Metal surface quality detection method and system

Abstract

The invention discloses a metal surface quality detection method and system, and relates to the technical field of metal surface quality detection. The method is used for solving the problems of low microdefect detection precision, weak technological parameter relevance and closed-loop control deficiency of the high-reflection surface. The metal surface is irradiated through multi-angle coherent light field serialization, the phase offset of interference fringes is analyzed to generate three-dimensional shape data, and reflection noise interference is restrained. Defect depth gradient is extracted based on dynamic segmentation of process parameter constraint, deposition temperature and pressure deviation are quantified through deconvolution calculation, and process deviation feature distribution is constructed. Finite element simulation is utilized to generate a process-morphology mapping atlas library, cross-domain invariance features are extracted through depth constraint manifold alignment and comparative learning, and a causal correlation model of defect types and process parameters is established. And dynamically adjusting process parameters according to the weight gradient, and reflowing data to update the manifold rule. And high-precision three-dimensional defect detection, process deviation traceability and adaptive parameter optimization are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of metal surface quality detection, and in particular to a method and system for detecting metal surface quality. Background Art

[0002] In contemporary manufacturing, metal materials, as key basic materials, are widely used in high-end equipment fields such as aerospace, precision manufacturing, and the automotive industry. Their surface quality has a direct impact on the service life and reliability of components. As high-end manufacturing develops towards intelligence and high precision, the industry has placed higher requirements on the detection of microscopic defects on metal surfaces. Not only must high-sensitivity identification of tiny defects be achieved, but further exploration of the causes behind the defects is also required to support quality traceability and process optimization. Therefore, the development of metal surface quality detection methods with in-depth process correlation analysis capabilities has become an important supporting technology for improving the level of intelligence in the manufacturing process.

[0003] Existing metal surface inspection methods mainly include visual image-based defect recognition, laser profile scanning, interferometry, and other methods. Although these methods have made certain progress in feature extraction and surface morphology reconstruction, they generally have several technical bottlenecks: First, most methods remain at the level of two-dimensional image recognition and lack in-depth analysis of the three-dimensional structure of defects; second, there is a disconnect between detection and process parameters, making it difficult to achieve causal analysis of defects and manufacturing processes; third, in complex environments, the detection system is not robust enough to light interference and material changes, resulting in fluctuations in defect recognition accuracy; fourth, traditional detection methods lack data closure and feedback capabilities, making it difficult to guide the dynamic adjustment and iterative optimization of process parameters. Therefore, there is an urgent need for an intelligent detection method that integrates three-dimensional morphology analysis, frequency domain modeling, physical simulation, and causal reasoning to achieve a systematic improvement from defect identification to defect cause analysis and process optimization. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method and system for detecting the surface quality of metals, which solve the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A surface quality detection method for metals, comprising the following steps: S1. irradiating the metal surface with a multi-angle coherent light field sequence, collecting interference fringes caused by the defect depth, analyzing the phase offset of the defect area in the interference fringes according to the nonlinear mapping relationship between the metal dielectric properties and the optical path difference, and generating phase domain data containing the three-dimensional morphology of the defect; S2. performing multi-scale frequency domain decomposition and dynamic segmentation of the phase domain data of the three-dimensional morphology of the defect with process parameter constraints, extracting the defect depth gradient distribution, and performing deconvolution calculation based on the metal processing history parameters and the defect depth gradient distribution. , quantify the deposition temperature deviation and pressure fluctuation, and generate the defect topology feature distribution that quantifies the process deviation through multi-scale topological decomposition and defect geometric feature statistics; S3. Construct a physical simulation atlas library of process parameter-defect morphology mapping relationship based on finite element simulation, align the defect topology feature distribution with the simulation data in the depth-constrained manifold space, extract cross-domain invariant features through comparative learning, and output the defect type and its causal correlation weight distribution with the process parameters; S4. Dynamically adjust the process parameters according to the gradient direction of the causal correlation weight distribution, and return the detection data under the new parameters to the simulation atlas library to update the depth-constrained manifold alignment rules.

[0006] Furthermore, based on the nonlinear mapping relationship between the dielectric properties of metal and the optical path difference, the specific process of analyzing the phase offset of the defect area in the interference fringes is as follows: noise suppression and background correction are performed on the interference fringe images collected from multiple angles, and the fringe distortion caused by environmental interference is eliminated by the phase shift encoding method; based on the spatial variation characteristics of the metal dielectric constant, a nonlinear mapping model between the optical path difference and the defect depth is constructed, and the model is converted into a nonlinear integral expression with constrained boundary conditions; the integral expression is numerically solved through multi-scale guided filtering and iterative optimization algorithm to extract the phase offset value.

[0007] Furthermore, the specific process of generating phase domain data containing the three-dimensional morphology of the defect is as follows: the phase offset is restored to a continuous phase distribution through the phase unwrapping algorithm, and the three-dimensional height field of the defect surface is calculated through the three-dimensional coordinate transformation model in combination with the wavelength of the light wave and the incident angle parameters; the height field data is smoothed by surface interpolation to eliminate local distortion caused by noise, and the depth gradient field, local principal curvature distribution and normal vector direction of the defect area are calculated through geometric feature analysis to generate phase domain data containing three-dimensional morphology, gradient features and geometric properties.

[0008] Furthermore, the phase domain data of the defect's three-dimensional morphology is subjected to multi-scale frequency domain decomposition and dynamic segmentation constrained by process parameters to extract the defect depth gradient distribution. The specific process is as follows: multi-scale frequency domain decomposition is performed on the phase domain data, and the base reflection noise and the defect high-frequency signal are separated by adaptive frequency band selection; a dynamic segmentation threshold is generated according to the statistical distribution of metal processing historical parameters, and the segmentation sensitivity is adjusted in combination with the probability density function of the phase gradient amplitude; the depth gradient direction consistency of the segmented defect area is verified, and pseudo-defect signals that deviate from the material stress field direction by more than a preset angle are eliminated.

[0009] Furthermore, based on the deconvolution calculation of the metal processing history parameters and the defect depth gradient distribution, the specific process of quantifying the deposition temperature deviation and pressure fluctuation is as follows: establishing the physical response relationship between the process parameters and the defect depth gradient, and obtaining the temperature deviation and pressure fluctuation by iteratively solving the deconvolution equation under the constraint conditions; verifying the statistical significance of the deconvolution results based on the random sampling method, screening out the defect areas that are strongly correlated with the process parameter deviation, and marking their spatial position and degree of deviation.

