Intelligent monitoring method and system for quality of metal component based on machine vision

Abstract

The application relates to a kind of metal component quality intelligent monitoring method and system based on machine vision.The method comprises: based on micro facet bidirectional reflectance distribution function model, the reflection separation of multispectral polarized data cube is carried out, and the thermal radiation component data cube is obtained;Based on the global gray scale temperature estimation of thermal radiation component data cube by two-band colorimetric temperature measurement algorithm, the initial temperature field is obtained;Based on the preset temperature-emissivity joint inversion framework, the initial temperature field is taken as the starting benchmark, the thermal radiation component data cube is alternately iteratively optimized and solved, and the real temperature field and spectral emissivity spatial distribution are generated;Based on the double-layer oxide film model and the cavity radiation enhancement effect model, the inverse physical parameter mapping processing is carried out on the spectral emissivity spatial distribution, and the oxide layer thickness distribution and surface micro-topology height field distribution are obtained.The method can solve the missed detection and misjudgment problem caused by thermal radiation flooding active imaging in high-temperature visual detection.

Description

Technical Field

[0001] This invention belongs to the field of metal component inspection technology, and in particular relates to a machine vision-based intelligent monitoring method and system for the quality of metal components. Background Technology

[0002] With the development of high-end equipment manufacturing technology, surface quality testing technology for metal components in high-temperature service has emerged, which is used to test metal components in extreme high-temperature service environments above 800 degrees Celsius.

[0003] At high temperatures, the spontaneous infrared radiation from metal components can overwhelm the active illumination signal. Traditional techniques mainly employ cooling-type detection, which involves first cooling the high-temperature component to room temperature before conducting surface quality inspection, thereby avoiding interference from high-temperature thermal radiation. Alternatively, filtering-type detection can be used, which involves installing a narrow-band filter in front of the inspection camera to find an imageable "window" band in the thermal radiation spectrum, suppressing the intense thermal radiation as pure noise, and thus obtaining the imaging signal of the component surface.

[0004] However, the above methods, such as cooling-type inspection, change the thermal stress distribution inside the component during the cooling process, causing the micro-cracks that opened at high temperature to close, resulting in serious missed detections and failing to truly reflect the surface quality of the component under high-temperature service conditions; and filtering-type inspection treats the thermal radiation that carries rich information about the surface condition of the metal component as noise and filters it out, losing key surface micro-topological information such as cracks, oxide layer thickness, and micro-roughness. Summary of the Invention

[0005] Therefore, it is necessary to provide a machine vision-based intelligent monitoring method and system for the quality of metal components that can utilize thermal radiation to monitor the quality of high-temperature metal components, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a machine vision-based intelligent monitoring method for the quality of metal components, comprising:

[0007] The original multispectral polarization image data with different exposure times were acquired synchronously by a multi-channel imaging system under the condition of thermal radiation dominating the surface of a high-temperature metal component. Radiometric calibration correction, polarization calibration correction, and multi-exposure fusion processing were performed on the original multispectral polarization image data to obtain a multispectral polarization data cube. The multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band under at least four spectral bands. The multi-channel imaging system is constructed by combining multispectral spectroscopy and polarization analysis.

[0008] For each pixel in the multispectral polarization data cube, Stokes vector calculation is performed based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity to obtain the linear polarization degree value of each pixel. Based on the polarization degree-reflection component ratio mapping table of the micro-facet bidirectional reflection distribution function model, the reflection of the multispectral polarization data cube is separated pixel by pixel through the linear polarization degree value of each pixel to obtain the thermal radiation component data cube.

[0009] The initial temperature field is obtained by performing global gray-level temperature estimation on the thermal radiation component data cube based on the dual-band colorimetric temperature measurement algorithm; the dual-band colorimetric temperature measurement algorithm estimates the temperature of the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube respectively.

[0010] Based on the preset temperature-emissivity joint inversion framework, the initial temperature field is used as the starting point, and the thermal radiation component data cube is iteratively optimized to generate the real temperature field and spectral emissivity spatial distribution.

[0011] Based on the double-layer oxide film model and the cavity radiation enhancement effect model, the spatial distribution of spectral emissivity is subjected to inverse physical parameter mapping to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution.

[0012] The system integrates the real temperature field, oxide layer thickness distribution, and surface micro-topology height field distribution to construct a multi-channel physical feature map. This multi-channel physical feature map is then input into a pre-trained defect judgment network for real-time inference, generating intelligent defect judgment results for the target component.

[0013] Secondly, this application also provides a machine vision-based intelligent quality monitoring system for metal components, comprising:

[0014] The data acquisition module is used to acquire raw multispectral polarization image data with different exposure times synchronously acquired by the multi-channel imaging system under the environment dominated by thermal radiation on the surface of high-temperature metal components. It then performs radiometric calibration correction, polarization calibration correction, and multi-exposure fusion processing on the raw multispectral polarization image data to obtain a multispectral polarization data cube. The multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band in at least four spectral bands. The multi-channel imaging system is constructed by combining multispectral spectroscopy and polarization analysis.

[0015] The polarization-guided component separation module is used to perform Stokes vector calculations on each pixel in the multispectral polarization data cube based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity, to obtain the linear polarization degree value of each pixel. Based on the polarization degree-reflection component ratio mapping table of the micro-facet bidirectional reflection distribution function model, the module performs reflection separation pixel by pixel on the multispectral polarization data cube through the linear polarization degree value of each pixel to obtain the thermal radiation component data cube.

[0016] The temperature measurement module is used to perform global gray-level temperature estimation on the thermal radiation component data cube based on the dual-band colorimetric temperature measurement algorithm to obtain the initial temperature field; the dual-band colorimetric temperature measurement algorithm estimates the temperature of the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube respectively.

[0017] The inversion module is used to perform alternating iterative optimization and solution of the thermal radiation component data cube based on the preset temperature-emissivity joint inversion framework, with the initial temperature field as the starting point, to generate the real temperature field and the spatial distribution of spectral emissivity.

[0018] The feature module is used to perform inverse physical parameter mapping on the spatial distribution of spectral emissivity based on the double oxide film model and the cavity radiation enhancement effect model, so as to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution.

[0019] The quality module is used to fuse the real temperature field, oxide layer thickness distribution, and surface micro-topology height field distribution to construct a multi-channel physical feature map. The multi-channel physical feature map is then input into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described intelligent monitoring methods for the quality of metal components based on machine vision.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-described machine vision-based intelligent monitoring methods for the quality of metal components.

[0022] The aforementioned intelligent monitoring method and system for the quality of metal components based on machine vision synchronously acquires multi-band polarized radiation data of the surface of high-temperature components through multispectral polarization imaging. It uses the polarization characteristics of the micro-facet bidirectional reflection distribution model to separate the mixed signal into a pure thermal radiation component. The initial temperature field is obtained by dual-band colorimetric thermometry. Then, the thermal radiation data is iteratively optimized in the temperature-emissivity joint inversion framework to decouple the real temperature field and the spatial distribution of spectral emissivity from the coupled signal. Based on the double-layer oxide film model and the cavity radiation enhancement effect model, the spectral emissivity is inversely mapped to physical parameters to recover the oxide layer thickness distribution and the surface micro-topology height field distribution. The multi-physics information is fused into a feature map and input into the defect judgment network for real-time inference. By repositioning thermal radiation signals, which were originally considered interference noise in high-temperature environments, as physical carriers carrying surface state information, and by establishing a multi-dimensional inverse calculation link from radiation data to temperature field, material properties, and micro-geometry, the system simultaneously completes real temperature measurement, surface oxidation state assessment, micron-level defect geometric inversion, and intelligent classification and recognition in a single detection process. This fundamentally overcomes the problems of missed detection and misjudgment caused by active imaging being submerged by thermal radiation in traditional high-temperature visual inspection. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the intelligent quality monitoring method for metal components based on machine vision according to the present invention.

[0025] Figure 2 This is a structural diagram of the intelligent monitoring system for the quality of metal components based on machine vision, as presented in this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] In one embodiment, such as Figure 1As shown, a machine vision-based intelligent monitoring method for the quality of metal components is provided. This embodiment illustrates the application of this method to an intelligent monitoring terminal for the quality of metal components. It is understood that this method can also be applied to an intelligent monitoring server for the quality of metal components, and further to an intelligent monitoring system for the quality of metal components that includes both an intelligent monitoring terminal and an intelligent monitoring server, and is implemented through the interaction between the two. In this embodiment, the method includes the following steps:

[0028] S101. Acquire raw multispectral polarization image data of different exposure durations synchronously acquired by the multi-channel imaging system under the environment dominated by thermal radiation on the surface of high-temperature metal components, and perform radiometric calibration correction, polarization calibration correction and multi-exposure fusion processing on the raw multispectral polarization image data to obtain a multispectral polarization data cube; wherein, the multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band under at least four spectral bands; the multi-channel imaging system is constructed by combining multispectral spectral dispersion and polarization analysis.

[0029] This illustration demonstrates how an intelligent monitoring terminal for metal component quality acquires raw multispectral polarization image data of different exposure durations simultaneously collected by a multi-channel imaging system under conditions dominated by thermal radiation from the surface of a high-temperature metal component. The multi-channel imaging system is encapsulated within a stainless steel housing with a water-cooled jacket. The inner wall of the housing features continuous serpentine water-cooling channels, with the cooling water flow rate designed to match the surface temperature of the object being measured and the thermal conductivity of the housing material, ensuring that the temperature inside the housing remains within the permissible operating range of the electronic components. A quartz glass observation window is opened on the side of the housing facing the component under test. The window material is sapphire or fused silica, which have high transmittance and high temperature resistance in the 0.4 to 14 micrometer wavelength range. To prevent secondary reflections of thermal radiation from the component surface on the inner surface of the window, thus introducing stray light, the normal direction of the window is tilted relative to the optical axis of the camera, with an tilt angle between 5° and 10°, causing the thermal radiation reflected from the inner surface of the window to deviate from the camera's field of view. The thermal radiation and surface-reflected light entering the observation window are collimated by a forward-facing telescope system, converting light from different spatial positions on the measured surface into approximately parallel beams.

