Endoscopic image quantification method for gear protective film thickness and uniformity

Abstract

The present application relates to wind turbine gearbox protection technical field, especially to gear protection film thickness and uniformity endoscope image quantitative analysis method, including obtaining gear surface protection film original micrograph matrix and lens focal length, magnification, gear modulus, tooth width, optical axis inclination and other parameters;Extract the pixel and calculate the gradient amplitude matrix, remove the optical refraction artifact by threshold truncation and morphological processing, extract the pure protection film gray matrix;Based on the nonlinear mapping equation, the initial nominal physical thickness is deduced and the space thickness numerical array is established;Using the local curvature radius of the gear surface and the optical axis inclination, the dynamic curvature compensation factor is derived, and the real thickness coordinate array is generated;According to the logarithmic cumulative operation, the spatial distribution randomness index is calculated and the rating information is outputted.The present application realizes the accurate quantization and uniformity quantitative evaluation of the protection film thickness, and provides data support for process optimization and closed-loop control.

Description

Technical Field

[0001] This invention relates to the field of wind turbine gearbox protection technology, and in particular to an endoscopic image quantitative analysis method for the thickness and uniformity of gear protective film. Background Technology

[0002] Gear protective film is a protective coating applied to the surface of wind turbine gears to enhance wear resistance and corrosion resistance. Thickness characterizes the depth of coverage of the protective film on the gear surface and is a fundamental parameter for evaluating the coating's protective effect. Uniformity describes the degree of evenness in the distribution of the protective film, directly affecting the long-term operational reliability of the gears. Endoscopic images, obtained through endoscopic equipment, capture visual information of the gear's interior or concealed areas, providing intuitive evidence for non-invasive inspections. Quantitative analysis, based on endoscopic images, uses image processing technology to convert visual data into numerical results, thereby objectively assessing the thickness distribution and uniformity of the protective film.

[0003] Existing endoscopic image analysis technology suffers from the following technical challenges: In the in-situ, non-destructive testing of the domestically produced coal-based PAO and hydrogenated base oil-supporting protection system for megawatt-level wind turbines, the complex, non-linear geometric curvature of the gear teeth causes non-uniform reflection of light collected by the endoscope between the protective film surface and the metal substrate. Simultaneously, the residual coal-based base oil lubricating medium on the tooth surface possesses unique optical refractive properties, creating severe optical coherence interference with the micron-level protective film. This results in a high degree of coupling between the protective film boundary features and the tooth surface substrate texture in the image, making accurate feature stripping difficult using standard image segmentation algorithms. For example, when inspectors perform in-situ observation of the hydrogenated repair and reinforcement protective film at the edge of the meshing zone, the complex gear module geometry causes local light and shadow distortion. Combined with the refractive and masking effect of the residual oil film, this leads the system to misjudge microscopic scratches on the tooth surface substrate as fluctuations in the protective film thickness. This results in significant mapping deviations during the micron-level thickness quantification calculation, ultimately failing to meet the quantitative accuracy requirements for precise control of the protection system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an endoscopic image quantitative analysis method for the thickness and uniformity of gear protective films. This invention solves the technical problem that the complex tooth surface curvature geometry of heavy-duty wind turbine gearboxes and the optical refraction interference caused by residual lubricating media make it difficult to effectively separate the micron-level protective film from the metal substrate image features and cause the thickness measurement accuracy deviation to exceed ±1 micron.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: The endoscopic image quantitative analysis method for the thickness and uniformity of gear protective film provided by this invention includes: Step 1: Obtain the original microscopic image matrix, lens focal length, magnification, gear module, tooth width, optical axis tilt angle, and physical chord length of the observation field of view of the protective film on the tooth surface with residual lubricating medium inside the heavy-duty gearbox. Step 2: Extract the pixels from the original microscopic image matrix, calculate the gradient magnitude matrix of the pixels, truncate the gradient magnitude matrix based on a preset segmentation threshold to obtain a binarized mask, perform morphological processing of the binarized mask by first erosion and then dilation, remove the optical refraction artifacts generated by the lubricating residual medium, and extract the pure protective film grayscale matrix. Step 3: Extract the gray values ​​of each pixel in the gray matrix of the pure protective film, substitute the gray values, the lens focal length and the magnification into the nonlinear mapping equation, deduce the initial nominal physical thickness value of the corresponding pixel, and summarize all the initial nominal physical thickness values ​​to establish a spatial thickness value array. Step 4: Analyze the local radius of curvature of the target tooth surface at the center point of the observation field of view based on the gear module and the tooth width. Derive the dynamic curvature compensation factor based on the local radius of curvature and the optical axis tilt angle. Multiply the spatial thickness numerical array with the dynamic curvature compensation factor to generate a true thickness coordinate array. Step 5: Divide the real thickness coordinate array into multiple equidistant sub-frequency bands, count the probability set of pixels in the real thickness coordinate array falling into the multiple equidistant sub-frequency bands, calculate the spatial distribution randomness index based on logarithmic accumulation operation, extract the main diagonal elements of the preset compliance threshold matrix to form a one-dimensional judgment benchmark vector, and compare the spatial distribution randomness index with the various level thresholds in the one-dimensional judgment benchmark vector.

[0006] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention obtains the original microscopic image matrix of the gear surface protective film containing residual lubricating medium inside a heavy-duty gearbox, including: Receive the intensity signal reflected from the tooth surface; Adjust the polarization angle according to the tooth surface reflection intensity signal; Capture surface-reflected photons that penetrate the lubricating residual medium at the adjusted polarization angle; The surface-reflected photons are converted into a discrete pixel array, and the discrete pixel array is packaged into the original microscopic image matrix.

[0007] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention includes, before calculating the gradient magnitude matrix of the pixel points: Extract the pixel grayscale values ​​from the original microscopic image matrix; The pixel grayscale values ​​are convolved according to a preset two-dimensional discrete pixel kernel matrix to obtain convolved grayscale values; The gray-level histogram distribution of the convolutional gray values ​​is statistically analyzed, and the maximum inter-class variance is calculated using an adaptive algorithm to lock the segmentation threshold. The gradient magnitude matrix is ​​truncated according to the segmentation threshold.

[0008] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention analyzes the local radius of curvature of the target tooth surface at the center point of the observation field of view based on the gear module and the tooth width, and derives the dynamic curvature compensation factor based on the local radius of curvature and the optical axis tilt angle, including: Substitute the gear module and the tooth width values ​​into the involute equation to solve for the local radius of curvature; Extract the optical axis tilt angle and the physical chord length of the observation field of view; Based on the local radius of curvature, the optical axis tilt angle, and the physical chord length of the observation field, a trigonometric function compensation model is constructed, and the dynamic curvature compensation factor is derived.

[0009] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention includes, before dividing the true thickness coordinate array into multiple equidistant sub-bands: Traverse the real thickness coordinate array and calculate the global thickness arithmetic mean; The standard deviation and range parameters are calculated based on the arithmetic mean of the global thickness. The true thickness coordinate array is divided into multiple equidistant sub-bands according to the range parameter.

[0010] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention, after extracting the main diagonal elements of the preset compliance threshold matrix to form a one-dimensional judgment benchmark vector, and comparing the spatial distribution randomness index with each level threshold in the one-dimensional judgment benchmark vector, includes: Extract the spatial thickness values ​​from the actual thickness coordinate array; The spatial thickness value is mapped to a preset red-blue pseudo-color spectrum; Render the red-blue pseudo-color spectrum and output a thickness distribution heatmap; Based on the pseudo-color extreme value regions in the thickness distribution heatmap, locate the locally excessively thick, excessively thin, and exposed areas of the tooth surface protective film.

[0011] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention includes, before receiving the tooth surface reflection intensity signal: Obtain the base oil type and refractive index properties from external input; Match the target polarization angle range and target brightness range corresponding to the refractive index attribute in the preset optical refractive index calibration library; The target polarization angle range and the target brightness range are converted into driving level signals; The lighting signal is output according to the driving level signal.

[0012] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention, which calculates the spatial distribution randomness index based on logarithmic accumulation, includes: Extract a single probability value from the probability set of the pixels; Calculate the natural logarithm of the single probability value; multiply the single probability value by the natural logarithm to obtain the discrete information content; The discrete information quantities within the multiple equidistant sub-frequency bands are summed and the negative value is taken to obtain the regional distribution entropy, which is used as the spatial distribution randomness index.

