention mechanism module. Through three independent parameterized linear mapping matrices of the underlying weights, the one-dimensional sequence features are projected into a query matrix, a key matrix, and a value matrix, respectively. A matrix dot product is performed between the query matrix and the transposed key matrix, and the result is divided by a dimension-based square root scaling factor to calculate the global topological similarity score between feature nodes in the sequence. Then, the Softmax normalized activation function is called to perform a probability distribution mapping on the similarity scores to generate an attention weight matrix. A weighted summation matrix multiplication operation is performed between the attention weight matrix and the value matrix to achieve high-dimensional association extraction of cross-spatial physical features. The weights at this point are a set of learnable floating-point parameter features that have been iteratively optimized during the pre-training phase using backpropagation combined with gradient descent of the loss function and ultimately solidified within the model.

[0047] The feature element values ​​in the target latent vector refer to the pure mathematical representation floating-point values ​​of the original optical phase and physical morphology information of the micro-bearing surface after multiple self-attention encodings and nonlinear dimensionality reduction processes by a deep neural network, mapped onto an extremely high-dimensional abstract manifold space.

[0048] In a specific embodiment, the spatial ordinary differential iterative evolution of the initial concentric annular coordinate system through the physical texture vector flow field specifically includes: Extract the coordinates of each grid point in the initial concentric annular coordinate system and construct the initial coordinate matrix; The physical texture vector flow field is used as a two-dimensional evolution velocity field. A preset Gaussian smoothing kernel is used to perform spatial convolution smoothing on the two-dimensional evolution velocity field to generate a smooth evolution velocity field. Extract the velocity vector of each grid point's coordinate in the smoothed evolving velocity field, and combine it with the grid point coordinates of the current iteration to calculate the updated grid point coordinates. The calculation expression is as follows: ; in, For the first The updated grid point coordinates obtained from the next iteration. For the first The coordinates of the current grid point at the next iteration, when hour, The initial grid point coordinates, To preset the evolution time step, To smooth the evolution of the velocity field, The coordinates of the current grid point The velocity vector at the corresponding position in the smooth evolution velocity field; Perform numerical truncation of the edges of the updated grid point coordinates in a two-dimensional Cartesian coordinate system; Determine if the current iteration count has reached the maximum iteration count: If the target is not reached, the updated grid point coordinates will be used as the input for the next iteration, and spatial convolution smoothing will be returned for the next iteration. If the target is reached, stop the iteration, output the final grid point coordinate matrix, and use the final grid point coordinate matrix as the flexible implicit sampling grid.

[0049] Specifically, the matrix size of the preset Gaussian smoothing kernel is preferably 5×5 pixels and its corresponding Gaussian distribution standard deviation parameter is preferably set to 2.0. In the process of ordinary differential fluid evolution, the topological consistency of the motion velocity between adjacent grid points is forcibly constrained to absolutely prevent the mathematical sampling grid from non-physical intersection or tearing when conforming to the knife pattern distortion. The spatial convolution smoothing of the two-dimensional evolution velocity field using a preset Gaussian smoothing kernel specifically includes: The two-dimensional evolution velocity field is split into a horizontal velocity matrix and a vertical velocity matrix. A preset Gaussian smoothing kernel is used as a sliding window to perform pixel-by-pixel two-dimensional spatial sliding on the horizontal and vertical velocity matrices. At each center-aligned position, the Gaussian weight value in the window is multiplied and accumulated element-by-element with the corresponding underlying evolution velocity value. The calculated weighted average smoothed velocity value is used to replace the original aggressive velocity value at the center point. Finally, the smoothed horizontal and vertical velocity components are reassembled to output a smoothed evolution velocity field.

[0050] The maximum number of iterations is preferably 20. The maximum number of iterations should not be too small, which would cause the flexible mesh to fail to fully fit the deformation trajectory of the actual machining tool marks, nor should it be too large, which would lead to unnecessary waste of computing resources or even cause overfitting and topological distortion collapse.