[0030] Specifically, a multi-channel imaging system can be constructed by combining a multispectral beam splitting module and a polarization analysis module. The multispectral beam splitting module separates the incident light into at least four spectral bands, while the polarization analysis module acquires the radiation intensity information in the four polarization directions for each spectral band. After beam splitting and polarization analysis, the four spectral channels, each with four polarization directions, resulting in a total of sixteen independent optical signals, are received by sixteen corresponding area array detectors. The detector type is selected based on the characteristics of each spectral band. For example, the visible-near-infrared channel uses a back-illuminated scientific-grade complementary metal-oxide-semiconductor (sCMOS) sensor, the short-wave infrared channel uses an indium gallium arsenide (InGaAs) focal plane array, and the mid-wave infrared and long-wave infrared channels use a mercury cadmium telluride (HgCdTe) or uncooled vanadium oxide microbolometer focal plane array. Furthermore, for each spectral channel, three frames of images with different exposure times are continuously captured within one acquisition cycle. Short exposures are used to capture details in high-intensity bands, medium exposures are used for normal operating bands, and long exposures are used to capture weak signals in low-intensity bands.

[0031] For example, radiometric calibration uses a blackbody radiation source of known temperature, placed on the detection plane, and collects the detector response values ​​of the blackbody in various spectral channels and polarization directions to establish a linear mapping relationship from the detector digital output value to the incident radiative flux, which follows the mathematical expression: ,in, Indicates wavelength Observation angle Azimuth The incident radiation flux at that point. Indicates wavelength The corresponding scaling factors, This represents the digital output value of the detector at the corresponding wavelength and angle. Indicates wavelength The corresponding dark current offset can be obtained by the intelligent monitoring terminal for the quality of metal components through a blackbody radiation source with a known temperature. and The specific values ​​are then determined. Furthermore, an integrating sphere light source is used in conjunction with a high extinction ratio polarizer with a known polarization direction as a standard light source. By rotating the angle of the standard polarizer, the detector response in each channel is recorded, and the Mueller matrix for each channel is calculated. Its mathematical expression is ,in, The Stokes vector representing the incident light. This represents the Stokes vector of the outgoing light after passing through the polarization analysis module. A 4×4 Mueller matrix is ​​used to correct systematic errors during polarization measurements. Optionally, the intelligent monitoring terminal for metal component quality performs multi-exposure fusion processing on images with different exposure times, following a weighted formula: ,in, This represents the pixel values ​​of the merged image. Indicates the number of exposures. Indicates the first Second exposure image at pixels The weight of the position, Indicates the first Second exposure image at pixels The intelligent monitoring terminal for metal component quality can calculate the signal-to-noise ratio of pixel grayscale values ​​based on the pixel values. Finally, a multispectral polarization data cube is obtained, which contains the spatial distribution of radiation intensity in the four polarization directions corresponding to each of the at least four spectral bands.

[0032] S102. For each pixel in the multispectral polarization data cube, Stokes vector calculation is performed based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity to obtain the linear polarization degree value of each pixel. Based on the polarization degree-reflection component ratio mapping table of the micro-facet bidirectional reflection distribution function model, the reflection of the multispectral polarization data cube is separated pixel by pixel through the linear polarization degree value of each pixel to obtain the thermal radiation component data cube.

[0033] Furthermore, the intelligent monitoring terminal for metal component quality calculates the Stokes vector for each pixel in the multispectral polarization data cube based on the radiation intensity in the four polarization directions of the spatial distribution of radiation intensity. The calculation expression is as follows ,in, Indicates the total radiation intensity. This represents the difference in radiation intensity between the 0° and 90° polarization directions. This represents the difference in radiation intensity between the 45° and 135° polarization directions. Represents the circular polarization component. , , , These represent the radiation intensity in the four polarization directions. The linear polarization degree value of each pixel is calculated based on the Stokes vector. The calculation formula is ,in, Represents the total intensity of the linearly polarized components. Represents total radiation intensity and degree of linear polarization. The value ranges from 0 to 1. The larger the value, the higher the proportion of polarization component in the radiation signal of that pixel.

[0034] Optionally, the intelligent monitoring terminal for metal component quality pre-constructs a polarization degree-reflection component ratio mapping table based on a micro-facet bidirectional reflection distribution function model. This model characterizes that at the same temperature, higher surface roughness results in a more dispersed distribution of the micro-facet normal vectors, leading to a lower degree of radiation polarization. At the same roughness, increased temperature causes a change in the complex refractive index, which in turn changes the Fresnel reflection coefficient, ultimately increasing the degree of polarization. The expression for the micro-facet bidirectional reflection distribution function model is as follows: ,in, Indicates the angle of incidence. Indicates the incident azimuth angle. Indicates the angle of reflection. Indicates the azimuth angle of the reflection. Indicates wavelength. Represents the Fresnel reflection coefficient. Represents the normal distribution function of the micro-element. This represents the angle between the normal of the micro-element and the normal of the macro-surface. This represents the surface roughness parameter. This represents the geometric attenuation factor. By simulating the proportion of reflection components corresponding to different degrees of linear polarization using a micro-facet bidirectional reflection distribution function model, and constructing a mapping table between polarization degree and reflection component proportion, the intelligent monitoring terminal for metal component quality can query this mapping table using the linear polarization degree value of each pixel to obtain the reflection component proportion of each pixel. and in accordance with Reflection separation is performed pixel-by-pixel, where... Indicates the thermal radiation component in a pixel ,wavelength polarization direction The radiation intensity at that location This represents the total radiation intensity at the corresponding position in the multispectral polarization data cube, ultimately yielding the thermal radiation component data cube.

[0035] S103. Based on the dual-band colorimetric thermometry algorithm, global gray-level temperature estimation is performed on the thermal radiation component data cube to obtain the initial temperature field; the dual-band colorimetric thermometry algorithm estimates the temperature of the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube, respectively.

[0036] Specifically, Planck's law of blackbody radiation is: ,in, Indicates the blackbody at wavelength ,temperature Monochromatic radiance at that time Denotes Planck's constant. This represents the speed of light in a vacuum. Represents Boltzmann's constant. This represents the absolute temperature of the blackbody. For the surface of a high-temperature metallic component, its thermal radiation component can be approximated as gray-body radiation, i.e. ,in Indicates the surface of the metal component at wavelength ,temperature Radiative exitance below This represents the spectral emissivity of the metal component surface at a corresponding wavelength and temperature. The dual-band colorimetric thermometry algorithm selects the short-wave infrared channel band. and mid-wave infrared channel band Assuming the ratio of the spectral emissivity of the two bands is... , Since is a constant and can be preset through material properties, the ratio of radiative exitance of the two bands is: The temperature estimation formula can be obtained. The intelligent monitoring terminal for the quality of metal components extracts the radiation intensity of the short-wave infrared channel and the mid-wave infrared channel in the thermal radiation component data cube, and substitutes them into the above formula to calculate the gray-scale temperature of each pixel. Optionally, spatial smoothing is performed on the gray-scale temperature of all pixels. The smoothing can be performed using an edge-preserving bilateral filtering algorithm to suppress noise and retain edge areas with large temperature gradients, and finally obtain the initial temperature field in the global range.

[0037] S104. Based on the preset temperature-emissivity joint inversion framework, the initial temperature field is used as the starting point to iteratively optimize and solve the thermal radiation component data cube to generate the real temperature field and spectral emissivity spatial distribution.

[0038] Specifically, the core of the temperature-emissivity joint inversion framework is to construct an optimized objective function. The intelligent monitoring terminal for the quality of metal components uses the alternating direction multiplier method for iterative optimization, and the temperature field can be fixed during the iteration process. Solve for the spatial distribution of spectral emissivity The optimal solution can be found using the gradient descent method; furthermore, the spatial distribution of spectral emissivity is fixed. Update the temperature field The optimal solution is obtained by combining the dual-band colorimetric temperature measurement results with constraints during the update process. The iterative process continues until the change in the objective function is less than a preset convergence threshold. The temperature field obtained at this point is the true temperature field, and the corresponding spectral emissivity distribution is the spatial distribution of spectral emissivity. The intelligent monitoring terminal for the quality of metal components can adjust the regularization coefficient... Balancing the accuracy of data fitting with the physical rationality of the solution.

[0039] S105. Based on the double-layer oxide film model and the cavity radiation enhancement effect model, the spatial distribution of spectral emissivity is subjected to inverse physical parameter mapping to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution.

[0040] Furthermore, the bilayer oxide film model is used to describe the effect of the oxide layer on spectral emissivity. Specifically, the oxide layer effect is particularly pronounced in the short-wavelength direction, and the affected wavelength range expands continuously with increasing oxidation temperature and time. The effective emissivity calculation formula for the bilayer oxide film model is as follows: ,in, This indicates that the bilayer oxide film model is at wavelength Oxide layer thickness Effective emissivity at the following levels This represents the effective reflectance of the double-layer oxide film model. The expression for effective reflectance is: ,in, This represents the reflectance of the upper surface of the oxide layer. This represents the reflectance at the interface between the oxide layer and the metal substrate. This indicates the wavenumber of light in the oxide layer. , Indicates the oxide layer at wavelength The complex refractive index is calculated. The intelligent monitoring terminal for metal component quality uses multi-band emissivity values ​​from the spatial distribution of spectral emissivity, substitutes them into the above formula, and performs nonlinear least-squares fitting to obtain the oxide layer thickness for each pixel. This leads to the formation of an oxide layer thickness distribution.