[0013] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention, after extracting the main diagonal elements of the preset compliance threshold matrix to form a one-dimensional judgment benchmark vector, and comparing the spatial distribution randomness index with each level threshold in the one-dimensional judgment benchmark vector, includes: Extract the rating information and the arithmetic mean of the global thickness; The difference between the global thickness arithmetic mean, thickness standard deviation, and range parameter and the preset process target baseline is calculated, and the extreme value normalization processing is performed on the difference to eliminate the numerical magnitude difference caused by the dimensionless information entropy and the micrometer-scale length physical quantity. The normalized data is spliced ​​together to generate a multidimensional deviation vector. The multidimensional deviation vector is input into a preset rule model. The rule model constructs a multi-level decision tree that includes a thickness deviation threshold, a dispersion deviation threshold, and a range deviation threshold. The node conditions of the feature dimensions in the multidimensional deviation vector are compared sequentially, and logical forward deduction is performed to obtain the deduction conclusion. Based on the derivation conclusion, output the preparation parameter correction command and send the preparation parameter correction command.

[0014] Furthermore, the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film described in this invention, based on the derived conclusion, outputs a preparation parameter correction command and sends the preparation parameter correction command, including: Thickness parameters, discrete parameters, and range parameters are extracted from the multidimensional deviation vector; when the thickness parameter exceeds the preset over-limit judgment condition, an instruction to reduce the additive concentration is generated; When the discrete parameter meets the preset uniformity failure judgment condition, an instruction to adjust the coating pressure is generated. When the range parameter meets the preset fluid agglomeration phenomenon judgment condition, an instruction is generated to increase the nozzle atomization pressure and increase the curing temperature. The generated preparation parameter correction instructions are encapsulated into a digital correction feedback message and sent.

[0015] Beneficial effects of this invention: This invention effectively suppresses specular reflection and coherent interference caused by residual lubricating media in heavy-duty gearboxes by constructing a deep collaborative mechanism between a high-precision polarized endoscope acquisition unit and a protective film image preprocessing unit, coupled with a dynamic polarization angle adjustment function. This allows for precise decoupling of the image features of the micron-level protective film and the metal substrate, significantly improving the quality of the raw data for subsequent quantitative analysis. In the thickness quantification calculation stage, by introducing the local radius of curvature obtained from the numerical analysis of gear module and tooth width, as well as the generated dynamic curvature compensation factor, the system corrects the projection scaling distortion caused by nonlinear tooth surface geometry. This ensures that the thickness quantification calculation results can accurately map the actual physical depth of the protective film on complex curved surfaces, effectively solving the problem of large measurement accuracy deviations in conventional testing. By constructing a multidimensional evaluation system using a real thickness coordinate array, comprising the global thickness arithmetic mean, standard deviation, range parameters, and spatial distribution randomness index obtained based on logarithmic accumulation, a comprehensive quantitative control over the coating distribution state from macroscopic benchmarks to microscopic disorder is achieved. Combined with the thermal mapping function of red-blue pseudo-color spectrum rendering, it can accurately locate local overly thick, overly thin, or exposed areas, greatly improving the objectivity of uniformity assessment. Finally, through the generated multidimensional deviation vector including thickness parameters, discrete parameters, and range parameters, the expert rule model is driven to output closed-loop preparation parameter correction instructions for additive concentration, coating pressure, and curing temperature. This enables feedback guidance of the detection data on the production process, significantly enhancing the long-term operation and maintenance reliability of the wind turbine gearbox protection system. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the endoscopic image quantitative analysis method for the gear protective film thickness and uniformity of the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1 The endoscopic image quantitative analysis method for the thickness and uniformity of gear protective film provided by the present invention includes: Step 1: Obtain the original microscopic image matrix, lens focal length, magnification, gear module, tooth width, optical axis tilt angle, and physical chord length of the observation field of view of the protective film on the tooth surface with residual lubricating medium inside the heavy-duty gearbox. Step 2: Extract the pixels from the original microscopic image matrix, calculate the gradient magnitude matrix of the pixels, truncate the gradient magnitude matrix based on a preset segmentation threshold to obtain a binarized mask, perform morphological processing of the binarized mask by first erosion and then dilation, remove the optical refraction artifacts generated by the lubricating residual medium, and extract the pure protective film grayscale matrix. Step 3: Extract the gray values ​​of each pixel in the gray matrix of the pure protective film, substitute the gray values, the lens focal length and the magnification into the nonlinear mapping equation, deduce the initial nominal physical thickness value of the corresponding pixel, and summarize all the initial nominal physical thickness values ​​to establish a spatial thickness value array. Step 4: Analyze the local radius of curvature of the target tooth surface at the center point of the observation field of view based on the gear module and the tooth width. Derive the dynamic curvature compensation factor based on the local radius of curvature and the optical axis tilt angle. Multiply the spatial thickness numerical array with the dynamic curvature compensation factor to generate a true thickness coordinate array. Step 5: Divide the real thickness coordinate array into multiple equidistant sub-frequency bands, count the probability set of pixels in the real thickness coordinate array falling into the multiple equidistant sub-frequency bands, calculate the spatial distribution randomness index based on logarithmic accumulation operation, extract the main diagonal elements of the preset compliance threshold matrix to form a one-dimensional judgment benchmark vector, and compare the spatial distribution randomness index with the various level thresholds in the one-dimensional judgment benchmark vector.

[0020] In the in-situ, non-disassembly inspection of heavy-duty gearboxes in megawatt-class wind turbines, the tooth surfaces often have residual lubricating media such as coal-based PAO or hydrogenated base oil. To investigate the microscopic state of the coating, the original microscopic image matrix of the protective film on the tooth surface with residual lubricating media inside the heavy-duty gearbox, along with lens focal length, magnification, gear module, tooth width, optical axis tilt angle, and physical chord length of the observation field of view were obtained. The lens focal length and magnification were obtained using a polarized endoscope, with the probe penetrating deep into the gearbox to capture surface-reflected photons through the residual lubricating media. These surface-reflected photons underwent photoelectric conversion to form a discrete pixel array, which was then packaged to construct the original microscopic image matrix. The gear module and tooth width values ​​were extracted from the gearbox design drawings, while the optical axis tilt angle and physical chord length of the observation field of view were derived from the spatial positioning parameters of the probe in a specific orientation.

[0021] The original microscopic image matrix contains optical refraction artifacts from residual lubricating medium. Pixels in the original microscopic image matrix are extracted, and the gradient magnitude matrix of each pixel is calculated. The gradient magnitude matrix reflects the degree of gray-level abrupt change between adjacent pixels in the orthogonal direction. The gradient magnitude matrix is ​​truncated to obtain a binarized mask, converting the continuous gray-level space into a binary space that is either black or white. Since the binarized mask contains isolated noise points caused by oil film reflection, a morphological process of erosion followed by dilation is performed on the binarized mask. The erosion operation cuts off the pixel connectivity regions caused by the residual lubricating medium, and the dilation operation restores the physical boundaries of the real protective film, removing the optical refraction artifacts caused by the residual lubricating medium to extract the pure protective film gray-level matrix.

[0022] The grayscale matrix of the pure protective film only includes the optical features of a two-dimensional plane. The grayscale values ​​of each pixel within the grayscale matrix are extracted, and the corresponding pixel grayscale values, lens focal length, and magnification are substituted into a nonlinear mapping equation. This nonlinear mapping equation is based on the attenuation law of photons penetrating the protective film medium. The lens focal length and magnification are used as correction variables for the optical attenuation coefficient in the calculation to deduce the initial nominal physical thickness value of the corresponding pixel. A spatial thickness value array is established by collecting data from a toothed surface region covering a macroscopic field of view and summarizing all initial nominal physical thickness values.

[0023] The spatial thickness numerical array has not yet incorporated the complex nonlinear curvature geometry of the tooth surface in heavy-duty gearboxes, inevitably resulting in quantization bias caused by projection perspective scaling effects. The local radius of curvature of the target tooth surface at the center point of the observation field is numerically analyzed based on the gear module and tooth width. The local radius of curvature differs at different meshing positions of the involute tooth surface; the dynamic curvature compensation factor is derived by combining the local radius of curvature with the optical axis tilt angle. The dynamic curvature compensation factor reflects the degree of optical projection distortion caused by the curved tooth surface under a specific observation angle. Multiplying the spatial thickness numerical array by the dynamic curvature compensation factor generates the true thickness coordinate array.