[0051] It should be noted that in current technologies, polar coordinate unfolded image matrices representing the machining texture of circular workpieces are typically obtained through bilinear interpolation and polar coordinate geometric transformation. However, this invention integrates large deformation differential homeomorphism measurement technology and ordinary differential iteration mechanism into the topological evolution of the initial concentric ring coordinate system. This results in a flexible implicit sampling grid that corresponds to the elastic deformation of the actual physical clamping of the micro-bearing and the actual machining tool marks. The data in this flexible implicit sampling grid can simultaneously characterize the topological mapping relationship of the physical morphology space and the meaning of the original pixel extraction guidance for high-frequency weak defects. This solves the problems of submicron level weak defects being smoothed out by interpolation and the serious mismatch between rigid grids and actual deformations in the traditional pixel-level resampling process. By directly extracting pixels in the original two-dimensional Cartesian coordinate system and directly restoring them to the corresponding positions, the innovative effect of maximally preserving the characteristics of high-frequency physical micro-defects and accurately locating them is achieved.

[0052] In a specific embodiment, step S4 further includes: Step S44: Obtain the set of extreme values ​​for the normal state. This set includes the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the Mahalanobis distance sequence. The Mahalanobis distance scalar of the micro-bearing under test is calculated using the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the target latent vector. Its calculation expression is as follows: ; in, The Mahalanobis distance scalar of the miniature bearing under test. Let be the target latent vector. The mean of the high-dimensional latent vector. For normal covariance matrix, superscript For transpose operation, superscript To perform the inverse operation; Step S45: Extract the Mahalanobis distance sequence from the set of extreme values ​​of the normal state, and truncate the Mahalanobis distance sequence that exceeds the set baseline to obtain the tail distribution sequence. Using the maximum likelihood estimation algorithm, the distribution parameters of the generalized Pareto distribution are solved based on the tail distribution sequence to obtain the target extreme value distribution function. Combined with the preset confidence level constant, the distribution quantile value corresponding to the target extreme value distribution function is solved, and the distribution quantile value is used as the extreme value threshold. If the Mahalanobis distance scalar is greater than the extreme value threshold, subtract the Cartesian reconstructed phase tensor from the initial phase tensor in the two-dimensional Cartesian coordinate system and take the absolute value to generate the Cartesian anomalous thermodynamic matrix. Step S46: Extract the row and column index subscripts of the abnormal points in the Cartesian abnormal thermal matrix that are greater than the set value in the two-dimensional Cartesian coordinate system, and perform coordinate transformation by combining the physical space sampling equivalent and the coordinates of the rotational geometric center to obtain the absolute defect coordinates of the surface of the micro bearing to be tested.

[0053] Specifically, the set of extreme values ​​for the normal state is a prior dataset stored in a cloud-based in-memory database. This dataset is created during the offline database building phase by extracting latent vectors and calculating mathematical distributions from thousands of known good-quality micro-bearing images that have been confirmed by physical flaw detection to be absolutely free of microscopic defects. The cloud server then uses an internal high-speed data bus to call a high-concurrency data query interface to directly load and extract the set of extreme values ​​for the normal state from the prior dataset in the cloud-based in-memory database.

[0054] The baseline is preferably the distance value corresponding to the 95th percentile of the Mahalanobis distance sequence of historical normal samples arranged in ascending order. In the extreme value theory, the peak value method is used to forcibly remove most of the data samples that belong to the normal processing tolerance fluctuation and only extract a very small number of extreme fluctuation data at the edge of the probability distribution for tail distribution modeling.

[0055] The method of solving the distribution parameters of the generalized Pareto distribution based on the tail distribution sequence using the maximum likelihood estimation algorithm specifically includes: All extreme Mahalanobis distance differences exceeding the set baseline in the tail distribution sequence are extracted as actual observation samples. These are then multiplied in the probability density analytical function of the generalized Pareto distribution to construct a joint likelihood objective functional. Subsequently, the natural logarithm of this joint likelihood objective functional is performed to convert it into a computationally stable log-likelihood function. Next, the first-order partial derivatives of the unknown shape and scale parameters in the log-likelihood function are calculated to form a system of nonlinear gradient differential equations. Finally, a quasi-Newton iterative optimization algorithm is called to perform a multi-step gradient ascent search in the parameter solution space until the convergence coordinates of the point where the log-likelihood function reaches its absolute maximum are found. These coordinate values ​​are then used as the optimal shape and scale parameters for output.