[0041] Optionally, a cavity radiation enhancement effect model is used to describe the enhancement effect of surface micro-topological defects on emissivity. Reflected light rays are generated on the inner wall of the defect cavity. Each ray propagates within the cavity; upon each collision with the inner wall, part of the energy is absorbed, and the remainder is reflected. The reflection direction is determined by the bidirectional reflection distribution function model of the micro-surface elements on the inner wall surface. The light ray continues to track until its energy decays to less than one ten-thousandth of its initial value, or escapes from the cavity opening. The effective emissivity formula is: ,in, This represents the effective emissivity in the presence of microcavity defects. Indicates the emissivity of a flat surface. Represents the geometric parameters of the microcavity. The cavity enhancement factor is represented by a parameterized expression that can be pre-fitted through Monte Carlo ray tracing simulation. The intelligent monitoring terminal for metal component quality uses emissivity anomalies in the spatial distribution of spectral emissivity, combined with a cavity radiation enhancement effect model, to perform inverse solving to obtain the microcavity geometric parameters for each emissivity anomaly region. Furthermore, by mapping the microcavity geometric parameters to the surface microtopological height field, the surface microtopological height field distribution is reconstructed. This distribution accurately reflects the microtopological morphology of the metal component surface.

[0042] S106. The real temperature field, oxide layer thickness distribution and surface micro-topology height field distribution are integrated to construct a multi-channel physical feature map, and the multi-channel physical feature map is input into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component.

[0043] The intelligent monitoring terminal for metal component quality integrates the real temperature field, oxide layer thickness distribution, and surface micro-topology height field distribution to construct a multi-channel physical feature map. This multi-channel physical feature map is then input into a pre-trained defect determination network for real-time inference. This defect determination network can employ an encoder-decoder structure, with its output layer using the softmax function. The output expression is as follows: ,in, Represents pixels Belongs to the defect category The probability, Indicates the defect category corresponding to the network output layer. The output value, This indicates the total number of defect categories.

[0044] The aforementioned intelligent monitoring method for the quality of metal components based on machine vision effectively solves the problems of low detection accuracy and single information extraction caused by thermal radiation interference in high-temperature environments. It realizes the synchronous extraction of multi-dimensional physical information, improves the accuracy and reliability of defect detection, and has strong physical interpretability and real-time performance, meeting the needs of industrial online inspection.

[0045] In one embodiment, based on a preset temperature-emissivity joint inversion framework, using the initial temperature field as the starting point, the thermal radiation component data cube is iteratively optimized to generate the true temperature field and spectral emissivity spatial distribution, including:

[0046] S21. A data fidelity term is constructed based on the forward coupling mapping function of the micro-surface element bidirectional reflection distribution function model, the thermal radiation component data cube, and the initial temperature field; the data fidelity term is used to characterize the difference between the measured radiation flux and the predicted value.

[0047] Schematic, the forward coupling mapping function of the micro-facet bidirectional reflectance distribution function model is the forward modeling result of the micro-facet bidirectional reflectance distribution function in the joint inversion objective function on the temperature field and the spectral emissivity field, and its expression is: ,in, This represents the forward modeling result of the two-way reflectance distribution function of the micro-surface element with respect to the temperature field and the spectral emissivity field. Represents the spectral emissivity field at wavelength and temperature field Spectral emissivity under the following conditions Indicates the blackbody at wavelength and temperature Monochromatic radiance at that time The reflection coefficient represents the bidirectional reflection distribution function of the micro-surface element. Indicates the angle of incidence. Indicates the angle of reflection. Indicates ambient radiation at wavelength The radiative exitance below.

[0048] Specifically, the intelligent monitoring terminal for metal component quality substitutes the initial temperature field into the forward coupling mapping function to calculate the predicted radiative flux of each pixel under different wavelengths and polarization directions. Furthermore, it combines the measured radiative flux in the thermal radiation component data cube to construct a data fidelity term. The expression for the data fidelity term is as follows: ,in, A cube representing the thermal radiation component data. The weighted L2 norm, weighted by the reciprocal of the measurement noise variance, is calculated using the following formula: In the formula, Indicates wavelength polarization direction Pixels The corresponding weighting coefficient is the reciprocal of the measurement noise variance. The measurement noise variance can be calculated by repeatedly collecting radiation data from the same location. The intelligent monitoring terminal for the quality of metal components can adjust the value of the weighting coefficient to highlight the fitting weight of high signal-to-noise ratio areas and improve the rationality of the data fidelity terms.

[0049] S22. Based on the emissivity spectral smoothing constraint, the roughness-emissivity correlation constraint, and the oxide layer thickness spatial smoothing constraint, a joint inversion objective function is generated by combining the data fidelity term.

[0050] The expression for the joint inversion objective function is as follows:

[0051]

[0052] In the formula, For temperature field, For the spectral emissivity field, This represents the forward modeling results of the bidirectional reflectance distribution function of the micro-surface element with respect to the temperature field and spectral emissivity field. For the thermal radiation component data cube, This represents the weighted L2 norm, which is calculated by weighting the inverse of the measurement noise variance. For emissivity spectrum smoothing constraint terms, For roughness-emissivity related constraints, This is the total variational regularization term of the spatial smoothing constraint term for oxide layer thickness. , and This is the regularization parameter.

[0053] For example, emissivity spectral smoothing constraint term To ensure the spectral continuity and smoothness of the spectral emissivity field, its expression is: ,in, Represents pixel coordinates, and Indicates different spectral bands, and Representing pixels At wavelength and The spectral emissivity is calculated using a roughness-emissivity correlation constraint. The emissivity spectral smoothing constraint penalizes sharp jumps in spectral emissivity across different wavelength bands, ensuring that the retrieved spectral emissivity conforms to the physical properties of the metallic material. The expression used to establish the coupling relationship between the spectral emissivity field and surface roughness is as follows: ,in, Indicates the reference wavelength. This represents the reference emissivity of a flat surface at a reference wavelength. Represents the correlation coefficient. Represents pixels The surface roughness parameters at the location can be predetermined by the intelligent monitoring terminal for metal component quality through measurement using standard samples. and The range of values ​​for . The total variational regularization term of the oxide layer thickness spatial smoothness constraint. The expression used to ensure the spatial continuity of the oxide layer thickness distribution is as follows: ,in, Represents pixels Oxide layer thickness at that location and These represent the oxide layer thickness at... direction and Spatial partial derivatives in the direction.

[0054] S23. Based on the joint inversion objective function, the temperature field variables and spectral emissivity field variables are updated alternately and iteratively through the alternating direction multiplier method until the convergence condition is met, and the true temperature field and spectral emissivity spatial distribution are output.

[0055] Specifically, the alternating direction multiplier method solves for the optimal solution of the spectral emissivity field variable by fixing the temperature field variable, then updates the optimal solution of the temperature field variable while fixing the spectral emissivity field variable, repeating these two steps until the convergence condition is met. When solving for the spectral emissivity field variable by fixing the temperature field variable, the intelligent monitoring terminal for metal component quality substitutes the temperature field into the joint inversion objective function. At this point, the joint inversion objective function only relates to the spectral emissivity field variable, and the solution process can use the gradient descent method. The gradient calculation is based on the differentiability of the forward coupling mapping function, and its gradient expression is: ,in, Describes the joint inversion objective function. Represents the weighted coefficient matrix. This represents the partial derivative of the forward coupling mapping function with respect to the spectral emissivity field. and These represent the gradients of the emissivity spectral smoothing constraint term and the roughness-emissivity correlation constraint term with respect to the spectral emissivity field, respectively. When updating the temperature field variable while keeping the spectral emissivity field variable fixed, the intelligent monitoring terminal for metal component quality substitutes the spectral emissivity field into the joint inversion objective function and uses the same gradient descent method to solve for the optimal solution of the temperature field. Its gradient expression is: ,in, This represents the partial derivative of the forward coupling mapping function with respect to the temperature field.

[0056] In one embodiment, based on the joint inversion objective function, the temperature field and spectral emissivity field are iteratively updated alternately using the alternating direction multiplier method until the convergence condition is met, outputting the true spatial distribution of the temperature field and spectral emissivity, including:

[0057] S31. Based on auxiliary variables, the joint inversion problem corresponding to the joint inversion objective function is decomposed into auxiliary variable update subproblems, emissivity and temperature update subproblems, and Lagrange multiplier update subproblems.

[0058] Schematic, the auxiliary variables introduced by the intelligent monitoring terminal for metal component quality include surface roughness, oxide layer thickness, and radiation flux, which are used to correspond to different constraint terms in the joint inversion objective function. The set of auxiliary variables can be represented as follows: ,in, Represents the auxiliary variable of surface roughness. Indicates the auxiliary variable of oxide layer thickness. This represents the auxiliary variable for radiation flux. It is achieved by introducing the Lagrange multiplier matrix. ,in, Denotes the Lagrange multipliers corresponding to the auxiliary variable of surface roughness. Denotes the Lagrange multipliers corresponding to the auxiliary variable of oxide layer thickness. Represents the Lagrange multipliers corresponding to the auxiliary variables of radiation flux, transforming the joint inversion objective function into an augmented Lagrange function, the expression of which is: ,in, Indicates the penalty parameter. This represents the inner product operation. This represents the L2 norm. Based on this augmented Lagrangian function, the intelligent monitoring terminal for metal component quality decomposes the joint inversion problem into three independent sub-problems, namely, the auxiliary variable update sub-problem. Emissivity and Temperature Update Subproblem And the Lagrange multiplier update subproblem .

[0059] S32. Fix the temperature field variable and spectral emissivity field variable of the current iteration step, solve the auxiliary variable update subproblem based on the total variational regularization term through the projection solution algorithm, and smooth and denoise the current intermediate solution to obtain the updated auxiliary variables.