[0024] The true thickness coordinate array characterizes the actual coverage depth of the micron-level protective film on the toothed surface, and is divided into multiple equidistant sub-frequency bands. These equidistant sub-frequency bands constitute statistical intervals for the thickness range, and the probability sets of pixels falling within these sub-frequency bands are statistically analyzed. These pixel probability sets reflect the clustering state of the film thickness in the spatial frequency domain. A spatial distribution randomness index is obtained by processing the pixel probability sets using logarithmic summation. This index quantifies the degree of disorder in the coating distribution, and rating information is output by comparing the index with a preset compliance threshold matrix.

[0025] Before the endoscope probe penetrates deep into the heavy-duty gearbox, the system receives external input regarding the base oil type and refractive index properties. Different lubricant formulations produce specific refractions to the probe beam. The system's preset optical refractive index calibration library stores target polarization angle ranges and target brightness ranges corresponding to various refractive index properties. After the main control program completes the matching process in the calibration library, it converts the target polarization angle range and target brightness range into drive level signals. The underlying drive module outputs an illumination signal based on the drive level signal, activating the light source assembly and polarization filter mechanism at the probe's front end. Under the illumination signal, the probe projects a probe beam, and the probe front end then receives the tooth surface reflection intensity signal. The closed-loop control circuit inside the system adjusts the polarization angle based on the tooth surface reflection intensity signal to avoid high-intensity specular reflection points formed on the surface of the lubricant. At the adjusted polarization angle, the optical sensor captures surface-reflected photons that penetrate the residual lubricant medium. The photosensitive element array converts the surface-reflected photons into a discrete pixel array, and the processor ultimately packages the discrete pixel array into a raw microscopic image matrix.

[0026] To eliminate stray light spots hidden in the original microscopic image matrix, the processor extracts the pixel grayscale values ​​from the original microscopic image matrix. The digital field composed of pixel grayscale values ​​is often accompanied by high-frequency abrupt noise. The processor calls a preset two-dimensional discrete pixel kernel matrix to perform convolution operations on the pixel grayscale values. A neighborhood-weighted smoothing process in the spatial domain filters out isolated bright spots caused by scattering from the lubricating residual medium, resulting in smoothly transitioned convolutional grayscale values. The algorithm module performs frequency-level analysis on pixels across the entire image and statistically analyzes the grayscale histogram distribution of the convolutional grayscale values. The grayscale histogram distribution visually presents a bimodal characteristic of the target coating and the metal background. The system uses an adaptive algorithm to calculate the maximum inter-class variance to lock the segmentation threshold. The segmentation threshold located at the trough of the bimodal peak represents the critical brightness of the medium boundary. The system performs a truncation gradient magnitude matrix operation based on the segmentation threshold.

[0027] After extracting features from the two-dimensional plane of the image, the system addresses the projection distortion caused by the three-dimensional curvature of the tooth surface. The processor substitutes the gear module and tooth width values ​​into the involute equation. The involute equation reflects the geometric characteristics of the tooth profile at different unfolding angles, and the system solves for the local radius of curvature along a given mathematical path. The spatial mapping relationship formed by the probe posture also affects thickness quantization; the system extracts the optical axis tilt angle and the physical chord length of the observation field. The local radius of curvature reflects the degree of surface curvature at the current observation point, and the optical axis tilt angle and the physical chord length of the observation field together define the relative projection posture of the endoscope lens and the curved surface. The processor constructs a trigonometric function compensation model based on the local radius of curvature, the optical axis tilt angle, and the physical chord length of the observation field. The trigonometric function compensation model flattens the curved surface features into a planar mapping, deriving a dynamic curvature compensation factor to correct projection distortion.

[0028] After obtaining the 3D data points corrected for curvature, the statistical analysis module traverses the real thickness coordinate array and calculates the global thickness arithmetic mean. The global thickness arithmetic mean reflects the macroscopic coverage level of the coating in the current observation area. The system further calculates the standard deviation and range parameters based on the global thickness arithmetic mean. The standard deviation characterizes the overall dispersion of numerical fluctuations, while the range parameter anchors the maximum span between thickness extrema. The system divides the real thickness coordinate array into multiple equidistant sub-bands based on the range parameter, segmenting the continuous thickness distribution into discrete statistical intervals. To transform the abstract data matrix into an image that can be interpreted by maintenance personnel, the rendering engine extracts the spatial thickness values ​​from the real thickness coordinate array. Variations in spatial thickness values ​​correspond to specific color codes, and the system maps these values ​​to a preset red-blue pseudo-color spectrum. The graphics processor renders the red-blue pseudo-color spectrum and outputs a thickness distribution heatmap. Different shades of pseudo-color patches correspond to different physical thicknesses. Based on the pseudo-color extrema regions in the thickness distribution heatmap, the system locates locally excessively thick, excessively thin, and exposed areas of the protective film on the tooth surface.

[0029] To quantify coating quality from the perspective of spatial disorder, the system delves into the probabilistic characteristics of pixel distribution, extracting single probability values ​​from the pixel probability set. Each single probability value reflects the density of residing pixels within a specific thickness range. The processor calculates the natural logarithm of each single probability value. The natural logarithm operation amplifies the information content differences caused by small probability variations. The system multiplies the single probability value by the natural logarithm to obtain discrete information content. Discrete information content characterizes the contribution weight of a single frequency band to the overall disorder level. The accumulator sums the discrete information content across multiple equidistant sub-frequency bands and takes the negative value. This summation and negative taking process integrates the spatial distribution state of all frequency bands to obtain the regional distribution entropy, which the system uses as an indicator of spatial distribution randomness.

[0030] The key to the quantitative assessment's transition to practical engineering applications lies in the closed-loop feedback mechanism. The processor extracts rating information and the global thickness arithmetic mean. The rating information and the thickness mean represent the current actual technological level. The system calculates the difference between the rating information, the global thickness arithmetic mean, and a preset technological target baseline. The direction and magnitude of the difference are aggregated to generate a multi-dimensional deviation vector. This multi-dimensional deviation vector is input into a preset rule model for logical forward derivation, yielding the derivation conclusion. The preset rule model employs a multi-level decision tree architecture. Using the multi-dimensional deviation vector as input, the processor sequentially extracts the thickness parameter, discrete parameter, and range parameter from the multi-dimensional deviation vector. The first-level node of the multi-level decision tree sets a thickness deviation threshold, and the processor determines whether the thickness parameter exceeds the preset threshold. The second-level node sets a dispersion deviation threshold, and the processor determines whether the discrete parameter exceeds the preset threshold. The third-level node sets a range deviation threshold, and the processor determines whether the range parameter exceeds the preset threshold. The processor sequentially compares node conditions according to the preset multi-level decision tree conditional branch paths, executes logical forward derivation, and outputs the derivation conclusion of the corresponding preparation parameter correction instruction. During the offline training phase, the deep neural network uses the mean squared error loss function as the convergence criterion. When the loss function value remains below 0.001 for 50 consecutive iterations, the processor stops training and locks the model parameters.

[0031] The multidimensional deviation vector specifically includes thickness parameters, discrete parameters, and range parameters. When the thickness parameter exceeds the preset over-limit judgment condition, the system determines that the solid content of the coating liquid is too high and generates an instruction to reduce the additive concentration. When the discrete parameter meets the preset uniformity non-compliance judgment condition, the system determines that the coating force distribution is uneven and generates an instruction to adjust the coating pressure. When the range parameter meets the preset fluid agglomeration phenomenon judgment condition, it indicates that the surface tension of the high-viscosity medium is unbalanced, and the system generates instructions to increase the nozzle atomization pressure and raise the curing temperature. The system packages and integrates the generated preparation parameter correction instructions, encapsulates them into a digital correction feedback message, and sends it.

[0032] The internal operating environment of the heavy-duty gearbox of a megawatt-class wind turbine is extremely complex. The composition of the residual lubricating medium that adheres to the gear surface over a long period directly determines the fundamental boundary of optical refraction characteristics. Specifically, the residual lubricating medium encompasses fluid materials of different viscosity grades, including coal-based PAO synthetic oil and hydrotreated base oil, which are widely used in the process of domestic production. The hydrocarbon molecular chain structures contained in different base oil categories differ significantly, causing the detection beam penetrating the residual lubricating medium to undergo deflection and scattering at specific angles. Externally input base oil category and refractive index attribute data clearly define the optical physical identity of the residual lubricating medium. The refractive index attribute data directly corresponds to the target polarization angle range and target brightness range fixed in the calibration library. The underlying hardware unit receives the converted drive level signal and outputs an illumination signal adapted to the current fluid characteristics.