[0056] Based on a preset confidence level constant, the specific steps for solving the distribution quantile values ​​corresponding to the target extreme value distribution function include: The preset confidence level constant is preferably set to 99.9% of the extremely high statistical confidence interval. In the specific solution operation, the optimal shape parameter and scale parameter are substituted into the cumulative distribution function model of the generalized Pareto distribution to construct the target extreme value distribution function with determined parameters. Then, the preset confidence level constant is used as the known cumulative probability ordinate and substituted into the inverse function equation of the target extreme value distribution function for algebraic inverse solution to calculate the relative critical distance increment beyond the set baseline. The set baseline value and the relative critical distance increment are added together to obtain the absolute numerical boundary, which is the distribution quantile value.

[0057] If the Mahalanobis distance scalar is less than or equal to the extreme value threshold, the current micro-bearing under test is determined to be a qualified product. The system will then terminate the calculation process and encapsulate and send the qualified product determination signal to the production line to execute the green light release and robot good product sorting instructions.

[0058] The preferred setting is an adaptive relative floating-point value represented by 35% of the peak value of the global maximum residual of the Cartesian anomalous thermodynamic matrix, in order to effectively filter out the global uniform error caused by minor environmental vibrations without missing extremely weak dot-like scratches.

[0059] By combining the physical space sampling equivalent with the coordinates of the rotational geometric center, the absolute defect coordinates of the surface of the micro-bearing under test are obtained through coordinate transformation, specifically including: The integer row and column indices of outliers exceeding a set value in the pixel matrix of a two-dimensional Cartesian coordinate system are extracted. The horizontal pixel offset of the defect is obtained by subtracting the pre-calculated horizontal pixel value of the rotation geometry center coordinate from the column index. The vertical pixel offset of the defect is obtained by subtracting the vertical pixel value of the rotation geometry center coordinate from the row index. Then, the horizontal and vertical pixel offsets are multiplied by the physical space sampling equivalent value containing physical dimensions calculated by the physical unfolding self-supervised network. The physical meaningless relative pixel distance is strictly converted into the real physical distance in the space with micrometers or nanometers as the absolute unit of measurement. The absolute physical two-dimensional position coordinates of the surface defect of the micro-bearing under test relative to its own physical rotation center are calculated.

[0060] In a specific embodiment, step S5 specifically includes: Step S51: The differentiable forward derivation layer receives the Cartesian reconstructed phase tensor, combines the physical space sampling equivalent and the center wavelength of the light source, and uses the Fresnel diffraction integral formula to calculate the forward and backward light wave transfer functions in the two-dimensional Cartesian coordinate system to generate the predicted negative defocus image and the predicted positive defocus image. Step S52: Calculate the sum of the mean square errors between the negative defocus image and the positive defocus image and the predicted negative defocus image and the predicted positive defocus image, respectively. Step S53: Calculate the original reconstruction confidence score using mean square error, normalize the original reconstruction confidence score to obtain the reconstruction confidence score, and calculate the control deviation between the preset confidence threshold and the reconstruction confidence score. Step S54: Substitute the control deviation into the preset discrete proportional-integral control equation to calculate the micro-motion defocusing step length compensation and exposure time compensation. Add the micro-motion defocus step size compensation and the exposure time compensation to the current batch's micro-motion defocus step size and the current batch's industrial camera exposure time, respectively, to obtain the updated micro-motion defocus step size and industrial camera exposure time.