[0060] Optionally, the solution to the auxiliary variable update subproblem needs to be performed independently for each auxiliary variable, where the oxide layer thickness is an example. The solution requires the use of the total variation regularization term, and its solution expression is: ,in, This represents the updated oxide layer thickness auxiliary variable. Indicates the oxide layer thickness distribution. Denotes the Lagrange multipliers corresponding to the auxiliary variable of oxide layer thickness. This represents the penalty parameter. The intelligent monitoring terminal for the quality of metal components can use the Chambolle projection algorithm to solve this optimization problem. The iterative formula for the projection solution algorithm is as follows: ,in, Indicates the current iteration step. This represents the total variation projection operator, used to project input values ​​into a feasible region that satisfies total variation constraints. For surface roughness auxiliary variables... and radiation flux auxiliary variables The solutions are as follows: and ,in, and These represent the updated auxiliary variables for surface roughness and radiation flux, respectively. and These represent the Lagrange multipliers of the corresponding auxiliary variables. After the solution is completed, the intelligent monitoring terminal for metal component quality obtains the updated set of auxiliary variables. .

[0061] S33. With the updated auxiliary variables fixed, the temperature field variables and spectral emissivity field variables are updated alternately based on the gradient descent algorithm and the partial derivative matrix through the Jacobian matrix coupling mechanism to solve the emissivity and temperature update subproblems, and the updated temperature field variables and updated spectral emissivity field variables are obtained.

[0062] Optionally, the core of solving the emissivity and temperature update subproblem is to minimize the values ​​of the augmented Lagrangian function with respect to the temperature field variable and the spectral emissivity field variable based on the gradient descent algorithm. Its gradient expression is as follows: ,in, This represents the gradient vector of the augmented Lagrangian function with respect to the temperature field variable and the spectral emissivity field variable. Represents the gradient of the data fidelity term. and These represent the gradients of the emissivity spectral smoothing constraint term and the roughness-emissivity correlation constraint term, respectively. The Jacobian matrix represents the partial derivatives of the forward coupling mapping function with respect to the temperature and spectral emissivity field variables. The Jacobian matrix coupling mechanism describes the interaction between the temperature and spectral emissivity field variables, and its expression is: ,in, This represents the matrix of partial derivatives of the forward coupling mapping function with respect to the temperature field variables. This represents the partial derivative matrix of the forward coupling mapping function with respect to the spectral emissivity field variable.

[0063] S34. Based on the updated temperature field variables and the updated spectral emissivity field variables, update the Lagrange multiplier matrix to solve the Lagrange multiplier update subproblem.

[0064] For example, solving the Lagrange multiplier update subproblem follows the gradient ascent criterion, maximizing the augmented Lagrange function by updating the Lagrange multiplier matrix. The update formulas for each multiplier in the Lagrange multiplier matrix are as follows: , and ,in , and They represent the updated Lagrange multipliers, , and These represent the Lagrange multipliers before the update. Indicates the penalty parameter. This represents the forward coupling mapping result calculated based on the updated temperature field variables and spectral emissivity field variables. and These represent the surface roughness parameters and oxide layer thickness distribution obtained by back-calculation based on the updated spectral emissivity field variables, respectively. , and These represent the updated auxiliary variables. The intelligent monitoring terminal for the quality of metal components can adjust the penalty parameters. The value of is used to adjust the update magnitude of the Lagrange multipliers, avoiding oscillations during the update process and ensuring that the Lagrange multiplier matrix can effectively guide the iterative process towards the optimal solution. After the update is completed, a new Lagrange multiplier matrix is ​​obtained. .

[0065] S35. Iteratively execute the auxiliary variable update subproblem, the emissivity and temperature update subproblem, and the Lagrange multiplier update subproblem until the relative change of the joint inversion objective function is less than the preset convergence threshold. Map the obtained temperature field variable to the real temperature field and the obtained spectral emissivity field variable to the spectral emissivity spatial distribution.

[0066] Optionally, the formula for calculating the relative change of the joint inversion objective function is as follows: ,in, This represents the relative change in the joint inversion objective function. and These represent the joint inversion objective function values ​​for the (k+1)th and kth iterations, respectively. The intelligent monitoring terminal for the quality of metal components can be set with a preset convergence threshold. ,when When the iteration reaches convergence, the loop stops. If convergence is not achieved, the updated auxiliary variables, temperature field variables, spectral emissivity field variables, and Lagrange multiplier matrix are used as the initial values ​​for the next iteration, and the iteration update step continues. When the iteration reaches convergence, the intelligent monitoring terminal for metal component quality maps the final temperature field variables to physical quantities using the following formula: ,in, Represents the true temperature field in pixels Temperature value at that location, This indicates the converged temperature field variable in pixels. The value at that location, Represents the temperature mapping coefficient. This represents the temperature mapping offset; simultaneously, the converged spectral emissivity field variables are mapped to physical quantities using the following formula: ,in, Indicates the spatial distribution of spectral emissivity in pixels ,wavelength Emissivity value at that location, This represents the value of the spectral emissivity field variable at the corresponding position after convergence. Represents the emissivity mapping coefficient. This represents the emissivity mapping offset.

[0067] In one embodiment, based on the bilayer oxide film model and the cavity radiation enhancement effect model, the spatial distribution of spectral emissivity is subjected to inverse physical parameter mapping to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution, including:

[0068] S41. Based on the emissivity sequence values ​​of each spatial pixel in multiple spectral bands in the spatial distribution of spectral emissivity, calculate the average emissivity of the neighboring bands to obtain the background emissivity matrix. Then, compare and analyze the spatial distribution of spectral emissivity and the background emissivity matrix to obtain the emissivity anomaly sequence.

[0069] Schematic, the spatial distribution of spectral emissivity can be represented as ,in, Represents spatial pixel coordinates, Indicates spectral band, Represents spatial pixels In the spectral band The emissivity value is calculated using the sliding neighborhood window method in the intelligent monitoring terminal for metal component quality. The calculation expression is as follows: ,in, Represents spatial pixels In the spectral band Mean emissivity of the neighboring band at that location Indicates the number of neighboring spectral bands. Indicates spectral band The intelligent monitoring terminal for the quality of metal components can adjust the neighborhood band set based on the number and distribution of spectral bands. The range is defined by combining the average emissivity of all spatial pixels in their neighborhood bands across each spectral band to obtain the background emissivity matrix. The intelligent monitoring terminal for the quality of metal components calculates the difference between the spatial distribution of spectral emissivity and the background emissivity matrix to obtain a sequence of emissivity anomalies. The calculation expression is as follows: ,in, Represents spatial pixels In the spectral band The following are emissivity anomalies; the emissivity anomaly sequence is all... The set of emissivity outliers can reflect the degree of deviation of the spatial distribution of spectral emissivity from the background.

[0070] S42. Input the spatial distribution of spectral emissivity into the double-layer oxide film model for inverse data fitting and solution to obtain the equivalent roughness parameters and oxide layer thickness distribution; the double-layer oxide film model is obtained by fitting the bottom metal substrate and the upper oxide film layer through a nonlinear least squares fitting algorithm.

[0071] Furthermore, the intelligent monitoring terminal for the quality of metal components employs a nonlinear least squares fitting algorithm to construct the fitting objective function, expressed as follows: ,in Represents the equivalent roughness parameter. This represents the effective emissivity considering the equivalent roughness parameter. The expression for the effective emissivity of the bilayer oxide film model is: In the formula, Represents spatial pixels In the spectral band Oxide layer thickness Effective emissivity at the following levels This represents the effective reflectance under the corresponding conditions. The expression for effective reflectance is: In the formula, This indicates the upper surface of the oxide layer in the spectral band. The reflection coefficient at the bottom, This indicates the interface between the oxide layer and the metal substrate in the spectral band. The reflection coefficient at the bottom, This indicates the wavenumber of light in the oxide layer. Indicates the oxide layer in the spectral band Complex refractive index at the following levels Represents spatial pixels The oxide layer thickness at each location is determined by solving the above optimization problem. The intelligent monitoring terminal for metal component quality obtains the oxide layer thickness for each spatial pixel. and equivalent roughness parameters .

[0072] S43. Mark the regions in the emissivity anomaly sequence whose emissivity anomalies exceed a preset threshold to obtain a set of candidate defect regions.

[0073] Specifically, an emissivity anomaly exceeding a preset threshold indicates that the spectral emissivity of the region deviates from the background level, potentially indicating surface defects. The preset threshold can be expressed as: The intelligent monitoring terminal for the quality of metal components can be adjusted according to the material properties of the metal components and the testing requirements. The value range of [value]. The intelligent monitoring terminal for metal component quality performs connectivity analysis on pixels with abnormal marked values. Connectivity analysis can use the eight-neighbor connectivity criterion to divide interconnected abnormal pixels into candidate defect regions. All candidate defect regions are then integrated to obtain a set of candidate defect regions. ,in, This indicates the number of candidate defect regions.

[0074] S44. Extract the average emissivity anomaly values ​​of each pixel in each region of the candidate defect region set in each spectral band, and input the average emissivity anomaly values ​​into the cavity radiation enhancement effect model for inversion to obtain the defect geometric morphology parameters of each region in the candidate defect region set; wherein, the defect geometric morphology parameters include at least the crack depth and opening width, the pit radius and pit depth, and the equivalent diameter of the pore.