[0033] The probe penetrates deep into the narrow inter-tooth meshing area to capture surface-reflected photons. A photoelectric conversion process maps these photons into a discrete pixel array. This discrete pixel array, along with spatial physical properties, constructs the original microscopic image matrix. This matrix not only includes coordinate information recording the object's geometric contours but also carries optical characteristic data characterizing the refractive index and transmittance of the lubricating residual medium. The lens focal length and magnification define the physical scaling ratio within the field of view, while the gear module and tooth width depict the macroscopic three-dimensional contour of the involute tooth surface. The positioning module acquires probe pose data in real time, outputting the optical axis tilt angle and the physical chord length of the observed field of view, calibrating the relative spatial relationship between the endoscope lens and the measured interface.

[0034] The image processing workflow requires extracting the features of the protective film adhered to the metal substrate. Pixel grayscale values ​​reflect the absorption and reflection intensity of light by different media. A two-dimensional discrete pixel kernel matrix performs spatial domain convolution operations on the pixel grayscale values ​​to smoothly filter out local high-frequency noise caused by uneven distribution of lubricating residual media. The partial derivatives of adjacent pixels in orthogonal directions converge to form a gradient magnitude matrix. The gradient magnitude matrix indicates the steepness of grayscale transitions at the interface of the media in the microscopic field of view. An adaptive algorithm locks the segmentation threshold and truncates the gradient magnitude matrix to obtain a binarized mask. Morphological processing, including erosion followed by dilation, cuts off isolated connected regions caused by residual oil droplets, removes optical refraction artifacts caused by impurities and substrate texture, and extracts the pure protective film grayscale matrix.

[0035] The average grayscale value of pixels within the grayscale matrix of the pure protective film reflects the energy attenuation caused by the probe beam penetrating the coating medium. A nonlinear mapping equation based on the photometric attenuation law transforms the average grayscale value into an initial nominal physical thickness value. The spatial thickness array established by summing the initial nominal physical thickness values ​​lacks correction for the three-dimensional morphology of the tooth surface. The involute equation, combined with the gear module and tooth width values, is used to solve for the local radius of curvature of the target tooth surface at the center point of the observation field. The dynamic curvature compensation factor derived from the local radius of curvature, optical axis tilt angle, and physical chord length of the observation field quantifies the perspective distortion caused by the projection of the curved surface onto the two-dimensional sensor plane. The product fusion operation maps the three-dimensional coordinates of the spatial thickness array according to the weights of the dynamic curvature compensation factor, generating a true thickness coordinate array that matches the curvature of the tooth surface.

[0036] Assessing coating quality often requires transcending the limitations of single-point thickness to explore the overall uniformity. The arithmetic mean of the global thickness outlines the macroscopic coverage baseline of the coating, the standard deviation quantifies the overall dispersion of numerical fluctuations, and the range parameter defines the maximum range of thickness extremes. The true thickness coordinate array is divided into multiple equidistant sub-bands based on the range parameter. The probability set of pixels falling within these equidistant sub-bands reflects the clustering of the film thickness in the spatial frequency domain. The discrete information is calculated by extracting a single probability value from the pixel probability set and multiplying it by the natural logarithm of that single probability value. The microprocessor sums the discrete information within the multiple equidistant sub-bands and takes the negative value to obtain the regional distribution entropy. This regional distribution entropy is then used as an indicator of spatial distribution randomness in the rating calculation.

[0037] The ultimate goal of defect identification is to guide parameter iteration in the manufacturing process. The rating information and the difference between the global thickness arithmetic mean and the preset process target baseline are aggregated into a multi-dimensional deviation vector. This vector includes thickness parameters reflecting the extent to which thickness exceeds limits, discrete parameters characterizing uneven thickness distribution, and range parameters reflecting the degree of local agglomeration. A preset rule model derives and outputs preparation parameter correction instructions based on the multi-dimensional deviation vector. These instructions package control data from various field devices, including instructions to reduce additive concentration to adjust the solid content of the coating liquid, instructions to adjust coating pressure to improve coating strength distribution, and instructions to increase nozzle atomization pressure and raise curing temperature to break the surface tension of high-viscosity media.

[0038] The internal operating environment of heavy-duty gearboxes in megawatt-class wind turbines is extremely complex, and the composition of residual lubricating media adhering to the tooth surfaces over long periods directly determines the fundamental boundaries of optical refraction characteristics. To address the technical challenges posed by the complex tooth surface curvature geometry and optical refraction interference caused by residual lubricating media in heavy-duty gearboxes, the system needs to acquire the original microscopic image matrix, lens focal length, magnification, gear module, tooth width, optical axis tilt angle, and physical chord length of the observation field of view of the protective film containing residual lubricating media on the tooth surfaces within the heavy-duty gearbox. To accommodate the refractive characteristics of different media, the system acquires the externally input base oil type and refractive index attribute. The base oil type and refractive index attribute clearly define the optical physical identity of the residual lubricating media. The system matches the target polarization angle range and target brightness range corresponding to the refractive index attribute in a pre-set optical refractive index calibration library. The main control program converts the target polarization angle range and target brightness range into drive level signals. The underlying drive module outputs an illumination signal based on the drive level signal to trigger the start-up of the light source assembly at the probe front end. Under the illumination signal, the probe projects a detection beam, and the detection front end then receives the tooth surface reflection intensity signal. The closed-loop control circuit adjusts the polarization angle based on the tooth surface reflection intensity signal to avoid strong specular reflection points. The optical sensor captures surface-reflected photons that penetrate the lubricating residual medium at the adjusted polarization angle. The photosensitive element array converts the surface-reflected photons into a discrete pixel array, and the processor packages the discrete pixel array into a raw microscopic image matrix. The lens focal length and magnification define the physical scaling ratio within the field of view, the gear module and tooth width depict the macroscopic three-dimensional profile of the tooth surface, and the optical axis tilt angle and the physical chord length of the observation field of view calibrate the relative spatial relationship between the lens and the measured interface.

[0039] The original microscopic image matrix contains optical refraction artifacts from residual lubricating media. The processor needs to peel off the protective film features adhering to the metal substrate. The processor extracts the pixel grayscale values ​​from the original microscopic image matrix. The digital field composed of pixel grayscale values ​​is often accompanied by high-frequency abrupt noise. The processor calls a preset two-dimensional discrete pixel kernel matrix to perform convolution operations on the pixel grayscale values. The neighborhood weighted smoothing process in the spatial domain filters out isolated bright spots caused by scattering from the residual lubricating media, resulting in smoothly transitioned convolutional grayscale values. The algorithm module statistically analyzes the grayscale histogram distribution of the convolutional grayscale values. The system calculates the maximum inter-class variance based on an adaptive algorithm to lock the segmentation threshold. The segmentation threshold located at a bimodal peak and trough represents the critical brightness of the medium boundary. The processor simultaneously extracts pixels from the original microscopic image matrix and calculates the gradient magnitude matrix of the pixels. The gradient magnitude matrix indicates the steepness of the grayscale transition at the medium interface in the microscopic field of view. The system performs a truncation operation on the gradient magnitude matrix based on the segmentation threshold to obtain a binarized mask. The binarization mask converts a continuous grayscale space into a binary space containing only discrete extremum pixels. A morphological process of erosion followed by dilation is then performed on the binarization mask. The erosion operation severs the pixel connectivity regions caused by the residual lubricating medium, while the dilation operation restores the physical boundaries of the true protective film. Optical refraction artifacts caused by the residual lubricating medium are removed to extract the pure protective film grayscale matrix.