[0010] Furthermore, through multi-scale topological decomposition and defect geometric feature statistics, the specific process of generating the defect topological feature distribution for quantifying process deviation is as follows: the defect area is spatially divided, and the hole density distribution, the consistency coefficient between the crack extension direction and the principal stress direction, and the surface curvature mutation characteristics of each sub-area are statistically analyzed; through statistical correlation analysis, a quantitative mapping relationship between the defect geometric characteristics and the process parameter deviation is constructed to generate a feature distribution map reflecting the degree of process deviation.

[0011] Furthermore, a physical simulation atlas library of process parameter-defect morphology mapping relationship is constructed based on finite element simulation, and the specific process of aligning the defect topological feature distribution with the simulation data in the depth-constrained manifold space is as follows: According to the finite element simulation model of multi-physical field coupling, the defect generation process of the metal surface under different process parameter combinations is simulated to generate a simulation data set containing temperature field, stress field distribution and defect morphology; the actual detected defect topological feature distribution and the simulation data are projected into the same manifold space through the depth-constrained manifold embedding algorithm, where the manifold distance calculation introduces the defect depth gradient similarity weight, and the feature distribution is aligned through geodesic optimization.

[0012] Furthermore, cross-domain invariant features are extracted through contrastive learning, and the specific process of outputting the causal association weight distribution of defect types and process parameters is as follows: construct positive and negative sample pairs of actual defect features and simulation data in the manifold space, design a contrast loss function constrained by process parameters, and extract cross-domain invariant features through gradient descent optimization; based on the optimized feature distribution, the association weights between defect types and process parameters are calculated through causal reasoning, where the weight distribution is dynamically updated through deconvolution gradient backpropagation.

[0013] Furthermore, the process parameters are dynamically adjusted according to the gradient direction of the causal weight distribution, and the detection data under the new parameters are fed back to the simulation atlas library. The specific process of updating the manifold alignment rules of the depth constraints is as follows: the process parameter adjustment amount is generated according to the negative gradient direction of the causal weight distribution, and the constraint optimization algorithm is used to ensure that the parameter adjustment conforms to the physical feasible domain; the detection data under the new parameters are added to the simulation atlas library, and the depth constraint weight matrix of the manifold space is optimized through the incremental parameter update method to achieve the coordinated iteration of the detection model and process parameters.

[0014] A surface quality inspection system for metals comprises the following modules: a multi-angle coherent imaging module, a process dynamic quantification module, a simulation manifold alignment module, and a process adjustment module; the multi-angle coherent imaging module is used to sequentially illuminate the metal surface through a multi-angle coherent light field, collect interference fringes caused by the depth of the defect, and analyze the phase offset of the defect area in the interference fringes according to the nonlinear mapping relationship between the dielectric properties of the metal and the optical path difference, thereby generating phase domain data containing the three-dimensional morphology of the defect; the process dynamic quantification module is used to perform multi-scale frequency domain decomposition and dynamic segmentation constrained by process parameters on the phase domain data of the three-dimensional morphology of the defect, extract the defect depth gradient distribution, and calculate the phase domain data containing the three-dimensional morphology of the defect according to the metal processing history parameters and the defect depth gradient distribution. The deconvolution calculation of the cloth is performed to quantify the deposition temperature deviation and pressure fluctuation, and the defect topology feature distribution that quantifies the process deviation is generated through multi-scale topological decomposition and defect geometric feature statistics; the simulation manifold alignment module is used to construct a physical simulation atlas library of process parameter-defect morphology mapping relationship based on finite element simulation, align the defect topology feature distribution with the simulation data in the depth-constrained manifold space, extract cross-domain invariant features through comparative learning, and output the defect type and its causal correlation weight distribution with the process parameters; the process adjustment module is used to dynamically adjust the process parameters according to the gradient direction of the causal correlation weight distribution, and return the detection data under the new parameters to the simulation atlas library to update the depth-constrained manifold alignment rules.

[0015] The present invention has the following beneficial effects: (1) A surface quality inspection method for metals, which can achieve high-precision three-dimensional morphology extraction of metal surface defect areas by introducing a multi-angle coherent light field interference mechanism and a nonlinear phase solution method; combined with multi-scale frequency domain decomposition and dynamic region segmentation under process parameter constraints, not only improves the separability of defect depth gradients, but also realizes the quantitative inversion of deposition temperature deviation and pressure fluctuation based on morphology data for the first time, effectively enhancing the interpretability and process sensitivity of the inspection method. By integrating the process parameter-defect morphology map library generated by finite element simulation, using a depth-constrained manifold alignment mechanism and a contrastive learning algorithm, high-dimensional consistency alignment between the defect topology distribution and the simulation results is achieved. Further, based on the causal association weight distribution, the process parameters are dynamically adjusted, and a closed-loop feedback mechanism for the inspection data is constructed. The simulation alignment rules can be continuously optimized to form an intelligent inspection system with adaptive evolution capabilities, significantly improving the system's defect classification accuracy and process response capabilities.

[0016] (2) A surface quality inspection system for metals that integrates a multi-angle coherent imaging module, a process dynamic quantification module, a simulation manifold alignment module, and a process adjustment module to achieve full process automation from data acquisition to process feedback, adapting to the real-time inspection requirements of high-speed production lines. Through the physical simulation atlas library and the manifold space alignment mechanism, the system integrates actual inspection data and multi-physics field simulation results to enhance the system's generalization ability for complex working conditions. Based on causal association weights and parameter optimization algorithms, it directly drives the dynamic adjustment of process units such as deposition equipment and heat treatment furnaces, reducing manual intervention and promoting the implementation of an intelligent manufacturing closed-loop system.

[0017] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flow chart of a method for detecting surface quality of metal.

[0019] Figure 2 This is a flow chart of a surface quality detection system for metals according to the present invention. DETAILED DESCRIPTION

[0020] The present application solves the core problems of the prior art, such as low micro-defect detection accuracy under high reflection interference, weak correlation between defects and process parameters, and insufficient generalization capability for small sample scenarios, through a surface quality inspection method and system for metals. The following general ideas are formed:

[0021] Through multi-angle coherent light field serialization irradiation and phase domain analysis technology, the metal surface reflection noise is stripped away and the interference fringes are converted into three-dimensional morphology data including defect depth and curvature, breaking through the limitations of traditional two-dimensional image detection.

[0022] By combining process history parameters with multi-scale frequency domain decomposition, a dynamic segmentation and deconvolution model is constructed to directly invert process deviations such as temperature and pressure from defect morphology data, thereby achieving physically explainable analysis of defect causes.