[0061] Specifically, using the Fresnel diffraction integral formula, the calculation of the forward and backward light wave transfer functions in a two-dimensional Cartesian coordinate system includes: The Cartesian reconstructed phase tensor is used as the phase angular component of the complex exponent term, and combined with a preset uniform amplitude constant to construct an initial spatial complex amplitude matrix. A two-dimensional fast Fourier transform operator is invoked to transform the initial spatial complex amplitude matrix from the spatial domain to the spatial frequency domain. In the frequency domain, a free-space frequency domain transfer function matrix containing a quadratic phase bias factor is constructed based on the previously extracted light source center wavelength and the physical space sampling equivalent. The positive and negative micro-motion defocusing step sizes are substituted into the frequency domain transfer function matrix to generate two diffraction kernel matrices with opposite propagation directions. The spatial frequency domain complex amplitude matrix is ​​then multiplied by the two diffraction kernel matrices using a complex array of Hadamard dot products. The two dot products are then subjected to a two-dimensional fast Fourier inverse transform to restore the two-dimensional Cartesian spatial domain, and the squared-modulus matrix of the complex values ​​is extracted and calculated. This generates predicted negative defocus and predicted positive defocus images representing the interference intensity of the forward and backward physical propagation evolution of the light wave. The preset uniform amplitude constant is preferably set to 1.0.

[0062] The mean square error between the negative and positive out-of-focus images and the predicted negative and positive out-of-focus images is calculated, specifically including: The corresponding pixel intensity values ​​of the negative defocus image and the predicted negative defocus image in the two-dimensional Cartesian coordinate system are extracted and subjected to pixel-by-pixel matrix subtraction to obtain the negative defocus residual matrix. The value of each element in the negative defocus residual matrix is ​​squared, summed globally, and divided by the total number of pixels in the image to calculate the negative mean square error scalar representing the forward inference physical deviation. Similarly, the positive defocus image and the generated predicted positive defocus image are extracted and subjected to the same pixel-by-pixel subtraction, squaring, summation, and mean-averaging algebraic operations to calculate the positive mean square error scalar representing the backward inference physical deviation. Finally, the negative mean square error scalar and the positive mean square error scalar are arithmetically added together to calculate the mean square error sum.

[0063] The calculation of the original reconstruction confidence score using mean squared error specifically includes: A negative exponential decay function with the natural constant as the base is introduced. The calculated mean square error is multiplied by a preset sensitivity adjustment negative constant factor and then used as the exponent of the natural constant. This makes the function output value exhibit a non-linear, extremely rapid, smooth convergence decay as the mean square error and scalar increase linearly in the mathematical mapping topological characteristics. This ensures that in the process of physical light wave reverse deduction verification, even a small non-physical self-consistent residual will cause a sharp drop in the calculation score. Thus, the originally infinitely large pure absolute error scalar is scientifically and extremely sensitively mapped to the original reconstruction confidence score. The preset sensitivity adjustment negative constant factor is preferably set to -15.0.

[0064] The preset calculation expression for the discrete proportional-integral control equation is: ; in, For the micro-focusing step size compensation or exposure time compensation amount output by the kth batch of calculations, the micro-focusing step size and exposure time are respectively substituted into independent gain constants to perform two isomorphic equation calculations. To control the deviation, It is the proportional gain constant. This is the summation integral term of all historical control deviations from the start batch to the current k-th batch on the discrete time axis. The integral gain constant; Independent gain constants include proportional gain constant and integral gain constant.

[0065] Regarding the micro-motion defocusing step length, to absolutely prevent the underlying piezoelectric ceramic servo motor from generating mechanical oscillations exceeding its physical resonant frequency and long-term structural fatigue damage during rapid compensation response, the proportional gain constant is preferably set to 0.1. To only gently eliminate the zero-point drift of long-range mechanical thermal expansion caused by the alternating day and night temperature changes in the workshop environment, the integral gain constant is preferably set to 0.15. Regarding the exposure time, to enable rapid and delay-free compensation for the instantaneous decay of ambient light in the production line or the contamination of the lens dust cover with oil and dust, the proportional gain constant is preferably set to 0.1. To accurately address the extremely slow physical light decay curve of the LED lighting source for machine vision in the production line over its lifespan of tens of thousands of hours, the integral gain constant is preferably set to 0.07.