[0040] The pure protective film's grayscale matrix only includes the optical features of a two-dimensional plane. The processor extracts the pixel grayscale values ​​of each pixel within the pure protective film's grayscale matrix. The nonlinear mapping equation is based on the attenuation law of photons penetrating the protective film medium. The processor substitutes each pixel grayscale value, lens focal length, and magnification into the nonlinear mapping equation. The lens focal length and magnification are used as correction variables for the optical attenuation coefficient in the calculation to deduce the initial nominal physical thickness value of the corresponding pixel. The processor collects data covering the tooth surface region of the macroscopic field of view and establishes a spatial thickness value array by summing all the initial nominal physical thickness values. The spatial thickness value array lacks correction for the three-dimensional morphology of the tooth surface and inevitably has quantization deviations caused by projection perspective scaling effects. The processor substitutes the gear module and tooth width values ​​into the involute equation. The involute equation reflects the geometric law of the tooth profile under different unfolding angles. The processor solves for the local radius of curvature of the target tooth surface at the center point of the observation field of view. The spatial mapping relationship formed by the probe attitude also affects the thickness quantization. The system extracts the optical axis tilt angle and the physical chord length of the observation field of view. A trigonometric function compensation model is constructed based on the local radius of curvature, optical axis tilt angle, and physical chord length of the observation field. This model flattens the curved surface features into a planar mapping, deriving the dynamic curvature compensation factor. The spatial thickness numerical array is then multiplied by the dynamic curvature compensation factor to generate the true thickness coordinate array.

[0041] The true thickness coordinate array characterizes the actual coverage depth of the micron-level protective film on the tooth surface. Assessing coating quality requires investigating the global uniformity. The statistical analysis module traverses the true thickness coordinate array, calculating the global arithmetic mean of the thickness. The system calculates the standard deviation and range parameters based on the global thickness arithmetic mean. The standard deviation characterizes the overall dispersion of numerical fluctuations, while the range parameter anchors the maximum span between thickness extremes. The system divides the true thickness coordinate array into multiple equidistant sub-bands based on the range parameter. These equidistant sub-bands divide the continuous thickness distribution into discrete statistical intervals. The probability set of pixels falling within these equidistant sub-bands is statistically analyzed. This pixel probability set reflects the clustering state of the film thickness in the spatial frequency domain. Individual probability values ​​are extracted from the pixel probability set. The processor calculates the natural logarithm of each individual probability value. Multiplying the individual probability value by the natural logarithm yields the discrete information. The accumulator sums the discrete information within the multiple equidistant sub-bands and takes the negative value to obtain the regional distribution entropy. The system uses the regional distribution entropy as an indicator of spatial distribution randomness. The rating information is output by comparing the spatial distribution randomness index with the preset compliance threshold matrix.

[0042] The abstract data matrix needs to be transformed into an image that can be interpreted by maintenance personnel. The rendering engine extracts the spatial thickness values ​​from the real thickness coordinate array. Variations in spatial thickness values ​​correspond to specific color codes, and the system maps these values ​​to a preset red-blue pseudo-color spectrum. The graphics processor renders the red-blue pseudo-color spectrum and outputs a thickness distribution heatmap. Based on the pseudo-color extreme regions in the thickness distribution heatmap, the system locates locally excessively thick, excessively thin, and exposed areas of the protective film on the tooth surface. The processor extracts rating information and the global thickness arithmetic mean. The rating information and the global thickness arithmetic mean represent the current actual process level. The difference between the rating information, the global thickness arithmetic mean, and the preset process target baseline is calculated, generating a multi-dimensional deviation vector. This multi-dimensional deviation vector is input into a preset rule model for logical forward derivation, yielding a derivation conclusion. Based on the derivation conclusion, a preparation parameter correction instruction is output. Thickness parameters, discrete parameters, and range parameters are extracted from the multi-dimensional deviation vector. When the thickness parameter exceeds the preset over-limit judgment condition, the system determines that the solid content of the coating liquid is too high and generates an instruction to reduce the additive concentration. When the discrete parameters meet the preset uniformity non-compliance criteria, the system identifies uneven coating force distribution and generates an instruction to adjust the coating pressure. When the range parameter meets the preset fluid agglomeration phenomenon criteria, it indicates an imbalance in the surface tension of the high-viscosity medium, generating instructions to increase the nozzle atomization pressure and raise the curing temperature. The generated preparation parameter correction instructions are encapsulated into a digital correction feedback message and sent.

[0043] The pixel grayscale values ​​are extracted from the original microscopic image matrix and fed into the digital processing channel. These pixel grayscale values ​​constitute the basic data field in two-dimensional space. A convolution operation is performed on the pixel grayscale values ​​according to a preset two-dimensional discrete pixel kernel matrix. The specific mathematical expression equation for the spatial domain neighborhood-weighted convolution operation is as follows:

[0044] in, This represents the grayscale value obtained from the convolution operation. Represents the lateral spatial coordinate position of the microscopic image. Represents the vertical spatial coordinate position of the microscopic image. Represents the discrete accumulation operator. The horizontal index parameter represents the preset two-dimensional discrete pixel kernel matrix. The vertical index parameter represents the preset two-dimensional discrete pixel kernel matrix. This represents the grayscale value of a specific pixel in the extracted original microscopic image matrix. This represents the weight coefficients within a predefined two-dimensional discrete pixel kernel matrix. After smoothing out local speckle patterns using the discrete pixel kernel matrix, a smooth convolutional grayscale value is obtained. The spatial partial derivatives of adjacent pixels in orthogonal directions are calculated based on the convolutional grayscale values. These spatial partial derivatives converge to construct the gradient magnitude matrix. The analytical model for the gradient magnitude matrix is ​​as follows:

[0045] in, This represents the magnitude elements in the calculated gradient magnitude matrix. Represents the lateral spatial coordinate position of the microscopic image. Represents the vertical spatial coordinate position of the microscopic image. Represents the square root operator. This represents the partial derivative of the convolutional gray value with respect to the horizontal dimension. This represents the partial derivative of the convolutional gray value in the vertical dimension.

[0046] The processor statistically analyzes the distribution pattern of the gray-level histogram of convolutional gray values. Based on the gray-level histogram distribution contour and an adaptive algorithm, the maximum inter-class variance is calculated. The core algorithm model for finding the maximum inter-class variance is as follows:

[0047] in, This represents the maximum inter-class variance calculated dynamically. A probability parameter representing the proportion of pixel regions on the background metal substrate. The probability parameter representing the proportion of pixel regions in the target medium protective film. The average brightness grayscale of the pixel area representing the background metal substrate. This represents the average brightness grayscale of the pixel region of the target medium protective film. The processor establishes a loop logic that iterates from grayscale values ​​0 to 255, assuming a segmentation threshold. The values ​​are incremented sequentially from 0 to 255 and fed into the core algorithm model, comparing the inter-class variance values ​​output in each iteration. When the iteration ends, the processor locks the assumed segmentation threshold corresponding to the maximum inter-class variance value. The assumed segmentation threshold will be locked. The optimal segmentation threshold is used. A binary mask is obtained by truncating the gradient magnitude matrix based on the segmentation threshold. A morphological process of erosion followed by dilation is performed on the binary mask. This morphological process cuts off isolated connected regions caused by residual oil droplets and removes optical refraction artifacts caused by residual lubricating medium, thereby extracting the pure protective film grayscale matrix. The grayscale mean of the pixels within the pure protective film grayscale matrix is ​​extracted. The grayscale mean, lens focal length, and magnification are substituted into the nonlinear mapping equation. The nonlinear mapping equation is based on the attenuation law of photons penetrating the protective film medium. The specific structure of the nonlinear mapping equation is as follows:

[0048] in, This represents the initial nominal physical thickness value derived from the calculation. This represents the system's preset optical energy transmission attenuation coefficient. This represents the previously extracted lens focal length. This represents the magnification factor extracted above. Represents the natural logarithm operator. This represents the average grayscale value of the pixels within the grayscale matrix of the pure protective film. The system represents the pre-calibrated reference grayscale mean of the metal substrate. The initial nominal physical thickness value is derived through a nonlinear mapping equation. A spatial thickness value array is established by summing all initial nominal physical thickness values ​​covering the entire map area.