[0075] For example, the intelligent monitoring terminal for the quality of metal components calculates the average emissivity anomaly value of each pixel in each spectral band within each candidate defect region, and the calculation expression is as follows: ,in, Indicates the first The candidate defect regions are in the spectral band The average emissivity anomaly is below. Indicates the first The number of pixels within each candidate defect region. Indicates the first One candidate defect region. The intelligent monitoring terminal for metal component quality converts average emissivity anomalies into effective emissivity deviations, i.e. Inverse inversion is performed using a cavity radiation enhancement effect model, with the inversion objective function being: By solving this optimization problem, the intelligent monitoring terminal for the quality of metal components obtains the first... Defect geometric morphology parameters of each candidate defect region The defect geometry parameters include at least the crack depth and opening width, the pit radius and pit depth, and the equivalent diameter of the pores.

[0076] S45. Based on the equivalent roughness parameter, oxide layer thickness distribution and defect geometry parameter, the spatial distribution of spectral emissivity is converted into the surface micro-topological height field distribution.

[0077] Optionally, the surface micro-topological height field distribution comprehensively reflects the surface roughness, oxide layer thickness, and defect morphology of the metal component, and its conversion formula is as follows: ,in, Represents spatial pixels The surface microtopological height value at that location, This represents the height component contributed by the equivalent roughness parameter. This represents the height component contributed by the oxide layer thickness distribution. This represents the height component contributed by the defect geometry parameters. ,in, This represents the roughness-height conversion factor, and the intelligent monitoring terminal for the quality of metal components can determine its value by measuring standard samples. ,in, This represents the oxide layer thickness-to-height conversion factor, which is related to the material density of the oxide layer; ,in, Indicates by the first Defect geometric morphology parameters of each candidate defect region The calculated defect height components can be expressed as a function of crack depth for crack defects, and as a function of the corresponding depth parameter for pit and porosity defects. The intelligent monitoring terminal for metal component quality superimposes the three height components to obtain the surface micro-topological height value of each spatial pixel. By combining the surface micro-topological height values ​​of all spatial pixels, the final surface micro-topological height field distribution is formed.

[0078] In one embodiment, a multi-channel physical feature map is constructed by fusing the real temperature field, oxide layer thickness distribution, and surface micro-topological height field distribution. This multi-channel physical feature map is then input into a pre-trained defect determination network for real-time inference, generating intelligent defect determination results for the target component, including:

[0079] S51. Based on the surface micro-topological height field distribution, extract the local surface height, local surface normal angle component, local surface root mean square roughness, local surface average curvature and Gaussian curvature to obtain three-dimensional geometric features.

[0080] Schematic, the local surface height is directly obtained from the surface micro-topological height field distribution, i.e. ,in, Represents spatial pixels The local surface height at a given location. The local surface normal angle component is obtained by calculating the spatial gradient of the surface micro-topological height field distribution; specifically, the gradient vector is calculated. ,in, and These represent the distribution of the surface microtopological height field at... direction and The partial derivative of the direction converts the gradient vector into angular components, i.e. and ,in, This represents the angle between the local surface normal and the z-axis. This represents the angle between the projection of the local surface normal onto the xy-plane and the x-axis. The expression for calculating the root mean square roughness of the local surface is: ,in, Represents spatial pixels The local root mean square roughness of the surface at that location. This indicates the number of pixels within a local sliding window. Represented in spatial pixels A local sliding window centered on the center. This represents the average height of the surface microtopology within the local sliding window. The local surface mean curvature and Gaussian curvature are calculated using second-order partial derivatives, and their expressions are as follows: and ,in, , , , , Let the first and second partial derivatives of the surface microtopological height field distribution be represented, respectively. This represents the average curvature of a local surface. This represents the Gaussian curvature of the local surface. The intelligent monitoring terminal for the quality of metal components combines all the features extracted above to obtain the three-dimensional geometric features.

[0081] S52. Extract the local temperature anomaly deviation and temperature spatial gradient vector in the real temperature field to obtain thermal characteristics, and extract the local emissivity anomaly data and emissivity spectral slope parameters in the spatial distribution of spectral emissivity, as well as the oxide layer thickness gradient in the oxide layer thickness distribution to obtain material characteristics.

[0082] Optionally, the calculation expression for the local temperature anomaly deviation is as follows: ,in, Represents spatial pixels Local temperature anomaly deviation at the location, For a real temperature field, Represented in spatial pixels The intelligent monitoring terminal for metal component quality can select local neighborhoods of different sizes for calculation, based on the true average temperature within a local neighborhood centered on the component. The expression for the temperature space gradient vector is: ,in, and Let and represent the partial derivatives of the real temperature field in the x and y directions, respectively. This vector reflects the spatial rate and direction of change of the real temperature field. The intelligent monitoring terminal for the quality of metal components can use the Sobel or Prewitt operator to calculate the spatial temperature gradient vector. The spatial distribution of spectral emissivity can be expressed as... The calculation expression for local emissivity anomaly data is as follows: ,in, Represents spatial pixels Local emissivity anomaly data at the location, Indicates the number of spectral bands. Represents spatial pixels In the spectral band The background emissivity. The expression for calculating the emissivity spectral slope parameter is: ,in, Represents spatial pixels The emissivity spectrum slope parameter at that location, and This represents two adjacent spectral bands. The oxide layer thickness distribution can be represented as... The expression for the oxide layer thickness gradient is: ,in, and These represent the partial derivatives of the oxide layer thickness distribution in the x and y directions, respectively.

[0083] S53. Tensor-based merging of three-dimensional geometric features, thermal features, and material features is performed to construct a multi-channel physical feature map.

[0084] For example, the expression for tensor concatenation and merging is: ,in, Represents a multi-channel physical feature map. This indicates a tensor splicing operation. This indicates splicing along the channel dimension. For three-dimensional geometric feature tensors, Indicates the image height. Indicates the image width. The number of channels represents the three-dimensional geometric features, with each channel corresponding to the local surface height, local surface normal angular component, local surface root mean square roughness, local surface mean curvature, and Gaussian curvature, respectively. For thermal characteristic tensor, The number of channels represents the thermal characteristics, with each channel corresponding to one of the two components of the local temperature anomaly deviation and the temperature spatial gradient vector. For material characteristic tensor, The number of channels represents the material characteristics, with each channel corresponding to two components: local emissivity anomaly data, emissivity spectral slope parameter, and oxide layer thickness gradient.

[0085] S54. Input the multi-channel physical feature map into the defect determination network based on the encoder and decoder architecture. The encoder performs continuous forward convolution and pooling downsampling operations to obtain a multi-scale deep physical feature map representing the target state. The decoder restores the spatial resolution of the multi-scale deep physical feature map through upsampling and skip connections to obtain the intelligent defect determination result. The intelligent defect determination result includes the defect existence probability value and classification label corresponding to each pixel. The defect determination network based on the encoder and decoder architecture includes a physical feature attention guidance module. The physical feature attention guidance module adjusts the distribution weight of the deep attention channel according to the local surface average curvature, local temperature anomaly deviation and local emissivity anomaly data.

[0086] Specifically, the expression for the encoder's forward convolution operation is as follows: ,in, Indicates the encoder's first... The output feature map of the layer, The activation function can be either ReLU or LeakyReLU, and the intelligent monitoring terminal for metal component quality can be selected. Indicates the encoder's first... The convolution kernel of the layer, This represents the convolution operation. Indicates the encoder's first... The input feature map of the layer, Indicates the encoder's first... The layer's bias term. The expression for the pooling downsampling operation is: ,in, This represents the feature map after pooling downsampling. This indicates a pooling operation; the intelligent monitoring terminal for metal component quality can select either max pooling or average pooling. Indicates the pooling window size. This represents the pooling stride. After multiple rounds of convolution and pooling downsampling operations, a multi-scale deep physical feature map is obtained. ,in, This indicates the number of layers in the encoder. The expression for the upsampling operation in the decoder is: ,in, Indicates the decoder's first Upsampled feature map of layer, This indicates the transposed convolution upsampling operation. Indicates the size of the transposed convolution kernel. Indicates the stride of the transposed convolution. Indicates the decoder's first The input feature map of the layer. The expression for the skip connection operation is: This is used to concatenate the multi-scale depth physical feature map of the corresponding layer of the encoder with the upsampled feature map of the decoder, preserving detailed features. The defect determination network based on the encoder and decoder architecture includes a physical feature attention guidance module, whose weight adjustment expression is: ,in, Indicates the first Each channel in spatial pixels Attention weights at the location Indicates the attention coefficient. The attention feature function is expressed as follows: , , , This represents the weighting coefficient; the intelligent monitoring terminal for metal component quality can adjust its value through training. This represents the average curvature of a local surface. This indicates an abnormal deviation in local temperature. This indicates abnormal local emissivity data.

[0087] In one embodiment, a multi-channel physical feature map is constructed by fusing the real temperature field, oxide layer thickness distribution, and surface micro-topological height field distribution. This multi-channel physical feature map is then input into a pre-trained defect determination network for real-time inference to generate intelligent defect determination results for the target component. The method further includes:

[0088] S61. Based on the probability value of the existence of defects, the independent closed connected defect regions are extracted from the intelligent defect judgment results to obtain the binary analysis mask tensor. The binary analysis mask tensor is then mapped to the surface micro-topological height field distribution to obtain the absolute depth extremum, surface projected area, and concave structure volume of the connected defect regions.

[0089] Indicatively, the intelligent monitoring terminal for metal component quality employs a connected component analysis algorithm to determine the connectivity of non-zero pixels (defect candidate pixels) in the binary analysis mask tensor, thus extracting independent, closed, connected defect regions. Each connected defect region corresponds to an independent defect candidate target. Furthermore, the intelligent monitoring terminal spatially maps the binary analysis mask tensor to the surface micro-topological height field distribution using a formula... Calculate the absolute depth extremum of the connected defect region, where, Indicates the connectivity defect region in the pixel The absolute depth value at that location. This represents the height value of the surface microtopological height field distribution at that pixel. This represents the average surface micro-topological height of the local neighborhood of the pixel. The intelligent monitoring terminal for metal component quality calculates the absolute depth extremum of each connected defect region by traversing all pixels within that region. Surface projected area ,in, This represents a single connected defect region.