[0066] It should be noted that in current technologies, static decision thresholds and preset fixed hardware control parameters are typically set manually to obtain binary segmentation boundaries representing the criteria for good and defective products and the system's operating status, as well as the instructions for executing equipment. However, this invention integrates the generalized Pareto distribution fitting of extreme value theory and discrete proportional-integral control equations into high-dimensional manifold metric decision and optical forward deduction physical self-checking feedback. This results in extreme value thresholds and micro-motion defocusing step size and exposure time compensation amounts that can control the deviation amount based on the normal processing tolerance tail distribution and reconstructed confidence. The data in these adaptive extreme values ​​and compensation values ​​can simultaneously characterize the meaning of the pure mathematical objective quality interception extreme value boundary and the instantaneous suppression increment of the production line's underlying software and hardware error coupling. Through rigorous mathematical statistical evolution and classic PID steady-state compensation mechanisms, an innovative effect is achieved that maintains extremely high detection robustness even when the system experiences environmental noise drift and material property changes.

[0067] Example 2, refer to Figure 2 It provides a cloud-based micro-bearing manufacturing and inspection data processing system, including an image acquisition module, a data encapsulation module, a phase calculation module, a defect location module, and a closed-loop control module; The image acquisition module is used to control the micro-displacement platform at the edge based on the execution instruction packets sent by the cloud server, and to acquire the defocus intensity image sequence of the micro bearing under test; The data encapsulation module is used to calculate the rotational geometric center coordinates of the in-focus image, encapsulate it together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; The phase calculation module is used to input the data packet to be tested into the pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in the two-dimensional Cartesian coordinate system based on the preset optical parameter data, and construct the initial concentric ring coordinate system. The defect localization module is used to perform spatial ordinary differential iteration on the initial concentric ring coordinate system through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and extract the target latent vector and fold it into a Cartesian reconstructed phase tensor; If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. The closed-loop control module is used to deduce the Cartesian reconstructed phase tensor into a predicted defocused image, calculate the reconstruction confidence score using the mean square error, and substitute it into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocusing step size and industrial camera exposure time.

[0068] This invention utilizes a flexible implicit sampling mesh generation architecture based on spatial ordinary differential iterative evolution of physical texture vector flow fields. This architecture enables the implicit sampling encoder and decoder layers to complete feature extraction and folding without altering the two-dimensional Cartesian coordinate system, achieving zero interpolation information loss extraction for submicron defects. Phase calculation is performed using the Poisson equation and light intensity derivative matrix built into the physical computation layer, and closed-loop verification is performed by substituting the Fresnel diffraction integral formula into the differentiable forward derivation layer, achieving three-dimensional phase reconstruction. The extreme value threshold is accurately generated by fitting the generalized Pareto distribution model using extreme value theory, and the compensation amount for micro-motion defocusing step size and exposure time is calculated by combining the discrete proportional integral control equation. Without introducing huge computational overhead, this invention endows the production line inspection system with adaptive capability without human intervention and possesses industrial robustness.

[0069] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A cloud computing-based method for processing manufacturing and inspection data of micro-bearings, characterized in that, Includes the following steps: Step S1: Based on the execution instruction packet sent by the cloud server, control the micro-motion displacement platform at the edge end to acquire the defocus intensity image sequence of the micro-bearing under test; Step S2: Calculate the rotational geometric center coordinates of the in-focus image, encapsulate them together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; Step S3: Input the data packet to be tested into the pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in the two-dimensional Cartesian coordinate system based on the preset optical parameter data, and construct the initial concentric ring coordinate system. Step S4: The initial concentric ring coordinate system is subjected to spatial ordinary differential iteration through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and the target latent vector is extracted and folded into a Cartesian reconstructed phase tensor. If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. Step S5: The Cartesian reconstructed phase tensor is derived into the predicted defocus image. The reconstruction confidence score is calculated using the mean square error. The score is then substituted into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocus step size and industrial camera exposure time.