[0023] Based on multi-physics field finite element simulation, a process-morphology mapping atlas library is generated. Through deep-constrained manifold alignment and comparative learning, feature migration between actual inspection data and simulation data is achieved to solve the problem of small sample defect classification.

[0024] The process parameters are dynamically adjusted according to the gradient direction of the causal weight distribution, and the adjusted new detection data is fed back to the simulation library to realize a closed-loop optimization process of detection-simulation-control.

[0025] See also Figure 1 , an embodiment of the present invention provides a technical solution: a surface quality detection method for metals, comprising the following steps: S1. irradiating the metal surface with a multi-angle coherent light field sequence, collecting interference fringes caused by the defect depth, analyzing the phase offset of the defect area in the interference fringes according to the nonlinear mapping relationship between the metal dielectric properties and the optical path difference, and generating phase domain data containing the three-dimensional morphology of the defect; S2. performing multi-scale frequency domain decomposition and dynamic segmentation of the phase domain data of the three-dimensional morphology of the defect with process parameter constraints, extracting the defect depth gradient distribution, and quantifying the deposition according to the deconvolution calculation of the metal processing history parameters and the defect depth gradient distribution. The temperature deviation and pressure fluctuation are accumulated, and the defect topology feature distribution that quantifies the process deviation is generated through multi-scale topological decomposition and defect geometric feature statistics; S3. Based on the finite element simulation, a physical simulation atlas library of the process parameter-defect morphology mapping relationship is constructed, and the defect topology feature distribution and the simulation data are aligned in the depth-constrained manifold space. The cross-domain invariant features are extracted through contrastive learning, and the defect type and its causal correlation weight distribution with the process parameters are output; S4. The process parameters are dynamically adjusted according to the gradient direction of the causal correlation weight distribution, and the detection data under the new parameters are returned to the simulation atlas library to update the depth-constrained manifold alignment rules.

[0026] In this implementation, step S1: irradiating a metal surface with a multi-angle coherent light field and acquiring interference fringes enables high-precision 3D topography detection of the depth and distribution of microscopic defects (such as scratches and dents) on the metal surface. The acquisition of phase domain data provides high-quality raw input for subsequent detailed defect analysis and process backtracking. Coherent light field: Laser illumination with correlated phases produces interference. Interference fringes: Fringe images formed by varying optical path lengths reflect surface elevation. Phase offset: Reflects local surface height variations and is key to 3D reconstruction. Phase domain data: The phase information set at each point can be mapped into a 3D surface model. Step S2: Multi-scale frequency domain decomposition of the phase domain data and dynamic segmentation combined with process parameters accurately identifies defect morphology trends and infers potential process anomalies such as temperature and pressure that cause defects during the manufacturing process. Topological analysis then generates statistically significant spatial distribution features of defects. Multi-scale frequency domain decomposition, such as wavelet transform, extracts morphological details at different scales. Dynamic segmentation: Flexible adjustment of the defect detection range based on the specific process context. Deconvolution: A mathematical deduction method for inferring original process fluctuations from defects. Defect topological feature distribution: A feature set that describes the morphology, structure, and spatial distribution of defects. Step S3: Construct a "process-defect" physical map based on finite element simulation. Align and match the topological features obtained from inspection. Utilize a contrastive learning algorithm to extract common features across different data sources. Ultimately, the defect type and its quantitative causal weights associated with process parameters are output, facilitating source identification and predictive control. Finite element simulation: Numerical methods are used to simulate the generation mechanisms of defects in metal processing. Manifold space alignment: Simulation data and actual inspection data are embedded in the same geometric space, making them comparable. Contrastive learning: Learning the essential data characteristics from "similar-dissimilar" pairs enhances cross-domain adaptability. Causal association weight distribution: Characterizes the causal strength and direction between a process variable and defect type. Step S4: Based on the causal association weights between defects and process parameters, key processes that affect the process are automatically identified and parameter fine-tuning and optimization are performed to improve product consistency. New inspection data is fed back into the simulation model to continuously optimize the alignment mechanism, enabling online model iteration and closed-loop process control. Dynamic adjustment of gradient direction: Optimizes process parameters based on the direction of maximum change in causal weight. Data reflow: Feeds new inspection data back into the model to enhance simulation prediction capabilities. Manifold alignment rule update: Dynamically calibrates mapping relationships to adapt to new process scenarios or new material changes.

[0027] Specifically, based on the nonlinear mapping relationship between the dielectric properties of metal and the optical path difference, the specific process of analyzing the phase offset of the defect area in the interference fringes is as follows: noise suppression and background correction are performed on the interference fringe images collected from multiple angles, and the fringe distortion caused by environmental interference is eliminated through the phase shift encoding method; based on the spatial variation characteristics of the metal dielectric constant, a nonlinear mapping model between the optical path difference and the defect depth is constructed, and the model is converted into a nonlinear integral expression with constrained boundary conditions; the integral expression is numerically solved through multi-scale guided filtering and iterative optimization algorithm to extract the phase offset value.

[0028] In this embodiment, the interference fringe image is usually affected by environmental noise, lighting changes, etc., which affect the accuracy of the image. The purpose of noise suppression is to remove irrelevant noise signals in the image and enhance the recognizability of the interference fringes. Common noise suppression methods include Gaussian filtering, mean filtering, or median filtering. Through these methods, the image can be smoothed, thereby reducing the interference of random noise on the interference fringe analysis. Since the metal surface may be affected by uneven lighting or surface reflection, background correction is to remove the light intensity changes in the image caused by these factors to ensure that the morphology of the interference fringes is related to the surface defects of the metal. Background smoothing methods, such as low-pass filtering, are used to correct the background light intensity and eliminate these effects. Interference image expression model in the noise suppression and background correction stage (frame k): ; Parameter explanation: : interference image intensity at the kth phase shift; : Background light intensity (i.e. static component); : Modulation light intensity (i.e. interference fringe amplitude); : the initial phase distribution to be determined; : The kth phase shift phase amount, usually ; : Noise term, including environmental interference and electronic noise Phase recovery formula (based on least squares method): ; Parameter explanation: : The total number of phase shift frames, the other parameters are the same as above. Nonlinear mapping model construction (metal dielectric properties and optical path difference), the relationship between the optical path difference distribution on the metal surface and the defect depth is expressed as follows: ; Parameter explanation: :Optical path difference caused by metal defects; : local depth of the defect; : The equivalent dielectric function of the defect area varies along the depth; : The reference dielectric constant of the metal defect-free region. Constraints: , ; Parameter explanation: , : Dielectric range allowed by material physics; : The maximum possible defect depth. Discretization of nonlinear integral expression and phase shift inversion discretize the integral model into a linear algebra problem and construct a linear mapping matrix: ; Parameter explanation: : coefficient matrix generated by discrete dielectric differences; : Defect depth vector (unknown); : The vector corresponding to the optical path difference (which can be extracted from the interference fringes) is used to solve the phase offset and introduce the regularized iterative optimization objective function: ; Parameter explanation: : The final phase offset value; : The laser wavelength used; : Regularization weight coefficient; : Gradient smoothing term (used to prevent excessive phase oscillation). Multi-scale guided filtering and iterative solution process, multi-scale guided filter in scale The bootstrap formula is: ; Parameter explanation: :scale The guided filter output phase under ; : Phase estimate of the input; : Mean filter output; :Guided image (grayscale) at scale The local mean under ; : Guide filter response coefficient. Iterative optimization uses a stepwise approximation method: ; Parameter explanation: : the phase solution of the nth iteration; : learning rate / step size coefficient; : Gradient of the loss function.