[0090] S62. For connected defect regions whose classification labels belong to the crack category in the intelligent defect judgment results, the core backbone centerline vector of the connected defect region is extracted by using the central skeletonization tracking algorithm. Based on the core backbone centerline vector, the total crack length, average crack width and extreme crack width of the connected defect region are obtained by scanning the normal cross-section span.

[0091] Specifically, the center skeletonization tracking algorithm is based on the distance transformation principle, that is, performing a distance transformation on the binary analysis mask tensor of the crack-type connected defect region. The distance transformation formula is as follows: ,in, Represents pixels The minimum distance to the boundary of the crack-connected defect region. This represents the boundary pixel coordinates of the crack-connected defect region. This indicates the type of crack and the connected defect region. The intelligent monitoring terminal for metal component quality uses iterative corrosion operations to gradually peel away the edge pixels of the crack-connected defect region, retaining the core backbone pixels. This process is repeated until no edge pixels can be peeled away, resulting in the centerline vector of the crack core backbone. ,in, Indicates the number of pixels in the center line. Indicates the center line number The coordinates of each pixel. Based on the centerline vector, the intelligent monitoring terminal for metal component quality calculates the total crack length, i.e. .

[0092] Optionally, the intelligent monitoring terminal for the quality of metal components adopts a normal cross-sectional span scanning method, sampling the span at each pixel along a direction perpendicular to the centerline, using the following sampling formula: ,in, Indicates the angle between the scanning direction and the normal of the center line. and These represent the boundary pixel coordinates along the scanning direction. By traversing all sampling points, the average crack width is obtained. and crack extreme width .

[0093] S63. Cross-search to extract the spectral emissivity waveform data sequence mapped by the connected defect region. When the spectral emissivity waveform data sequence shows a sudden drop in data value and simultaneously meets the preset spectral slope abrupt change rule, the connected defect region is forcibly marked as an oxide layer peeling defect type.

[0094] For example, the spectral emissivity waveform data sequence is represented as follows: ,in, Representing spectral channels of different wavelengths, This represents the spectral emissivity value at the corresponding wavelength. The intelligent monitoring terminal for the quality of metal components calculates the first-order difference sequence of the spectral emissivity waveform using the following formula: ,in, Indicates the first The spectral channel and the first The emissivity difference of each spectral channel, and This represents two adjacent spectral channels. The intelligent monitoring terminal for metal component quality is pre-set with a spectral slope abrupt change rule; when there are two or more consecutive adjacent spectral channels with emissivity differences that satisfy... And the corresponding emissivity value At that time, it was determined that the spectral emissivity waveform data sequence showed a sudden drop, in which Indicates the threshold for abrupt changes in spectral slope. This represents the threshold for a sudden drop in spectral emissivity. When a connected defect region simultaneously satisfies a sudden drop in spectral emissivity waveform data sequence and conforms to a preset rule for abrupt changes in spectral slope, the intelligent monitoring terminal for metal component quality will forcibly mark the connected defect region as an oxide layer peeling defect type.

[0095] In one embodiment, based on the joint inversion objective function, the temperature field variables and spectral emissivity field variables are iteratively updated alternately using the alternating direction multiplier method until convergence is reached, outputting the true spatial distribution of the temperature field and spectral emissivity. The method further includes:

[0096] S71. Perform spatial downsampling on the thermal radiation component data cube to obtain a downsampled thermal radiation component data cube. Then, perform alternating direction multiplier method on the downsampled thermal radiation component data cube to iteratively update the temperature field variable and the spectral emissivity field variable until the convergence condition is met, thus obtaining the coarse-grained temperature field and the coarse-grained spectral emissivity field.

[0097] Indicatively, the intelligent monitoring terminal for metal component quality performs spatial downsampling on the thermal radiation component data cube. The aim is to reduce data dimensionality and iterative computation while preserving core radiation characteristics. The intelligent monitoring terminal for metal component quality performs spatial downsampling on the thermal radiation component data cube. Perform downsampling, where, Indicates spectral band, Represents pixel coordinates, This represents the thermal radiation intensity of the corresponding wavelength and pixel. The downsampling operation uses a fixed-step sampling mode, and the sampling formula is... ,in, This represents the downsampled thermal radiation component data. Indicates the downsampling step size. , These represent the coordinates of the pixels after downsampling. This indicates a round-down operation. , This represents the coordinates of the corresponding sampled pixel in the original data cube. This represents the offset within the sampling window. Based on the downsampled data cube, a complete iterative process using the alternating direction multiplier method is executed, alternately updating the temperature field variable and the spectral emissivity field variable until the coarse-grained convergence condition is met, thus obtaining the coarse-grained temperature field. and coarse-grained spectral emissivity field ,in, Indicates the pixel coordinates after downsampling. Indicates the spectral band.

[0098] S72. The coarse-grained temperature field and coarse-grained spectral emissivity field are upsampled to the original image resolution using a bicubic interpolation method to obtain an initial thermal radiation image. The initial thermal radiation image is used to set the initial values ​​of the auxiliary variables for iterative solution using the alternating direction multiplier method.

[0099] Optionally, the core formula for bicubic interpolation is: and ,in, This represents the initial temperature value obtained after upsampling. This represents the initial spectral emissivity value obtained after upsampling. This represents the bicubic interpolation kernel function, and its expression is: , This represents the temperature value of the corresponding offset pixel in the coarse-grained temperature field. This represents the emissivity value of the corresponding offset pixel in the coarse-grained spectral emissivity field. , , This represents the pixel interpolation offset. This represents the coarse-grained pixel coordinates. The intelligent monitoring terminal for metal component quality uses the above interpolation formula to accurately map the coarse-grained temperature field and coarse-grained spectral emissivity field to the original image resolution. The resulting initial thermal radiation image is directly used as the initial values ​​of auxiliary variables in the alternating direction multiplier method, including temperature and emissivity auxiliary variables, ensuring rapid convergence of subsequent iterative solutions.

[0100] S73. During the iterative process of the alternating direction multiplier method, the regularization parameter in the joint inversion objective function is dynamically adjusted according to the residual decay rate of the current iteration step. In the early stage of the alternating direction multiplier method iteration, the regularization parameter is dominated by the data fitting term to quickly approximate the neighborhood of the true value.

[0101] Optionally, the residual calculation formula is as follows: ,in, Indicates the first The residual of the next iteration This represents the product of the total number of pixels involved in the residual calculation and the number of spectral bands. This represents the measured radiation intensity within the thermal radiation component data cube. Indicates the first In the next iteration, based on the current temperature field and spectral emissivity field The predicted radiation intensity is calculated. Subsequently, the intelligent monitoring terminal for the quality of metal components calculates the residual decay rate. ,in Indicates the first The residual of the next iteration This represents the residual decay rate. Based on the residual decay rate, the intelligent monitoring terminal for the quality of metal components dynamically updates the regularization parameter, using the following update formula: ,in, Indicates the first The regularization parameter for the next iteration Indicates the first The regularization parameter for the next iteration This represents the minimum threshold for the regularization parameter, preventing excessive regularization from causing fitting bias. In the initial stages of the alternating direction multiplier method iteration, the residual decay rate... Larger, regularization parameter The value is relatively small, with the data fitting term dominating, ensuring that the iterative process quickly approximates the neighborhood of the true value; as the iteration progresses, the residual decay rate gradually decreases, and the regularization parameter... Gradually increase the regularization constraint to strengthen it and suppress overfitting.

[0102] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0103] Based on the same inventive concept, this application also provides a machine vision-based intelligent monitoring system for metal component quality, which implements the aforementioned machine vision-based intelligent monitoring method for metal component quality. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more machine vision-based intelligent monitoring system embodiments provided below can be found in the limitations of the machine vision-based intelligent monitoring method for metal component quality described above, and will not be repeated here.

[0104] In one exemplary embodiment, such as Figure 2 As shown, a machine vision-based intelligent monitoring system 200 for the quality of metal components is provided, comprising:

[0105] The data acquisition module 201 is used to acquire raw multispectral polarization image data with different exposure times synchronously acquired by the multi-channel imaging system under the environment dominated by thermal radiation on the surface of high-temperature metal components, and to perform radiometric calibration correction, polarization calibration correction and multi-exposure fusion processing on the raw multispectral polarization image data to obtain a multispectral polarization data cube; wherein, the multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band under at least four spectral bands; the multi-channel imaging system is constructed by combining multispectral spectral dispersion and polarization analysis;

[0106] The polarization-guided component separation module 202 is used to perform Stokes vector calculations on each pixel in the multispectral polarization data cube based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity, to obtain the linear polarization degree value of each pixel, and based on the polarization degree-reflection component ratio mapping table of the micro-surface element bidirectional reflection distribution function model, to perform reflection separation pixel by pixel on the multispectral polarization data cube through the linear polarization degree value of each pixel, to obtain the thermal radiation component data cube.

[0107] Temperature measurement module 203 is used to perform global gray-level temperature estimation on the thermal radiation component data cube based on the dual-band colorimetric temperature measurement algorithm to obtain the initial temperature field; the dual-band colorimetric temperature measurement algorithm estimates the temperature of the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube respectively.

[0108] Inversion module 204 is used to perform alternating iterative optimization and solution of thermal radiation component data cube based on a preset temperature-emissivity joint inversion framework, with the initial temperature field as the starting point, to generate the real temperature field and spectral emissivity spatial distribution.

[0109] Feature module 205 is used to perform inverse physical parameter mapping on the spatial distribution of spectral emissivity based on the double oxide film model and the cavity radiation enhancement effect model to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution.