[0049] Spatial thickness numerical arrays are thickness projections onto a two-dimensional plane, necessitating the introduction of a dynamic curvature compensation factor in the physical three-dimensional dimension to eliminate distortion. The local radius of curvature of the target tooth surface at the center point of the observation field is determined numerically based on the gear module and tooth width. The involute surface equation is then introduced to solve for the local radius of curvature. The geometric transformation form of the involute surface equation is as follows:

[0050] in, The radius of curvature of the target tooth surface at the center point of the observation field of view, as determined by analysis, is represented. This represents the gear module obtained in the previous step. This represents the gear tooth count parameter set in conjunction with the gear module. Represents the sine trigonometric function operator. This represents the pressure angle parameter in gear design specifications. This represents the tooth width value obtained previously. Represents the tangent trigonometric function operator. This represents the helix angle parameter in gear design specifications. The dynamic curvature compensation factor is derived based on the local radius of curvature and the optical axis tilt angle. A trigonometric function compensation model is constructed by integrating the physical chord length of the observed field of view. The specific correction equations of the trigonometric function compensation model are as follows:

[0051] in, This represents the derived dynamic curvature compensation factor. This represents the physical chord length of the previously acquired observation field of view. The radius of curvature of the target tooth surface at the center point of the observation field of view, as determined by analysis, is represented. This represents the cosine trigonometric function operator. This represents the optical axis tilt angle obtained previously. The spatial thickness numerical array is multiplied by the dynamic curvature compensation factor. The expression for the spatial lattice mapping product operation is as follows:

[0052] in, This represents the element thickness value at a specific coordinate point in the generated true thickness coordinate array. The horizontal dimension index representing the data array. The vertical dimension index representing the data array. This represents the initial thickness element at the corresponding coordinate point in the completed spatial thickness numerical array. This represents the derived dynamic curvature compensation factor. The product operation corrects the pixel stretching distortion caused by surface projection and generates a true thickness coordinate array.

[0053] The obtained true thickness coordinate array is traversed to calculate the global arithmetic mean of thickness. The statistical summation formula for the global thickness arithmetic mean is as follows:

[0054] in, This represents the calculated arithmetic mean of the global thickness. This represents the total number of valid pixel data points included in the true thickness coordinate array. This represents the summation operator. The sequence index order representing the valid pixel data points. This represents a single thickness value corresponding to a specific sequence index in the true thickness coordinate array. The standard deviation and range parameters are calculated based on the global thickness arithmetic mean. The standard deviation characterizes the dispersion of the thickness data from the arithmetic mean. The formula for calculating the standard deviation is as follows:

[0055] in, This represents the obtained standard deviation of the thickness. Represents the square root operator. This represents the total number of valid pixel data points included in the true thickness coordinate array. This represents the summation operator. The sequence index order representing the valid pixel data points. This represents a single thickness value corresponding to a specific sequence index in the true thickness coordinate array. This represents the arithmetic mean of the calculated global thickness. The range parameter defines the limiting boundary range that the actual coating thickness crosses. The formula for calculating the range parameter is as follows:

[0056] in, This represents the range parameter obtained through calculation. This represents the maximum extreme value of thickness detected in the true thickness coordinate array. This represents the minimum thickness extremum detected in the true thickness coordinate array. The true thickness coordinate array is divided into multiple equidistant sub-frequency bands based on the range parameter. The probability set of pixels in the true thickness coordinate array falling within these equidistant sub-frequency bands is calculated. Individual probability values ​​are extracted from the pixel probability set, and the natural logarithm of each individual probability value is calculated. The discrete information content is obtained by multiplying the individual probability value by the natural logarithm. The discrete information content across multiple equidistant sub-frequency bands is summed and the result is negative. The information theory conversion formula for obtaining the regional distribution entropy is as follows:

[0057] in, This represents the calculated regional distribution entropy. This represents the total number of the multiple equidistant sub-bands that have been defined. This represents the summation operator. The ascending sequence number representing the equidistant sub-bands. This represents a single probability value corresponding to a specific sub-frequency band within the probability set of pixels. The logic for obtaining the single probability value is as follows: the processor statistically analyzes the data in the actual thickness coordinate array that falls within the first... Number of pixels in each equidistant sub-band And obtain the total number of valid pixels contained in the true thickness coordinate array. Through calculation and The ratio determines the single probability value. Regional distribution entropy The calculation process is strictly limited to 0 to 1. Within the range of values, where The closer the values ​​are The more the protective film is distributed in space on the tooth surface, the more it tends to be absolutely uniform. The closer the value is to 0, the more severe the local aggregation phenomenon of the protective film is. The calculation model of the single probability value is the number of pixels falling within a specific equidistant sub-frequency band divided by the total number of effective pixel data points. This represents the natural logarithm operator. The regional distribution entropy is used as a spatial distribution randomness indicator reflecting the degree of disorder in the coating. The spatial distribution randomness indicator is compared with a preset compliance threshold matrix to output rating information. The rating information and the global thickness arithmetic mean are extracted and used in closed-loop control simulation. The difference between the rating information, the global thickness arithmetic mean, and the preset process target baseline is calculated. Various detection difference data are aggregated and spliced ​​to generate a multidimensional deviation vector. Since the rating information is a dimensionless probability entropy feature, while the global thickness arithmetic mean is a micrometer-level physical length feature, the processor must perform scale alignment processing of the feature space before generating the multidimensional deviation vector. The processor obtains the historical maximum and minimum values ​​of the rating information and the global thickness arithmetic mean from the past 1000 samples, and uses the extreme value normalization formula to linearly map the rating information and the global thickness arithmetic mean to the [0,1] interval. Normalization can eliminate the differences in numerical magnitude caused by different physical dimensions, ensuring that the absolute fluctuation of the thickness value will not overwhelm the probability distribution characteristics when the deep neural network calculates the weight gradient through backpropagation, thereby guaranteeing the global convergence performance and inference accuracy of the algorithm.

[0058] The multidimensional deviation vector forms a quantitative link between the detection and production ends. The construction matrix form of the multidimensional deviation vector is as follows:

[0059] in, This represents the multidimensional bias vector of the generated output. This represents the extracted global thickness arithmetic mean. This represents the thickness reference target parameter specified in the preset process target baseline. This represents the calculated standard deviation of the thickness. This represents the standard deviation of the baseline target parameter specified in the preset process target baseline. This represents the range parameter obtained through calculation. The multidimensional deviation vector represents the range baseline target parameter specified in the preset process target baseline. The multidimensional deviation vector internally maps the specific states of the thickness parameter, discrete parameter, and range parameter. The multidimensional deviation vector is input into a preset rule model for logical forward derivation, relying on an expert experience rule base for matching and retrieval to obtain the derivation conclusion. Based on the derivation conclusion, a preparation parameter correction command is output. The thickness parameter, discrete parameter, and range parameter are extracted from the multidimensional deviation vector for control logic judgment. When the thickness parameter exceeds the preset over-limit judgment condition, the system's underlying logic determines that the current coating solid content is redundant and generates a command to reduce the additive concentration. When the discrete parameter meets the preset uniformity non-compliance judgment condition, a command to adjust the coating pressure is generated to improve spatial coating unevenness. When the range parameter meets the preset fluid agglomeration judgment condition, it is determined that severe surface tension contraction and accumulation of local high-viscosity media occurs, generating commands to increase the nozzle atomization pressure and raise the curing temperature. The generated preparation parameter correction commands are integrated and encapsulated into a digital correction feedback message and sent.

[0060] To verify the practical application effect of the endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film of this invention, researchers conducted a comparative experiment on the involute tooth surface of a heavy-duty gearbox of a megawatt-level wind turbine. The experimental group used the quantitative analysis method provided by this invention, which includes dynamic curvature compensation factor correction and morphological processing, while the control group used a conventional image analysis method without curvature compensation and removal of residual lubricating media. Under test conditions of an ambient temperature of 25°C, a gear module of 12, and a tooth width of 180 mm, the researchers pre-coated the tooth surface with a protective film of nominal physical thickness of 15.0 μm and covered it with a layer of coal-based polyalphaolefin lubricating residual media with a thickness of 2.0 mm.

[0061] Experimental data show that the initial nominal physical thickness measured in the control group was affected by the nonlinear curvature of the tooth surface and the refraction of residual lubricating medium, resulting in a maximum measurement deviation of ±3.5 μm at the edge of the meshing zone, and significant artifact misjudgments occurred at the accumulation of residual lubricating medium. In contrast, the experimental group, through morphological processing of corrosion followed by expansion, eliminated more than 98% of the optical refraction artifacts caused by residual lubricating medium. Simultaneously, by introducing a dynamic curvature compensation factor to correct projection distortion, the physical thickness measurement value displayed by the real thickness coordinate array generated by the experimental group fluctuated between 14.8 μm and 15.2 μm, with the measurement accuracy deviation strictly locked within ±0.2 μm, significantly better than the ±1.0 μm technical target set in this invention. The experimental results are compared in Table 1 below:

[0062] Furthermore, to assess the effectiveness of uniformity evaluation, researchers compared the consistency between spatial distribution randomness indices and human experience ratings. In quantitative analysis of 20 tooth surface protective film samples with different fluid aggregation degrees, the spatial distribution randomness index calculated by regional distribution entropy accurately identified local accumulation regions with a range parameter greater than 2.5 μm, achieving an accuracy rate of 96.5%. Based on a multi-level decision tree-based rule model, and using multi-dimensional deviation vectors to output preparation parameter correction instructions, the global thickness arithmetic mean deviation of the protective film was reduced by 75% in subsequent production iterations, with the thickness standard deviation decreasing from 0.85 μm to 0.22 μm.