[0029] Specifically, the specific process of generating phase domain data containing the three-dimensional morphology of the defect is as follows: the phase offset is restored to a continuous phase distribution through the phase unwrapping algorithm, and the three-dimensional height field of the defect surface is calculated through the three-dimensional coordinate transformation model in combination with the wavelength of the light wave and the incident angle parameters; the height field data is smoothed by surface interpolation to eliminate local distortion caused by noise, and the depth gradient field, local principal curvature distribution and normal vector direction of the defect area are calculated through geometric feature analysis to generate phase domain data containing three-dimensional morphology, gradient features and geometric properties.

[0030] In this embodiment, the phase unwrapping purpose is to convert the wrapped phase obtained from the original fringe image into Restore to physically continuous phase distribution . Processing method: Use mass-guided phase unwrapping. 3D coordinate transformation model calculates the surface height field and converts the continuous phase Convert to relative height value on the defect surface , the calculation model is as follows: ; Parameter explanation: : In pixel coordinates The corresponding relative three-dimensional height value; : Actual observed light wavelength (632.8 nm, HeNe laser commonly used in metal interferometry); : The angle between the incident light and the metal surface normal; : The unwrapped continuous phase. Surface interpolation and smoothing, the height field Interpolation (B-spline) and smoothing (bidirectional Gaussian filtering) operations are performed to reduce local distortion caused by stripe blur and image noise. Interpolation method: Construct a polynomial spline surface , maintain the main trend smoothing filter: use anisotropic diffusion filtering or multi-scale wavelet denoising to improve surface continuity. The gradient field calculation of the defect area is based on the smoothed height field , calculate the local depth gradient field : ; Parameter explanation: : depth gradient vector on the two-dimensional plane; :along The rate of change of elevation in the axial direction; :along The local principal curvature is calculated by constructing the Hessian matrix through the second-order derivative to obtain the principal curvatures κ1 and κ2: ; Parameter explanation: : Hessian matrix of the second derivative of local elevation; : principal curvature (maximum, minimum) at (u, v); :Eigenvalue operation of Hessian matrix, solving the two real eigenvalues ​​of the matrix; normal vector direction estimation calculates the normal vector of the metal surface from the gradient field : ,in, ; Parameter explanation: : Unit normal vector, indicating the surface orientation; :Height field is First derivative of direction; : Normalization factor to ensure that the normal vector is a unit vector. Phase domain data construction, finally, the following information is integrated into a structured phase domain dataset: 3D height field: ; Local gradient field: ; Principal curvature: ; Normal vector direction: The phase domain data can be used as input for subsequent metal defect geometry identification, material damage identification, and defect source tracing analysis.

[0031] Specifically, the phase domain data of the defect's three-dimensional morphology is subjected to multi-scale frequency domain decomposition and dynamic segmentation constrained by process parameters to extract the defect depth gradient distribution. The specific process is as follows: multi-scale frequency domain decomposition is performed on the phase domain data, and the base reflection noise and the defect high-frequency signal are separated by adaptive frequency band selection; a dynamic segmentation threshold is generated according to the statistical distribution of metal processing historical parameters, and the segmentation sensitivity is adjusted in combination with the probability density function of the phase gradient amplitude; the depth gradient direction consistency of the segmented defect area is verified, and pseudo-defect signals that deviate from the material stress field direction by more than a preset angle are eliminated.

[0032] In this embodiment, multi-scale frequency domain decomposition is used to separate the background reflection noise (low frequency component) and the defect information (high frequency component) from the phase domain height field. Perform multi-scale frequency domain decomposition (such as wavelet transform or discrete cosine transform), commonly used expressions are as follows: ; Parameter explanation: : Original 3D height field; : frequency domain components of the sth layer (including low-frequency base and high-frequency defects); : total number of decomposition scales; : Undecomposable high-frequency residual; Adaptive band selection automatically divides high and low frequency bands according to frequency domain amplitude statistics, and selects a set of frequency band indexes that meet the high-frequency energy aggregation conditions , used to extract defect signals: ; Parameter explanation: : the average amplitude of the frequency domain components of the sth layer; : standard deviation of the amplitude of the sth layer; : The sensitivity coefficient set by experience is usually 1.5~3; : The set of frequency bands that are judged to contain defect signals. Dynamic segmentation threshold setting (process parameter constraints) combined with metal processing history parameters (such as cutting depth, feed rate, etc.) to form a joint probability distribution , and dynamically set the defect segmentation threshold based on the statistical quantile : ; Parameter explanation: : depth gradient amplitude after band synthesis; : Joint distribution function of metal processing parameters; : p-quantile function of the distribution (such as the 85th percentile); : Dynamic threshold for defect determination under process constraints; Segmentation sensitivity adjustment (based on gradient probability density) to obtain the probability density function (PDF) of the gradient amplitude , and adjust the split sensitivity coefficient by its skewness or kurtosis , update the decision threshold to: ,in, ; Parameter explanation: : probability density function of depth gradient amplitude; : Kurtosis or skewness of the PDF; : Sensitivity amplification factor, enhancing the response to small defects; :Adjusted threshold. Depth gradient direction consistency verification (removal of pseudo defects) For the preliminary segmentation results, calculate the normal vector direction in each defect area Main direction of material stress Angle : ; Eliminate areas that meet the following conditions: ; Parameter explanation: : average normal vector in the defect area; : direction vector of the known stress field of the material; : Angle between vectors; : Deviation from the threshold angle, generally ;like If it exceeds the set range, it is judged as processing noise or surface texture false detection.