2. The cloud computing-based micro-bearing manufacturing and testing data processing method as described in claim 1, characterized in that, Step S1 specifically includes: Step S11: The cloud server obtains the micro-motion defocus step size and industrial camera exposure time updated in the previous batch, and uses them as the micro-motion defocus step size and industrial camera exposure time for the current batch, respectively. Step S12: Package the micro-motion defocusing step size and the industrial camera exposure time of the current batch into an execution instruction package and send it to the edge controller. Step S13: The edge controller parses and executes the instruction packet to control the micro-motion platform arranged along the detection axis to run to the displacement point, which includes a negative displacement point, a zero point, and a positive displacement point. Step S14: When the micro-motion displacement platform moves to the displacement point, the industrial camera is triggered to sequentially acquire the negative defocus image, positive focus image and positive defocus image of the micro bearing under test, and generate a defocus intensity image sequence in a two-dimensional Cartesian coordinate system.

3. The cloud computing-based micro-bearing manufacturing and testing data processing method as described in claim 2, characterized in that, Step S2 specifically includes: Step S21: Perform multi-scale log-Gapper filtering on the positive focus image to extract the real part convolution response matrix and the imaginary part convolution response matrix of the positive focus image at each filter scale. Based on the real part convolution response matrix and the imaginary part convolution response matrix, the local amplitude and local phase of each pixel in the positive focus image are calculated at each filter scale; Calculate the sum of local amplitudes of each pixel in the in-focus image at all filter scales, and calculate the phase alignment energy of the local phase at all filter scales; Divide the phase alignment energy by the sum of local amplitudes to calculate the phase consistency value of each pixel in the positive focus image. Obtain the phase consistency values ​​of all pixels to form a two-dimensional phase consistency feature map. Non-maximum suppression calculation is performed on the two-dimensional phase consistency feature map to extract the connected extremum skeleton line of the outer contour of the micro bearing under test, and the rotational geometric center coordinates of the connected extremum skeleton line are calculated by the least squares ellipse fitting operator. Step S22: Read preset optical parameter data, which includes the center wavelength of the light source, the physical pixel size of the industrial camera, and the optical magnification. Step S23: Encapsulate the defocus intensity image sequence, rotation geometric center coordinates, micro-motion defocus step size of the current batch, and preset optical parameter data to obtain the test data package, and send it to the cloud server.

4. The cloud computing-based method for processing micro-bearing manufacturing and testing data as described in claim 3, characterized in that, Step S3 specifically includes: Step S31: The cloud server decompresses the data packet to be tested, inputs the defocus intensity image sequence and the micro-motion defocus step size of the current batch into the pre-trained physical unfolding self-supervised network, and calculates the physical space sampling equivalent through the physical computing layer of the pre-trained physical unfolding self-supervised network. The pre-trained physics unfolding self-supervised network includes a physics computation layer, a differential homeomorphism evolution layer, an implicit sampling encoder layer, a decoder layer, and a differentiable forward inference layer. Step S32: Using the center difference algorithm, calculate the light intensity derivative matrix of the defocus intensity image sequence in the physical calculation layer along the direction perpendicular to the two-dimensional Cartesian coordinate system; Step S33: Based on the center wavelength of the light source and the light intensity derivative matrix, perform phase calculation to obtain the initial phase tensor; Step S34: The differential homeomorphic evolution layer obtains the initial phase tensor and the coordinates of the rotational geometric center. Using the coordinates of the rotational geometric center as the pole, an initial concentric ring sampling coordinate system is constructed in a two-dimensional Cartesian coordinate system.