[0110] The quality module 206 is used to fuse the real temperature field, oxide layer thickness distribution and surface micro-topology height field distribution to construct a multi-channel physical feature map, and input the multi-channel physical feature map into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component.

[0111] In one embodiment, the inversion module 204 is further configured to:

[0112] A data fidelity term is constructed based on the forward coupling mapping function of the micro-facet bidirectional reflection distribution function model, the thermal radiation component data cube, and the initial temperature field; the data fidelity term is used to characterize the difference between the measured radiation flux and the predicted value.

[0113] Based on the emissivity spectral smoothing constraint, the roughness-emissivity correlation constraint, and the oxide layer thickness spatial smoothing constraint, a joint inversion objective function is generated by combining the data fidelity term.

[0114] Based on the joint inversion objective function, the temperature field variables and spectral emissivity field variables are updated alternately and iteratively through the alternating direction multiplier method until the convergence condition is met, and the true spatial distribution of temperature field and spectral emissivity is output.

[0115] In one embodiment, the inversion module 204 is further configured to:

[0116] Based on auxiliary variables, the joint inversion problem corresponding to the joint inversion objective function is decomposed into auxiliary variable update subproblems, emissivity and temperature update subproblems, and Lagrange multiplier update subproblems;

[0117] With the temperature field variable and spectral emissivity field variable fixed in the current iteration step, the auxiliary variable update subproblem is solved by projection solution algorithm based on the total variational regularization term, and the current intermediate solution is smoothed and denoised to obtain the updated auxiliary variables;

[0118] With the updated auxiliary variables fixed, the temperature field variables and spectral emissivity field variables are updated alternately based on the gradient descent algorithm and the partial derivative matrix through the Jacobian matrix coupling mechanism to solve the emissivity and temperature update subproblems, and the updated temperature field variables and updated spectral emissivity field variables are obtained.

[0119] Based on the updated temperature field variables and the updated spectral emissivity field variables, update the Lagrange multiplier matrix to solve the Lagrange multiplier update subproblem;

[0120] The auxiliary variable update subproblem, emissivity and temperature update subproblem, and Lagrange multiplier update subproblem are executed iteratively until the relative change of the joint inversion objective function is less than the preset convergence threshold. The obtained temperature field variable is then mapped to the real temperature field, and the obtained spectral emissivity field variable is mapped to the spectral emissivity spatial distribution.

[0121] In one embodiment, the feature module 205 is further configured to:

[0122] Based on the emissivity sequence values ​​of each spatial pixel in multiple spectral bands in the spatial distribution of spectral emissivity, the mean emissivity of the neighboring bands is calculated to obtain the background emissivity matrix. The spatial distribution of spectral emissivity and the background emissivity matrix are compared and analyzed to obtain the emissivity anomaly sequence.

[0123] The spatial distribution of spectral emissivity is input into the bilayer oxide film model for inverse data fitting and solution to obtain the equivalent roughness parameters and oxide layer thickness distribution; the bilayer oxide film model is obtained by fitting the bottom metal substrate and the upper oxide film layer using a nonlinear least squares fitting algorithm.

[0124] Regions in the emissivity anomaly sequence whose emissivity anomalies exceed a preset threshold are marked to obtain a set of candidate defect regions.

[0125] The average emissivity anomalies of each pixel in each region of the candidate defect region set in each spectral band are extracted, and the average emissivity anomalies are input into the cavity radiation enhancement effect model for inversion to obtain the defect geometric morphology parameters of each region in the candidate defect region set; wherein, the defect geometric morphology parameters include at least the crack depth and opening width, the pit radius and pit depth, and the equivalent diameter of the pore;

[0126] Based on the equivalent roughness parameter, oxide layer thickness distribution, and defect geometry parameter, the spatial distribution of spectral emissivity is converted into the surface micro-topological height field distribution.

[0127] In one embodiment, the quality module 206 is further configured to:

[0128] Based on the surface micro-topological height field distribution, local surface height, local surface normal angle component, local surface root mean square roughness, local surface average curvature and Gaussian curvature are extracted to obtain three-dimensional geometric features;

[0129] Local temperature anomaly deviation and temperature spatial gradient vector in the real temperature field are extracted to obtain thermal characteristics. Local emissivity anomaly data and emissivity spectral slope parameters in the spatial distribution of spectral emissivity, as well as oxide layer thickness gradient in the oxide layer thickness distribution, are extracted to obtain material characteristics.

[0130] By tensor-based splicing and merging of three-dimensional geometric features, thermal features, and material features, a multi-channel physical feature map is constructed.

[0131] Multi-channel physical feature maps are input into a defect determination network based on an encoder-decoder architecture. The encoder performs continuous forward convolution and pooling downsampling operations to obtain a multi-scale deep physical feature map representing the target state. The decoder then restores the spatial resolution of the multi-scale deep physical feature map through upsampling and skip connections, resulting in a defect intelligent determination result. The defect intelligent determination result includes the defect presence probability value and classification label for each pixel. The defect determination network based on the encoder-decoder architecture includes a physical feature attention guidance module, which adjusts the distribution weights of deep attention channels based on local surface average curvature, local temperature anomaly deviation, and local emissivity anomaly data.

[0132] In one embodiment, the quality module 206 is further configured to:

[0133] Based on the probability value of the existence of defects, the independent closed connected defect regions are extracted from the intelligent defect judgment results to obtain the binary analysis mask tensor. The binary analysis mask tensor is then mapped to the surface micro-topological height field distribution to obtain the absolute depth extremum, surface projected area, and concave structure volume of the connected defect regions.

[0134] For connected defect regions whose classification labels belong to crack type morphology in the intelligent defect judgment results, the core backbone centerline vector of the connected defect region is extracted by the central skeletonization tracking algorithm. Based on the core backbone centerline vector, the total crack length, average crack width and extreme crack width of the connected defect region are obtained by scanning the normal cross section span.

[0135] Cross-referencing is used to extract the spectral emissivity waveform data sequence mapped to the connected defect region. When the spectral emissivity waveform data sequence shows a sudden drop in data value and simultaneously conforms to the preset spectral slope abrupt change rule, the connected defect region is forcibly marked as an oxide layer peeling defect type.

[0136] In one embodiment, the inversion module 204 is further configured to:

[0137] Spatial downsampling is performed on the thermal radiation component data cube to obtain a downsampled thermal radiation component data cube. The temperature field variable and the spectral emissivity field variable are then iteratively updated using the alternating direction multiplier method on the downsampled thermal radiation component data cube until the convergence condition is met, resulting in a coarse-grained temperature field and a coarse-grained spectral emissivity field.

[0138] The coarse-grained temperature field and coarse-grained spectral emissivity field are upsampled to the original image resolution using a bicubic interpolation method to obtain an initial thermal radiation image; the initial thermal radiation image is used to set the initial values ​​of the auxiliary variables for the iterative solution of the alternating direction multiplier method.

[0139] During the iterative process of the alternating direction multiplier method, the regularization parameter in the joint inversion objective function is dynamically adjusted according to the residual decay rate of the current iteration step. In the early stage of the alternating direction multiplier method iteration, the regularization parameter is dominated by the data fitting term to quickly approximate the neighborhood of the true value.

[0140] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0141] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0142] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0143] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A machine vision-based intelligent quality monitoring method for metal components, characterized in that, The method includes: The system acquires raw multispectral polarization image data of different exposure durations simultaneously collected by a multi-channel imaging system under a high-temperature metal component surface thermal radiation-dominated environment. The raw multispectral polarization image data undergoes radiometric calibration correction, polarization calibration correction, and multi-exposure fusion processing to obtain a multispectral polarization data cube. The multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band in at least four spectral bands. The multi-channel imaging system is constructed by combining multispectral spectral dispersion and polarization analysis. For each pixel in the multispectral polarization data cube, Stokes vector calculation is performed based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity to obtain the linear polarization degree value of each pixel. Based on the polarization degree-reflection component ratio mapping table of the micro-facet bidirectional reflection distribution function model, the reflection of the multispectral polarization data cube is separated pixel by pixel through the linear polarization degree value of each pixel to obtain the thermal radiation component data cube. The initial temperature field is obtained by performing global gray-level temperature estimation on the thermal radiation component data cube based on the dual-band colorimetric temperature measurement algorithm; the dual-band colorimetric temperature measurement algorithm estimates the temperature of the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube respectively. Based on the preset temperature-emissivity joint inversion framework, with the initial temperature field as the starting point, the thermal radiation component data cube is iteratively optimized and solved to generate the real temperature field and spectral emissivity spatial distribution. Based on the double-layer oxide film model and the cavity radiation enhancement effect model, the spatial distribution of the spectral emissivity is subjected to inverse physical parameter mapping to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution. The real temperature field, the oxide layer thickness distribution, and the surface micro-topology height field distribution are integrated to construct a multi-channel physical feature map. The multi-channel physical feature map is then input into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component.

2. The method according to claim 1, characterized in that, The preset temperature-emissivity joint inversion framework, using the initial temperature field as the starting point, iteratively optimizes and solves the thermal radiation component data cube to generate the true temperature field and spectral emissivity spatial distribution, including: A data fidelity term is constructed based on the forward coupling mapping function of the micro-facet bidirectional reflectance distribution function model, the thermal radiation component data cube, and the initial temperature field; the data fidelity term is used to characterize the difference between the measured radiation flux and the predicted value. Based on the emissivity spectral smoothing constraint, the roughness-emissivity correlation constraint, and the oxide layer thickness spatial smoothing constraint, a joint inversion objective function is generated by combining the data fidelity term. Based on the joint inversion objective function, the temperature field variables and spectral emissivity field variables are alternately and iteratively updated using the alternating direction multiplier method until the convergence condition is met, and the true temperature field and the spatial distribution of spectral emissivity are output. The expression for the joint inversion objective function is as follows: In the formula, For temperature field, For the spectral emissivity field, The forward modeling results of the bidirectional reflectance distribution function of the micro-surface element with respect to the temperature field and spectral emissivity field are given. The thermal radiation component data cube, This represents the weighted L2 norm, which is calculated by weighting the inverse of the measurement noise variance. For emissivity spectrum smoothing constraint terms, For the roughness-emissivity related constraint term, This is the total variational regularization term of the spatial smoothing constraint term for the oxide layer thickness. , and This is the regularization parameter.