[0033] Specifically, based on the deconvolution calculation of metal processing history parameters and defect depth gradient distribution, the specific process of quantifying deposition temperature deviation and pressure fluctuation is as follows: establish the physical response relationship between process parameters and defect depth gradient, and obtain temperature deviation and pressure fluctuation by iteratively solving the deconvolution equation under constraint conditions; verify the statistical significance of the deconvolution results according to the random sampling method, screen out defect areas that are strongly correlated with process parameter deviations, and mark their spatial positions and deviation degrees.

[0034] In this embodiment, the physical response relationship between the defect depth gradient and the processing temperature / pressure disturbance is established; assuming that the defect gradient distribution is the temperature field disturbance and pressure disturbance The convolution superposition response of is used to construct the following physical response model: ;in, : Depth gradient distribution of defect area; : Temperature response kernel function (thermal gradient-morphology response); : pressure response kernel function (stress field-morphology response); *: two-dimensional convolution operator; : local deposition temperature deviation distribution; : Local processing pressure fluctuation distribution; : System error terms (Gaussian noise, measurement error, etc.); iteratively solve the deconvolution problem under constraints to restore the disturbance source field; introduce Tikhonov regularization constraints (to suppress the ill-posedness of the solution) and establish the deconvolution objective function: ;in, :L2 norm, measuring the fitting error; : Regularization coefficient (control smoothness); : Gradient of temperature / pressure disturbance field; use alternating least squares or gradient descent method to perform numerical iteration and output Statistical sampling is used to verify the significance of the deconvolution results and quantify the strength of the correlation between the defect and the disturbance deviation for each defect area. , the set of perturbation values ​​after sampling deconvolution , calculate its Z score with the standard process value: , ; : the disturbance mean of the jth region; : mean value of historical process standard disturbance; : Corresponding standard deviation; if (like , significance level 0.05), the disturbance in the region is considered to be significantly abnormal. The significantly abnormal region is screened out, and its spatial coordinates and deviation degree are marked. The following marking function is used to output the coordinates and abnormality degree: ; will satisfy all Output the center position of the area Maximum deviation amplitude .

[0035] Specifically, the specific process of generating the defect topological feature distribution for quantifying process deviation through multi-scale topological decomposition and defect geometric feature statistics is as follows: the defect area is spatially divided, and the hole density distribution, the consistency coefficient between the crack extension direction and the principal stress direction, and the surface curvature mutation characteristics of each sub-area are statistically analyzed; through statistical correlation analysis, a quantitative mapping relationship between defect geometric features and process parameter deviations is constructed to generate a feature distribution map reflecting the degree of process deviation.

[0036] In this embodiment, the spatial region division and sub-region topological feature extraction are used to divide the defect area into Divided into multiple sub-areas , the following three types of geometric topological features are extracted from each sub-region: hole density distribution (number of defect holes per unit area): ; :Sub-area The pore density; : The number of holes identified in : The consistency coefficient between the crack extension direction and the principal stress direction (cosine of the angle between the directions): ;in, : Directional consistency coefficient; : the main direction vector of the crack in the i-th region (obtained by crack edge fitting); : The principal stress direction vector at the corresponding position of the material (derived from stress field simulation or actual measurement); : The angle between the two. Surface curvature mutation characteristics (ratio of maximum principal curvature to average principal curvature): ; :Curvature mutation index; : the maximum value among the principal curvatures of all points in the i-th region; : The average value of the principal curvature in the i-th region. Construct a mapping model between defect geometric characteristics and process parameter deviations. Use linear regression or multivariate statistical modeling to construct the response relationship between characteristics and deviations: ;in, : Deviation of process parameters inferred for the corresponding sub-area (e.g. temperature deviation or uneven load); : Regression coefficient, reflecting the sensitivity of each feature to the deviation : Residual term, including random disturbance and model error. The mapping relationship can be obtained by least square fitting to obtain the optimal parameters. Generate a topological feature distribution map reflecting the degree of process deviation. Mapping back to its original space coordinates, forming a two-dimensional deviation feature map: ;in, : Indicates location Strength of process deviation; :Indicator function, if Belong to the region 1 if yes, 0 otherwise.

[0037] Specifically, a physical simulation atlas library of process parameter-defect morphology mapping relationship is constructed based on finite element simulation, and the specific process of aligning the defect topological feature distribution with the simulation data in the depth-constrained manifold space is as follows: According to the finite element simulation model of multi-physical field coupling, the defect generation process of the metal surface under different process parameter combinations is simulated to generate a simulation data set containing temperature field, stress field distribution and defect morphology; the actual detected defect topological feature distribution and the simulation data are projected into the same manifold space through the depth-constrained manifold embedding algorithm, where the manifold distance calculation introduces the defect depth gradient similarity weight, and the feature distribution is aligned through geodesic optimization.

[0038] In this implementation, finite element simulation constructs a physical simulation atlas of process parameter defect morphology mapping relationships. First, a multi-physics field coupled finite element simulation model is established. This model simulates the generation process of metal surface defects under different process parameter combinations (such as temperature, stress, pressure, etc.). The simulation process generates a temperature field ( )、Stress Field( ) and defect morphology ( ), where , , are spatial coordinates respectively. Temperature field: Represents the temperature distribution of each point on the metal surface. The temperature distribution is usually affected by the thermal conductivity of the material, the heating method and the boundary conditions. Stress field: It is the stress distribution on the surface and inside of the material, which depends on the applied external force field and temperature field and characterizes the deformation process in metal processing. Defect morphology: The three-dimensional morphology data of the defects generated under different process parameters is the key output of the simulation model. Then, the topological features of the actual detected defects (such as ) and simulation data (such as ) are aligned to the same manifold space. In this process, the manifold distance calculation introduces the defect depth gradient ( ) similarity weight. Manifold distance calculation: Manifold distance The geometric distance between the actual defect data and the simulation data is calculated by the manifold learning algorithm. Depth gradient weight: The depth gradient information of the defect is used as a weighting term to affect the accuracy of the manifold alignment process. Used to measure local changes in defect surfaces and enhance alignment in areas with large depth variations. Geodesic optimization: In manifold space, geodesic optimization It is used to minimize the deviation of defect feature distribution and ensure the alignment of actual defects with simulation data.