[0063] The optimization effect of process closed loop is shown in Table 2 below:

[0064] The aforementioned quantitative experimental data fully demonstrates that this invention, by performing deep feature stripping on the original microscopic image matrix and deriving a dynamic curvature compensation factor based on gear module and tooth width values, effectively solves the technical problems caused by the complex tooth surface curvature geometry and residual lubricating medium in heavy-duty gearboxes. The experimental group's measurement accuracy is an order of magnitude higher than the control group, and the generated spatial distribution randomness index provides highly reliable feedback support for the manufacturing process, significantly improving the uniformity of the protective film and achieving the expected technical effect.

[0065] Embodiment 1 of this invention: For heavy-duty gearboxes of wind turbines operating in high-humidity coastal environments, the maintenance team uses coal-based polyalphaolefin synthetic oil with matching lubricating additives for surface protection. The system then receives external input of the coal-based polyalphaolefin base oil type and refractive index properties. The main control program searches and compares in a preset optical refractive index calibration library to match the target polarization angle range and target brightness range. Then, the underlying drive module converts the target range parameters into drive level signals to output lighting signals. After the polarized light endoscope probe penetrates into the gearbox, it captures surface-reflected photons that penetrate the lubricating residual medium at the adjusted polarization angle. Subsequently, the photosensitive element array converts the surface-reflected photons into a discrete pixel matrix and packages them into an original microscopic image matrix. At the same time, the lens focal length, magnification, optical axis tilt angle, and physical chord length of the observation field are simultaneously acquired. To eliminate image noise, the processor extracts the pixel gray values ​​from the original microscopic image matrix and performs neighborhood weighted convolution operations based on the two-dimensional discrete pixel kernel matrix to obtain the convolution gray values. The system statistically analyzes the grayscale histogram distribution of convolutional grayscale values ​​and locks the segmentation threshold based on the maximum inter-class variance. This prompts the processor to truncate the gradient magnitude matrix according to the segmentation threshold to obtain a binary mask. Then, a morphological processing step of erosion followed by dilation is performed to remove optical refraction artifacts and extract the pure protective film grayscale matrix. The processor extracts the mean grayscale value of each pixel within the pure protective film grayscale matrix. Substituting the mean grayscale value, lens focal length, and magnification into a nonlinear mapping equation, it derives the initial nominal physical thickness value of the corresponding pixel and establishes a spatial thickness numerical array. The processor substitutes the gear module and tooth width values ​​into the involute equation to solve for the local radius of curvature of the target tooth surface at the center point of the observation field. Combining the optical axis tilt angle and the physical chord length of the observation field, a trigonometric function compensation model is constructed to derive the dynamic curvature compensation factor. The spatial thickness numerical array is multiplied by the dynamic curvature compensation factor to generate a true thickness coordinate array. The system iterates through the true thickness coordinate array to calculate the global thickness arithmetic mean and obtains the standard deviation and range parameters. The true thickness coordinate array is divided into multiple equidistant sub-bands based on the range parameter. Single probability values ​​are extracted from the probability set of pixels falling within each band, and the natural logarithm of each single probability value is calculated and multiplied to obtain discrete information. The discrete information within multiple equidistant sub-bands is summed and the negative value is taken to obtain the regional distribution entropy as a spatial distribution randomness index. The spatial distribution randomness index is compared with a preset compliance threshold matrix to output rating information. The rating information is then subtracted from the global thickness arithmetic mean and a preset process target baseline to generate a multi-dimensional deviation vector. The expert rule model, through forward derivation, discovers that the thickness parameter in the multi-dimensional deviation vector exceeds the preset over-limit judgment condition, prompting the system to determine that the coating liquid solid content is too high. This directly generates an instruction to reduce the additive concentration, encapsulates it as a digital correction feedback message, and sends it.

[0066] Embodiment 2 of this invention: Addressing the microscopic fatigue state at the edge of the meshing zone of a wind turbine gearbox after long-term medium-load operation, technicians inject a repair-strengthening protective film prepared from hydrotreated base oil and domestically produced repair agents. The system simultaneously acquires the type of hydrotreated base oil and its corresponding refractive index properties. The main control program retrieves the target polarization angle range and target brightness range from the calibration library, converts them into drive level signals, and guides the probe front end to project a specific wavelength illumination signal. The probe penetrates deep into the meshing zone to capture surface-reflected photons. After photoelectric conversion, the surface-reflected photons are mapped into a discrete pixel array and packaged into an original microscopic image matrix. The system simultaneously records the lens focal length, magnification, optical axis tilt angle, and physical chord length of the observation field. The computation unit extracts pixel grayscale values ​​and performs convolution operations on a two-dimensional discrete pixel kernel matrix to smooth out local speckles. By statistically analyzing the grayscale histogram distribution of the convolution grayscale values, a segmentation threshold is locked. The gradient amplitude matrix is ​​then truncated, and the generated binarized mask undergoes morphological processing (corrosion followed by dilation) to extract the pure protective film grayscale matrix. The system extracts the average grayscale value of each pixel, along with the lens focal length and magnification, and substitutes them into a nonlinear mapping equation to deduce the initial nominal physical thickness value of the corresponding pixel, constructing a spatial thickness numerical array. The computational engine substitutes the gear module and tooth width values ​​into the involute equation to solve for the local radius of curvature. Using a trigonometric function compensation model, it integrates the local radius of curvature, optical axis tilt angle, and physical chord length of the observation field to derive a dynamic curvature compensation factor. The spatial thickness numerical array is multiplied by the dynamic curvature compensation factor to output a true thickness coordinate array. The system calculates the global thickness arithmetic mean to obtain the standard deviation and range parameters. The true thickness coordinate array is divided into multiple equidistant sub-bands based on the range parameters to statistically determine the pixel probability set. The natural logarithm of each individual probability value is calculated and multiplied to generate discrete information. The discrete information is summed and negatively evaluated to obtain the regional distribution entropy, which serves as a spatial distribution randomness index. The system compares the spatial distribution randomness index with a compliance threshold matrix to output rating information. The spatial thickness value is mapped onto a red-blue pseudo-color spectrum to render a thickness distribution heatmap, locating locally thin areas at the edges. The rating information is compared with the global thickness arithmetic mean and the process target baseline to calculate the difference, generating a multi-dimensional deviation vector. The rule model analyzes the multi-dimensional deviation vector and finds that the discrete parameters meet the preset uniformity non-compliance criteria. This prompts the system to determine that the coating pressure distribution at the edges is uneven, and a command to adjust the coating pressure is generated and encapsulated as a digital correction feedback message for transmission.