3. The method according to claim 2, characterized in that, The process, based on the joint inversion objective function, iteratively updates the temperature field and spectral emissivity field using the alternating direction multiplier method until convergence is achieved, outputting the true temperature field and the spatial distribution of the spectral emissivity, includes: Based on the auxiliary variables, the joint inversion problem corresponding to the joint inversion objective function is decomposed into auxiliary variable update sub-problems, emissivity and temperature update sub-problems, and Lagrange multiplier update sub-problems; With the temperature field variable and the spectral emissivity field variable fixed in the current iteration step, the auxiliary variable update subproblem is solved by projection solution algorithm based on the total variation regularization term, and the current intermediate solution is smoothed and denoised to obtain the updated auxiliary variable. With the updated auxiliary variables fixed, the temperature field variables and the spectral emissivity field variables are alternately updated through the Jacobian matrix coupling mechanism based on the gradient descent algorithm and the partial derivative matrix to solve the emissivity and temperature update subproblem, and the updated temperature field variables and the updated spectral emissivity field variables are obtained. Based on the updated temperature field variables and the updated spectral emissivity field variables, the Lagrange multiplier matrix is ​​updated to solve the Lagrange multiplier update subproblem; The auxiliary variable update subproblem, the emissivity and temperature update subproblem, and the Lagrange multiplier update subproblem are executed iteratively until the relative change of the joint inversion objective function is less than a preset convergence threshold. The obtained temperature field variable is then mapped to the true temperature field, and the obtained spectral emissivity field variable is mapped to the spectral emissivity spatial distribution.

4. The method according to claim 1, characterized in that, The method based on the bilayer oxide film model and the cavity radiation enhancement effect model performs inverse physical parameter mapping on the spatial distribution of the spectral emissivity to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution, including: Based on the emissivity sequence values ​​of each spatial pixel in the spectral emissivity spatial distribution across multiple spectral bands, the mean emissivity of the neighboring bands is calculated to obtain the background emissivity matrix. The spectral emissivity spatial distribution and the background emissivity matrix are then compared and analyzed to obtain an emissivity anomaly sequence. The spatial distribution of spectral emissivity is input into the double-layer oxide film model for inverse data fitting and solution to obtain the equivalent roughness parameter and the oxide layer thickness distribution; the double-layer oxide film model is obtained by fitting the bottom metal substrate and the upper oxide film layer using a nonlinear least squares fitting algorithm; Regions in the emissivity anomaly sequence whose emissivity anomalies exceed a preset threshold are marked to obtain a set of candidate defect regions. The average emissivity anomaly value of each pixel in each region of the candidate defect region set in each spectral band is extracted, and the average emissivity anomaly value is input into the cavity radiation enhancement effect model for inversion to obtain the defect geometric morphology parameters of each region in the candidate defect region set; wherein, the defect geometric morphology parameters include at least the crack depth and opening width, the pit radius and pit depth, and the equivalent diameter of the pore; Based on the equivalent roughness parameter, the oxide layer thickness distribution, and the defect geometry parameter, the spatial distribution of spectral emissivity is converted into the surface micro-topological height field distribution.

5. The method according to claim 1, characterized in that, The process involves fusing the real temperature field, the oxide layer thickness distribution, and the surface micro-topological height field distribution to construct a multi-channel physical feature map. This multi-channel physical feature map is then input into a pre-trained defect determination network for real-time inference, generating intelligent defect determination results for the target component, including: Based on the surface micro-topological height field distribution, local surface height, local surface normal angle component, local surface root mean square roughness, local surface average curvature and Gaussian curvature are extracted to obtain three-dimensional geometric features; The local temperature anomaly deviation and temperature spatial gradient vector in the real temperature field are extracted to obtain thermal characteristics. The local emissivity anomaly data and emissivity spectral slope parameters in the spatial distribution of spectral emissivity, as well as the oxide layer thickness gradient in the oxide layer thickness distribution, are extracted to obtain material characteristics. The three-dimensional geometric features, the thermal features, and the material features are tensor-stitched and merged to construct the multi-channel physical feature map. The multi-channel physical feature map is input into a defect determination network based on an encoder-decoder architecture. The encoder performs continuous forward convolution and pooling downsampling operations to obtain a multi-scale deep physical feature map representing the target state. The decoder then restores the spatial resolution of the multi-scale deep physical feature map through upsampling and skip connections to obtain the intelligent defect determination result. The intelligent defect determination result includes the defect presence probability value and classification label corresponding to each pixel. The defect determination network based on the encoder-decoder architecture includes a physical feature attention guidance module, which adjusts the distribution weights of the deep attention channels based on the local surface average curvature, the local temperature anomaly deviation, and the local emissivity anomaly data.

6. The method according to claim 5, characterized in that, The method of fusing the real temperature field, the oxide layer thickness distribution, and the surface micro-topological height field distribution to construct a multi-channel physical feature map, and inputting the multi-channel physical feature map into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component, further includes: Based on the probability value of the existence of the defect, the independent closed connected defect region is extracted from the intelligent defect judgment result to obtain the binary analysis mask tensor. The binary analysis mask tensor is then mapped to the surface micro-topological height field distribution to obtain the absolute depth extremum, surface projected area, and concave structure volume of the connected defect region. For the connected defect region whose classification label belongs to the crack category morphology in the intelligent defect judgment result, the core backbone centerline vector of the connected defect region is extracted by the center skeletonization tracking algorithm. Based on the core backbone centerline vector, the total crack length, average crack width and extreme crack width of the connected defect region are obtained by scanning the normal cross section span. Cross-retrieval extracts the spectral emissivity waveform data sequence mapped by the connected defect region. When the spectral emissivity waveform data sequence shows a sudden drop in data value and simultaneously conforms to the preset spectral slope abrupt change rule, the connected defect region is forcibly marked as an oxide layer peeling defect type.

7. The method according to claim 3, characterized in that, The process of iteratively updating the temperature field variables and spectral emissivity field variables using the alternating direction multiplier method based on the joint inversion objective function until convergence is reached, and outputting the true temperature field and the spatial distribution of the spectral emissivity, further includes: Spatial downsampling is performed on the thermal radiation component data cube to obtain a downsampled thermal radiation component data cube. The temperature field variable and the spectral emissivity field variable are then iteratively updated using the alternating direction multiplier method on the downsampled thermal radiation component data cube until the convergence condition is met, resulting in a coarse-grained temperature field and a coarse-grained spectral emissivity field. The coarse-grained temperature field and the coarse-grained spectral emissivity field are upsampled to the original image resolution using a bicubic interpolation method to obtain an initial thermal radiation image; the initial thermal radiation image is used to set the initial values ​​of the auxiliary variables solved iteratively by the alternating direction multiplier method; During the alternating direction multiplier method iteration process, the regularization parameter in the joint inversion objective function is dynamically adjusted according to the residual decay rate of the current iteration step; the regularization parameter is dominated by the data fitting term in the early stage of the alternating direction multiplier method iteration to quickly approximate the neighborhood of the true value.

8. A machine vision-based intelligent monitoring system for the quality of metal components, characterized in that, The system includes: The data acquisition module is used to acquire raw multispectral polarization image data of different exposure times synchronously acquired by the multi-channel imaging system under the environment dominated by thermal radiation on the surface of a high-temperature metal component. The module performs radiometric calibration correction, polarization calibration correction, and multi-exposure fusion processing on the raw multispectral polarization image data to obtain a multispectral polarization data cube. The multispectral polarization data cube includes the spatial distribution of radiation intensity in the four polarization directions corresponding to each spectral band under at least four spectral bands. The multi-channel imaging system is constructed by combining multispectral spectral dispersion and polarization analysis. The polarization-guided component separation module is used to perform Stokes vector calculations on each pixel in the multispectral polarization data cube based on the radiation intensity in the four polarization directions in the spatial distribution of radiation intensity, to obtain the linear polarization degree value of each pixel, and based on the polarization degree-reflection component ratio mapping table of the micro-surface element bidirectional reflection distribution function model, to perform reflection separation pixel by pixel on the multispectral polarization data cube through the linear polarization degree value of each pixel, to obtain the thermal radiation component data cube. The temperature measurement module is used to perform global gray-level temperature estimation on the thermal radiation component data cube based on the dual-band colorimetric temperature measurement algorithm to obtain the initial temperature field; the dual-band colorimetric temperature measurement algorithm estimates the temperature for the radiation intensity of the short-wave infrared channel band and the mid-wave infrared channel band in the thermal radiation component data cube respectively. The inversion module is used to perform alternating iterative optimization and solution of the thermal radiation component data cube based on a preset temperature-emissivity joint inversion framework, with the initial temperature field as the starting point, to generate the real temperature field and spectral emissivity spatial distribution. The feature module is used to perform inverse physical parameter mapping on the spatial distribution of the spectral emissivity based on the double oxide film model and the cavity radiation enhancement effect model, so as to obtain the oxide layer thickness distribution and the surface micro-topological height field distribution. The quality module is used to fuse the real temperature field, the oxide layer thickness distribution, and the surface micro-topology height field distribution to construct a multi-channel physical feature map, and input the multi-channel physical feature map into a pre-trained defect judgment network for real-time inference to generate intelligent defect judgment results for the target component.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images