[0039] Specifically, the specific process of extracting cross-domain invariant features through contrastive learning and outputting the causal association weight distribution of defect types and process parameters is as follows: constructing positive and negative sample pairs of actual defect features and simulation data in the manifold space, designing a contrast loss function constrained by process parameters, and extracting cross-domain invariant features through gradient descent optimization; based on the optimized feature distribution, the association weights between defect types and process parameters are calculated through causal reasoning, where the weight distribution is dynamically updated through deconvolution gradient backpropagation.

[0040] In this implementation, cross-domain invariant features are extracted through contrastive learning. In the manifold space, contrastive learning methods are used to extract cross-domain invariant features to facilitate the analysis of the relationship between defect types and process parameters. The specific steps are as follows: Construct positive and negative sample pairs First, construct positive and negative sample pairs based on actual defect features and simulation data. Positive sample pairs are composed of similar defect topology features and corresponding process parameters, while negative sample pairs are composed of samples with different defect types or process conditions. Positive sample pairs: ( )When the defect topology characteristics are consistent with the process parameter combination. Negative sample pair: ( ) When the defect topology characteristics are inconsistent with the process parameter combination. Design contrast loss function Design contrast loss function , used to optimize the feature embedding of positive and negative sample pairs. The contrast loss function measures that the distance between similar sample pairs in the embedding space should be small, while the distance between dissimilar sample pairs should be large. Common contrast loss functions are: ;in, is the embedding of the positive sample pair, is the embedding of the negative sample pair, Is the Euclidean distance in the embedding space. Optimize the contrast loss and minimize the contrast loss function through the gradient descent optimization algorithm (Adam optimizer) , thereby extracting cross-domain invariant features. These features can be applied to both actual detection data and simulation data, retaining important information related to process parameters. Causal reasoning and the calculation of the associated weights between defect types and process parameters, the optimized feature distribution can calculate the causal relationship between defect types and process parameters through causal reasoning methods. This process is based on deconvolution gradient back propagation to dynamically update the weight distribution. Causal reasoning model: through the causal reasoning model , infer the causal relationship between process parameters and defect types. The model uses feature representation ( ) to quantify the impact of different process parameters on defect types. Weight calculation: Calculate the causal weight distribution of process parameters ,The weight represents the contribution of each process parameter to the ,defect generation process.

[0041] Specifically, the process parameters are dynamically adjusted according to the gradient direction of the causal weight distribution, and the detection data under the new parameters are fed back to the simulation atlas library. The specific process of updating the manifold alignment rules of the depth constraint is as follows: the process parameter adjustment amount is generated according to the negative gradient direction of the causal weight distribution, and the constraint optimization algorithm is used to ensure that the parameter adjustment conforms to the physical feasible domain; the detection data under the new parameters are added to the simulation atlas library, and the depth constraint weight matrix of the manifold space is optimized through the incremental parameter update method to achieve the coordinated iteration of the detection model and process parameters.

[0042] In this embodiment, the process parameter adjustment amount is generated according to the negative gradient direction of the causal weight distribution. First, the negative gradient direction of the causal weight distribution calculated previously is used to determine the adjustment direction of the process parameter. The negative gradient direction indicates the direction in which the process parameter adjustment is required to reduce the inconsistency between defect generation and process parameters. Negative gradient direction: The negative gradient direction of the causal weight distribution Wcausal is calculated, that is, the adjustment direction of the process parameter. Assume that the process parameter is p and the negative gradient direction is: Where p = (p1, p2, ..., pn) represents the process parameter vector, and Wcausal is the causal weight distribution between the process parameters and the defect. Adjustment value generation: Based on the negative gradient direction, the process parameter adjustment value Δp is generated. This adjustment value represents the correction required for each process parameter. Generally speaking, the calculation form of the adjustment value is: ; where α is the step size coefficient, which is used to control the magnitude of each adjustment. A constrained optimization algorithm is used to ensure that parameter adjustments are within the physical feasible domain. To ensure that the results of process parameter adjustments are still within the physical feasible domain, a constrained optimization algorithm is used to correct the adjustment amount. Feasible domain constraints include physical limitations such as temperature, stress, and pressure. Constrained optimization problem: The constrained optimization problem for process parameter adjustment can be expressed as: ; Where f(p) is the objective function, which can be the error function of the process parameter adjustment amount, and Pfeasible is the physical feasible domain of the process parameter. Optimization process: Through the constrained optimization algorithm (such as the interior point method, quasi-Newton method, etc.), optimize the adjustment amount Δp to ensure that the adjusted process parameters pnew=p+Δp are within the feasible domain. Return the test data under the new parameters to the simulation atlas library. Once the new process parameters are determined within the physical feasible domain, the next step is to conduct production testing and return the test data under the new parameters to the simulation atlas library to further improve the accuracy of the simulation atlas. Test data generation: Use the new process parameters Inspect the actual production process and collect new defect data sets . Data reflow: New detection dataset Add to the existing simulation atlas library to enrich the diversity and accuracy of simulation data. Optimize the depth constraint weight matrix of the manifold space through the incremental parameter update method. After the new detection data flows back to the simulation atlas library, it is necessary to update the manifold alignment rules and the depth constraint weight matrix. The depth constraint weight matrix in the manifold space ( ) is used to accurately align the features between the actual detection data and the simulation data. Incremental update: The depth constraint weight matrix in the manifold space is updated using an incremental parameter update method. The update process can be expressed as: ;in, is the update step size, It is the update amount calculated based on the new data, which represents the new alignment requirement between the defect feature distribution and the simulation data. Optimization goal: By optimizing the depth constraint weight matrix of the manifold space , ensuring that the actual detection data and the simulation data are more accurately aligned in the manifold space, thereby improving the accuracy of the subsequent detection model. Collaborative iteration of detection model and process parameters Finally, based on the updated depth-constrained manifold alignment rules, the detection model and process parameters begin to collaboratively iterate to further accurately adjust the relationship between process parameters and defect characteristics. The iterative process continues until the degree of matching between process parameters and defect generation reaches the predetermined accuracy requirements. Collaborative iterative process: In each iteration, the adjustment amount of process parameters interacts with the feedback of detection data, and the error in the defect generation process is gradually reduced by adjusting the process parameters. The ultimate goal: to achieve iterative optimization through feedback of process parameters and detection data, accurately control the defect generation process, and improve production efficiency and product quality.