5. The cloud computing-based micro-bearing manufacturing and testing data processing method as described in claim 4, characterized in that, Step S33 specifically includes: The physical wave number of the light wave is calculated based on the center wavelength of the light source. The physical wave number is then multiplied by the light intensity derivative matrix. The result is then negative to obtain the source term matrix. By constructing the Poisson equation using the source term matrix, the initial phase tensor can be solved, and its calculation expression is as follows: ； in, For the initial phase tensor, For a two-dimensional Laplace operator, The source term matrix, The coordinates are in a two-dimensional Cartesian coordinate system. The optical intensity matrix of the positive focus image. For regularization parameters, For divergence operators, This is the spatial gradient operator.

6. The cloud computing-based micro-bearing manufacturing and testing data processing method as described in claim 5, characterized in that, Step S4 specifically includes: Step S41: Perform first-order partial derivative operation on the initial phase tensor to obtain the spatial gradient matrix, extract the tangential and normal components of the spatial gradient matrix, and construct the physical texture vector flow field on the two-dimensional Cartesian coordinate system. A flexible implicit sampling mesh is obtained by iterating the initial concentric annular coordinate system using the physical texture vector flow field through spatial ordinary differential equations. Step S42: The implicit sampling encoder layer obtains a flexible implicit sampling grid, uses the coordinates of the target grid points in the flexible implicit sampling grid as an index to obtain the target grid point coordinate index relationship, and extracts the pixel values ​​at the positions that coincide with the coordinates of the target grid points in the initial phase tensor and concatenates them to generate a one-dimensional sequence feature. The target latent vector is extracted by linear projection and attention weight calculation of one-dimensional sequence features. In step S43, the decoder layer receives the target latent vector and the flexible implicit sampling grid. Based on the coordinate index relationship of the target grid points in the flexible implicit sampling grid, the feature element values ​​in the target latent vector are filled into the corresponding coordinate positions in the two-dimensional Cartesian coordinate system one by one to obtain the Cartesian reconstructed phase tensor.

7. The cloud computing-based method for processing micro-bearing manufacturing and inspection data as described in claim 6, characterized in that, The spatial ordinary differential iterative evolution of the initial concentric annular coordinate system through the physical texture vector flow field specifically includes: Extract the coordinates of each grid point in the initial concentric annular coordinate system and construct the initial coordinate matrix; The physical texture vector flow field is used as a two-dimensional evolution velocity field. A preset Gaussian smoothing kernel is used to perform spatial convolution smoothing on the two-dimensional evolution velocity field to generate a smooth evolution velocity field. Extract the velocity vector of each grid point's coordinate in the smoothed evolving velocity field, and combine it with the grid point coordinates of the current iteration to calculate the updated grid point coordinates. The calculation expression is as follows: ； in, For the first The updated grid point coordinates obtained from the next iteration. For the first The coordinates of the current grid point at the next iteration, when hour, The initial grid point coordinates, To preset the evolution time step, To smooth the evolution of the velocity field, The coordinates of the current grid point The velocity vector at the corresponding position in the smooth evolution velocity field; Perform numerical truncation of the edges of the updated grid point coordinates in a two-dimensional Cartesian coordinate system; Determine if the current iteration count has reached the maximum iteration count: If the target is not reached, the updated grid point coordinates will be used as the input for the next iteration, and spatial convolution smoothing will be returned for the next iteration. If the target is reached, stop the iteration, output the final grid point coordinate matrix, and use the final grid point coordinate matrix as the flexible implicit sampling grid.