[0067] Embodiment 3 of this invention: The heavy-duty gearbox of an offshore wind turbine is exposed to a high-salt-spray corrosive environment year-round. Its surface is covered with a high-viscosity special anti-corrosion protective film. The system correspondingly records the type and refractive index properties of the high-viscosity base oil. The main control program searches the calibration library to obtain the target polarization angle range and target brightness range. In conjunction with the level conversion module, the target parameters are converted into a drive level signal to wake up the illumination signal of the detection front end. The probe enters the tooth profile region to capture surface-reflected photons. Subsequently, the photoelectric array converts the surface-reflected photons into a discrete pixel matrix, packaging them into an original microscopic image matrix. The central processing unit simultaneously reads the lens focal length, magnification, optical axis tilt angle, and physical chord length of the observation field. The pixel grayscale values ​​of the original microscopic image matrix are extracted and convolved to obtain convolution grayscale values. The maximum inter-class variance is calculated using the grayscale histogram distribution waveform to lock the segmentation threshold. The central processing unit truncates the gradient magnitude matrix according to the segmentation threshold to obtain a binary mask. A pre-etching followed by dilation morphological processing is performed to extract the pure protective film grayscale matrix. The grayscale mean is extracted and combined with the lens focal length and magnification, then substituted into a nonlinear mapping equation to deduce the initial nominal physical thickness values ​​of the corresponding pixels, summing them into a spatial thickness numerical array. The computational engine substitutes the gear module and tooth width values ​​into the involute equation to solve for the local radius of curvature of the target tooth surface. Combining the local radius of curvature, optical axis tilt angle, and physical chord length of the observation field of view, a dynamic curvature compensation factor is derived in a trigonometric function compensation model. Multiplication operations are performed on the spatial thickness numerical array and the dynamic curvature compensation factor to generate a true thickness coordinate array. The system iterates through the true thickness coordinate array to calculate the global thickness arithmetic mean, obtaining the standard deviation and range parameters. The true thickness coordinate array is divided into multiple equidistant sub-frequency bands based on the range parameters, and pixel probability sets are statistically analyzed. The natural logarithm of a single probability value is calculated to obtain discrete information. The summation and negative value are then used to output the regional distribution entropy as a spatial distribution randomness indicator. The system compares this indicator with a compliance threshold matrix to generate rating information. The rating information is extracted and compared with the global thickness arithmetic mean to form a multidimensional deviation vector. The rule model extracts the multidimensional deviation vector and finds that the range parameter meets the preset judgment condition for fluid agglomeration. The system determines that the severe imbalance of surface tension of the high-viscosity medium leads to local accumulation. The logic derivation automatically generates instructions to increase the nozzle atomization pressure and adjust the curing temperature, which are then packaged into a digital correction feedback message and sent.

Claims

1. A method for quantitative analysis of the thickness and uniformity of gear protective film in endoscopic images, characterized in that, include: Step 1: Obtain the original microscopic image matrix, lens focal length, magnification, gear module, tooth width, optical axis tilt angle, and physical chord length of the observation field of view of the protective film on the tooth surface with residual lubricating medium inside the heavy-duty gearbox. Step 2: Extract the pixels from the original microscopic image matrix, calculate the gradient magnitude matrix of the pixels, truncate the gradient magnitude matrix based on a preset segmentation threshold to obtain a binarized mask, perform morphological processing of the binarized mask by first erosion and then dilation, remove the optical refraction artifacts generated by the lubricating residual medium, and extract the pure protective film grayscale matrix. Step 3: Extract the gray values ​​of each pixel in the gray matrix of the pure protective film, substitute the gray values, the lens focal length and the magnification into the nonlinear mapping equation, deduce the initial nominal physical thickness value of the corresponding pixel, and summarize all the initial nominal physical thickness values ​​to establish a spatial thickness value array. Step 4: Analyze the local radius of curvature of the target tooth surface at the center point of the observation field of view based on the gear module and the tooth width. Derive the dynamic curvature compensation factor based on the local radius of curvature and the optical axis tilt angle. Multiply the spatial thickness numerical array with the dynamic curvature compensation factor to generate a true thickness coordinate array. Step 5: Divide the real thickness coordinate array into multiple equidistant sub-frequency bands, count the probability set of pixels in the real thickness coordinate array falling into the multiple equidistant sub-frequency bands, calculate the spatial distribution randomness index based on logarithmic accumulation operation, extract the main diagonal elements of the preset compliance threshold matrix to form a one-dimensional judgment benchmark vector, and compare the spatial distribution randomness index with the various level thresholds in the one-dimensional judgment benchmark vector.

2. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 1, characterized in that, Obtain the original microscopic image matrix of the protective film on the gear surface containing residual lubricating medium inside a heavy-duty gearbox, including: Receive the intensity signal reflected from the tooth surface; Adjust the polarization angle according to the tooth surface reflection intensity signal; Capture surface-reflected photons that penetrate the lubricating residual medium at the adjusted polarization angle; The surface-reflected photons are converted into a discrete pixel array, and the discrete pixel array is packaged into the original microscopic image matrix.

3. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 2, characterized in that, Before calculating the gradient magnitude matrix of the pixel, the following steps are included: Extract the pixel grayscale values ​​from the original microscopic image matrix; The pixel grayscale values ​​are convolved according to a preset two-dimensional discrete pixel kernel matrix to obtain convolved grayscale values; The gray-level histogram distribution of the convolutional gray values ​​is statistically analyzed, and the maximum inter-class variance is calculated using an adaptive algorithm to lock the segmentation threshold. The gradient magnitude matrix is ​​truncated according to the segmentation threshold.

4. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 3, characterized in that, Based on the gear module and the tooth width, the local radius of curvature of the target tooth surface at the center point of the observation field is analyzed. Based on the local radius of curvature and the optical axis tilt angle, the dynamic curvature compensation factor is derived, including: Substitute the gear module and the tooth width values ​​into the involute equation to solve for the local radius of curvature; Extract the optical axis tilt angle and the physical chord length of the observation field of view; Based on the local radius of curvature, the optical axis tilt angle, and the physical chord length of the observation field, a trigonometric function compensation model is constructed, and the dynamic curvature compensation factor is derived.

5. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 4, characterized in that, Before dividing the true thickness coordinate array into multiple equidistant sub-bands, the process includes: Traverse the real thickness coordinate array and calculate the global thickness arithmetic mean; The standard deviation and range parameters are calculated based on the arithmetic mean of the global thickness. The true thickness coordinate array is divided into multiple equidistant sub-bands according to the range parameter.

6. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 5, characterized in that, After extracting the main diagonal elements of the preset compliance threshold matrix to construct a one-dimensional judgment benchmark vector, and comparing the spatial distribution randomness index with each level threshold in the one-dimensional judgment benchmark vector, the process includes: Extract the spatial thickness values ​​from the actual thickness coordinate array; The spatial thickness value is mapped to a preset red-blue pseudo-color spectrum; Render the red-blue pseudo-color spectrum and output a thickness distribution heatmap; Based on the pseudo-color extreme value regions in the thickness distribution heatmap, locate the locally excessively thick, excessively thin, and exposed areas of the tooth surface protective film.

7. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 6, characterized in that, Before receiving the tooth surface reflection intensity signal, the process includes: Obtain the base oil type and refractive index properties from external input; Match the target polarization angle range and target brightness range corresponding to the refractive index attribute in the preset optical refractive index calibration library; The target polarization angle range and the target brightness range are converted into driving level signals; The lighting signal is output according to the driving level signal.

8. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 7, characterized in that, Spatial distribution randomness indices are obtained based on logarithmic summation, including: Extract a single probability value from the probability set of the pixels; Calculate the natural logarithm of the single probability value; multiply the single probability value by the natural logarithm to obtain the discrete information content; The discrete information quantities within the multiple equidistant sub-frequency bands are summed and the negative value is taken to obtain the regional distribution entropy, which is used as the spatial distribution randomness index.

9. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 8, characterized in that, After extracting the main diagonal elements of the preset compliance threshold matrix to construct a one-dimensional judgment benchmark vector, and comparing the spatial distribution randomness index with each level threshold in the one-dimensional judgment benchmark vector, the process includes: Extract the rating information and the arithmetic mean of the global thickness; The difference between the global thickness arithmetic mean, thickness standard deviation, and range parameter and the preset process target baseline is calculated, and the extreme value normalization processing is performed on the difference to eliminate the numerical magnitude difference caused by the dimensionless information entropy and the micrometer-scale length physical quantity. The normalized data is spliced ​​together to generate a multidimensional deviation vector. The multidimensional deviation vector is input into a preset rule model. The rule model constructs a multi-level decision tree that includes a thickness deviation threshold, a dispersion deviation threshold, and a range deviation threshold. The node conditions of the feature dimensions in the multidimensional deviation vector are compared sequentially, and logical forward deduction is performed to obtain the deduction conclusion. Based on the derivation conclusion, output the preparation parameter correction command and send the preparation parameter correction command.

10. The endoscopic image quantitative analysis method for the thickness and uniformity of the gear protective film according to claim 9, characterized in that, Based on the derivation conclusion, output a preparation parameter correction command, and send the preparation parameter correction command, including: Thickness parameters, discrete parameters, and range parameters are extracted from the multidimensional deviation vector; when the thickness parameter exceeds the preset over-limit judgment condition, an instruction to reduce the additive concentration is generated; When the discrete parameter meets the preset uniformity failure judgment condition, an instruction to adjust the coating pressure is generated. When the range parameter meets the preset fluid agglomeration phenomenon judgment condition, an instruction is generated to increase the nozzle atomization pressure and increase the curing temperature. The generated preparation parameter correction instructions are encapsulated into a digital correction feedback message and sent.

Images