[0043] See also Figure 2 , a surface quality inspection system for metals, comprising the following modules: a multi-angle coherent imaging module, a process dynamic quantification module, a simulation manifold alignment module, and a process adjustment module; the multi-angle coherent imaging module is used to sequentially illuminate the metal surface through a multi-angle coherent light field, collect interference fringes caused by the depth of the defect, and analyze the phase offset of the defect area in the interference fringes according to the nonlinear mapping relationship between the dielectric properties of the metal and the optical path difference, and generate phase domain data containing the three-dimensional morphology of the defect; the process dynamic quantification module is used to perform multi-scale frequency domain decomposition and dynamic segmentation of the phase domain data of the three-dimensional morphology of the defect with process parameter constraints, extract the defect depth gradient distribution, and analyze the defect depth gradient distribution according to the metal processing history parameters and the defect depth gradient distribution. The deconvolution calculation of the cloth quantifies the deposition temperature deviation and pressure fluctuation, and generates the defect topology feature distribution that quantifies the process deviation through multi-scale topological decomposition and defect geometric feature statistics; the simulation manifold alignment module is used to construct a physical simulation atlas library of process parameter-defect morphology mapping relationship based on finite element simulation, align the defect topology feature distribution with the simulation data in the depth-constrained manifold space, extract cross-domain invariant features through comparative learning, and output the defect type and its causal correlation weight distribution with the process parameters; the process adjustment module is used to dynamically adjust the process parameters according to the gradient direction of the causal correlation weight distribution, and return the detection data under the new parameters to the simulation atlas library to update the depth-constrained manifold alignment rules.

[0044] In this implementation, the Multi-Angle Coherent Imaging Module utilizes a multi-angle coherent light field sequential illumination method, combined with a nonlinear mapping model between the dielectric properties of metal materials and optical path difference, to analyze the phase shift of interference fringes caused by metal surface defects, thereby generating phase-domain data with three-dimensional depth information. Compared to existing single-angle or fixed-viewing angle coherent detection methods, this module significantly improves spatial resolution and depth recognition capabilities, making it particularly suitable for fine-grained imaging of complex morphologies or microstructural defects. The Process Dynamic Quantification Module utilizes a multi-scale frequency domain decomposition and process-constrained dynamic segmentation approach to separate high-frequency defect signals from low-frequency background noise in defect topography data. Dynamic segmentation thresholds are constructed by integrating the statistical distribution of metal processing history parameters, enabling the segmentation process to adapt to varying process conditions. Furthermore, deposition temperature and pressure fluctuations are quantified using a defect depth gradient deconvolution method. This approach of inferring process disturbances from topography significantly outperforms traditional empirical inference methods, enabling quantification, localization, and traceability of abnormal process conditions. This module offers high industrial practical value and represents a technological breakthrough. Simulation Manifold Alignment Module: Based on a multi-physics coupled finite element simulation atlas, this module, for the first time, embeds defect topology features derived from actual inspections into the simulation manifold space. It also introduces defect depth gradient similarity weights and a geodesic optimization mechanism to improve feature alignment accuracy. Through deep-constrained manifold embedding combined with a contrastive learning mechanism, it extracts invariant features across domains (inspection vs. simulation), overcoming the recognition bias caused by inconsistencies between actual working conditions and simulation data in traditional inspection systems. This provides a more stable feature foundation for defect identification and causal traceability, demonstrating significant technological advancement. Process Adjustment Module: This module leverages the causal weight distribution between defect types and process parameters generated in the previous step and employs a gradient directional optimization method to achieve directional and quantitative adjustments to process parameters. It also dynamically feeds the new inspection data back into the simulation atlas. The manifold embedding rules are then simultaneously optimized through an incremental update mechanism, forming a complete closed-loop iterative mechanism of "inspection-causal determination-adjustment-re-inspection-re-optimization." Different from the traditional process feedback method based on rules or experience adjustment, this module has the ability of data-driven, adaptive adjustment, and autonomous learning and evolution, and is a key supporting component for realizing closed-loop control of intelligent manufacturing.

[0045] In summary, this application has at least the following effects: A surface quality inspection method and system for metals uses multi-physics field coupled finite element simulation to establish a mapping between process parameters and defect morphology, making the defect generation process predictable and controllable under different process conditions. A depth-constrained manifold embedding algorithm combines defect depth gradient similarity with a geodesic optimization strategy to align actual detected and simulated defects within the feature space, improving defect recognition accuracy. A contrastive learning mechanism under process parameter constraints effectively extracts cross-domain invariant features, providing a stable representation basis for defect recognition and classification. Causal reasoning methods are used to dynamically calculate the association weights between defect types and process parameters, quantifying the impact of process factors on defect formation and supporting decision optimization for quality control. Process parameters are dynamically adjusted based on the gradient direction of the causal association weight distribution. Closed-loop optimization of the defect detection model and process design is achieved through the reflow of inspection data and incremental updates of manifold weights. By continuously reflowing new inspection data and updating simulation maps, the timeliness and representativeness of the simulation database are ensured, enhancing the system's adaptability to complex working conditions.

[0046] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0048] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1The function specified in one or more boxes.

[0049] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0050] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0051] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for detecting the surface quality of metal, characterized in that: The following steps are involved: S1. Illuminate the metal surface with a multi-angle coherent light field sequence, capturing interference fringes caused by the depth of the defect. Based on the nonlinear mapping relationship between the metal's dielectric properties and the optical path difference, the phase offset of the defect region in the interference fringes is analyzed to generate phase domain data containing the three-dimensional morphology of the defect. S2. Perform multi-scale frequency-domain decomposition and dynamic segmentation of process parameter constraints on the phase-domain data of the defect's three-dimensional morphology to extract the defect depth gradient distribution. Quantify the deposition temperature deviation and pressure fluctuation based on the deconvolution calculation of the metal processing history parameters and the defect depth gradient distribution. Generate a defect topological feature distribution that quantifies the process deviation through multi-scale topological decomposition and defect geometric feature statistics. S3. Based on finite element simulation, a physical simulation atlas library of process parameter-defect morphology mappings is constructed. The defect topology feature distribution is aligned with the simulation data in a depth-constrained manifold space. Cross-domain invariant features are extracted through contrastive learning, and the defect type and its causal weight distribution with the process parameters are output. S4. Dynamically adjust the process parameters according to the gradient direction of the causal association weight distribution, and return the inspection data under the new parameters to the simulation atlas library to update the manifold alignment rules of the depth constraint.