8. The cloud computing-based micro-bearing manufacturing and testing data processing method as described in claim 7, characterized in that, Step S4 also includes: Step S44: Obtain the set of extreme values ​​for the normal state. This set includes the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the Mahalanobis distance sequence. The Mahalanobis distance scalar of the micro-bearing under test is calculated using the mean of the high-dimensional latent vectors of historical normal samples, the normal covariance matrix, and the target latent vector. The calculation expression is as follows: ； in, The Mahalanobis distance scalar of the miniature bearing under test. Let be the target latent vector. The mean of the high-dimensional latent vector. For normal covariance matrix, superscript For transpose operation, superscript To perform the inverse operation; Step S45: Extract the Mahalanobis distance sequence from the set of extreme values ​​of the normal state, and truncate the Mahalanobis distance sequence that exceeds the set baseline to obtain the tail distribution sequence. Using the maximum likelihood estimation algorithm, the distribution parameters of the generalized Pareto distribution are solved based on the tail distribution sequence to obtain the target extreme value distribution function. Combined with the preset confidence level constant, the distribution quantile value corresponding to the target extreme value distribution function is solved, and the distribution quantile value is used as the extreme value threshold. If the Mahalanobis distance scalar is greater than the extreme value threshold, subtract the Cartesian reconstructed phase tensor from the initial phase tensor in the two-dimensional Cartesian coordinate system and take the absolute value to generate the Cartesian anomalous thermodynamic matrix. Step S46: Extract the row and column index subscripts of the abnormal points in the Cartesian abnormal thermal matrix that are greater than the set value in the two-dimensional Cartesian coordinate system, and perform coordinate transformation by combining the physical space sampling equivalent and the coordinates of the rotational geometric center to obtain the absolute defect coordinates of the surface of the micro bearing to be tested.

9. The cloud computing-based method for processing micro-bearing manufacturing and testing data as described in claim 8, characterized in that, Step S5 specifically includes: Step S51: The differentiable forward derivation layer receives the Cartesian reconstructed phase tensor, combines the physical space sampling equivalent and the center wavelength of the light source, and uses the Fresnel diffraction integral formula to calculate the forward and backward light wave transfer functions in the two-dimensional Cartesian coordinate system to generate the predicted negative defocus image and the predicted positive defocus image. Step S52: Calculate the sum of the mean square errors between the negative defocus image and the positive defocus image and the predicted negative defocus image and the predicted positive defocus image, respectively. Step S53: Calculate the original reconstruction confidence score using mean square error, normalize the original reconstruction confidence score to obtain the reconstruction confidence score, and calculate the control deviation between the preset confidence threshold and the reconstruction confidence score. Step S54: Substitute the control deviation into the preset discrete proportional-integral control equation to calculate the micro-motion defocusing step length compensation and exposure time compensation. Add the micro-motion defocus step size compensation and the exposure time compensation to the current batch's micro-motion defocus step size and the current batch's industrial camera exposure time, respectively, to obtain the updated micro-motion defocus step size and industrial camera exposure time.

10. A cloud computing-based micro-bearing manufacturing and inspection data processing system, applied in the cloud computing-based micro-bearing manufacturing and inspection data processing method as described in any one of claims 1-9, characterized in that, It includes an image acquisition module, a data encapsulation module, a phase calculation module, a defect location module, and a closed-loop control module; The image acquisition module is used to control the micro-displacement platform at the edge end based on the execution instruction package sent by the cloud server, and to acquire the defocus intensity image sequence of the micro bearing under test; The data encapsulation module is used to calculate the rotational geometric center coordinates of the in-focus image, encapsulate it together with the defocus intensity image sequence and preset optical parameter data into a test data package and upload it to the cloud server; The phase calculation module is used to input the data packet to be tested into a pre-trained physical unfolding self-supervised network, calculate and output the initial phase tensor in a two-dimensional Cartesian coordinate system based on preset optical parameter data, and construct an initial concentric ring coordinate system. The defect localization module is used to perform spatial ordinary differential iteration on the initial concentric ring coordinate system through the physical texture vector flow field to obtain a flexible implicit sampling mesh, and extract the target latent vector and fold it into a Cartesian reconstructed phase tensor. If the Mahalanobis distance of the target latent vector is greater than the extreme value threshold, the tensor difference is calculated to generate an abnormal thermodynamic matrix, and the absolute defect coordinates of the surface of the micro-bearing under test are extracted. The closed-loop control module is used to deduce the Cartesian reconstructed phase tensor into a predicted defocused image, calculate the reconstruction confidence score using the mean square error, and substitute it into the preset discrete proportional-integral control equation to obtain the updated micro-motion defocusing step size and industrial camera exposure time.

Images