2. A method for detecting surface quality of metal according to claim 1, characterized in that: Based on the nonlinear mapping relationship between the dielectric properties of metals and the optical path difference, the specific process of analyzing the phase offset of the defect area in the interference fringes is as follows: Noise suppression and background correction are performed on the interference fringe images collected from multiple angles, and fringe distortion caused by environmental interference is eliminated through phase shift encoding. Based on the spatial variation characteristics of the metal dielectric constant, a nonlinear mapping model between the optical path difference and the defect depth is constructed, and the model is converted into a nonlinear integral expression with constrained boundary conditions. The integral expression is numerically solved through multi-scale guided filtering and iterative optimization algorithm to extract the phase offset value.

3. A method for detecting surface quality of metal according to claim 2, characterized in that: The specific process of generating phase domain data containing the three-dimensional morphology of the defect is as follows: The phase offset is restored to a continuous phase distribution through the phase unwrapping algorithm. Combined with the wavelength and incident angle parameters of the light wave, the three-dimensional height field of the defect surface is calculated through the three-dimensional coordinate transformation model. The height field data is smoothed by surface interpolation to eliminate local distortion caused by noise. The depth gradient field, local principal curvature distribution and normal vector direction of the defect area are calculated through geometric feature analysis to generate phase domain data containing three-dimensional morphology, gradient features and geometric properties.

4. A method for detecting surface quality of metal according to claim 3, characterized in that: The specific process of performing multi-scale frequency domain decomposition and dynamic segmentation of process parameter constraints on the phase domain data of the defect 3D morphology to extract the defect depth gradient distribution is as follows: Perform multi-scale frequency domain decomposition on phase domain data and separate the background reflection noise and defect high-frequency signals through adaptive frequency band selection; Generate a dynamic segmentation threshold based on the statistical distribution of metal processing history parameters, and adjust the segmentation sensitivity by combining the probability density function of the phase gradient amplitude; The segmented defect area is verified for consistency in depth gradient direction, and false defect signals that deviate from the material stress field direction by more than a preset angle are eliminated.

5. The surface quality detection method for metal according to claim 4, characterized in that: The specific process of quantifying the deposition temperature deviation and pressure fluctuation is as follows: Establish the physical response relationship between process parameters and defect depth gradient, and obtain temperature deviation and pressure fluctuation by iteratively solving the deconvolution equation under constraint conditions; The statistical significance of the deconvolution results was verified using a random sampling method, defect areas that were strongly correlated with process parameter deviations were screened out, and their spatial locations and deviation levels were marked.

6. A method for detecting surface quality of metal according to claim 5, characterized in that: The specific process of generating defect topological feature distribution that quantifies process deviation through multi-scale topological decomposition and defect geometric feature statistics is as follows: The defect area is divided into spatial regions, and the hole density distribution, the consistency coefficient between the crack extension direction and the principal stress direction, and the surface curvature mutation characteristics of each sub-region are statistically analyzed; Through statistical correlation analysis, a quantitative mapping relationship between defect geometric characteristics and process parameter deviations is constructed to generate a characteristic distribution map reflecting the degree of process deviation.

7. A method for detecting surface quality of metal according to claim 6, characterized in that: The specific process of building a physical simulation atlas library of process parameter-defect morphology mapping relationships based on finite element simulation and aligning the defect topology feature distribution with the simulation data in the depth-constrained manifold space is as follows: Based on the multi-physics field coupled finite element simulation model, the defect generation process of the metal surface under different process parameter combinations is simulated to generate a simulation data set including temperature field, stress field distribution and defect morphology; The topological feature distribution of the actual detected defects and the simulation data are projected into the same manifold space through a depth-constrained manifold embedding algorithm, in which the manifold distance calculation introduces the defect depth gradient similarity weight, and the feature distribution is aligned through geodesic optimization.

8. A method for detecting surface quality of metal according to claim 7, characterized in that: The specific process of extracting cross-domain invariant features through contrastive learning and outputting the weight distribution of defect types and their causal associations with process parameters is as follows: Construct positive and negative sample pairs of actual defect features and simulation data in the manifold space, design a contrast loss function constrained by process parameters, and extract cross-domain invariant features through gradient descent optimization; Based on the optimized feature distribution, the association weights between defect types and process parameters are calculated through causal reasoning, where the weight distribution is dynamically updated through deconvolution gradient backpropagation.

9. A method for detecting surface quality of metal according to claim 8, characterized in that: The specific process of dynamically adjusting process parameters according to the gradient direction of the causal association weight distribution and returning the inspection data under the new parameters to the simulation atlas library and updating the manifold alignment rules of the depth constraint is as follows: Generate process parameter adjustments based on the negative gradient direction of the causal weight distribution, and use a constrained optimization algorithm to ensure that the parameter adjustments are in line with the physical feasible domain; The inspection data under the new parameters are added to the simulation atlas library, and the depth constraint weight matrix of the manifold space is optimized through the incremental parameter update method to achieve the coordinated iteration of the inspection model and process parameters.

10. A surface quality detection system for metal, applied to a surface quality detection method for metal according to any one of claims 1 to 9, characterized in that: It includes the following modules: multi-angle coherent imaging module, process dynamic quantification module, simulation manifold alignment module, and process adjustment module; The multi-angle coherent imaging module is used to sequentially illuminate the metal surface through a multi-angle coherent light field, collect interference fringes caused by the depth of the defect, and analyze the phase offset of the defect area in the interference fringes based on the nonlinear mapping relationship between the metal dielectric properties and the optical path difference to generate phase domain data containing the three-dimensional morphology of the defect; The process dynamic quantification module is used to perform multi-scale frequency domain decomposition and dynamic segmentation of process parameter constraints on the phase domain data of the defect three-dimensional morphology, extract the defect depth gradient distribution, quantify the deposition temperature deviation and pressure fluctuation based on the deconvolution calculation of the metal processing history parameters and the defect depth gradient distribution, and generate the defect topological feature distribution that quantifies the process deviation through multi-scale topological decomposition and defect geometric feature statistics; The simulation manifold alignment module is used to construct a physical simulation atlas library of process parameter-defect morphology mapping relationships based on finite element simulation, align the defect topological feature distribution with the simulation data in the depth-constrained manifold space, extract cross-domain invariant features through contrastive learning, and output the defect type and its causal association weight distribution with the process parameters; The process adjustment module is used to dynamically adjust process parameters according to the gradient direction of the causal association weight distribution, and return the detection data under the new parameters to the simulation atlas library to update the manifold alignment rules of the depth constraint.

Images