Coating defect diagnosis method and device based on physical mechanism constraint

Abstract

The application discloses a coating defect diagnosis method and device based on physical mechanism constraint, and relates to the technical field of coating detection. The method comprises the following steps: synchronously acquiring multi-modal detection data in the same detection area of a to-be-detected coating, wherein the multi-modal detection data at least comprises hyperspectral data, Fourier transform infrared spectrum data and electrochemical impedance spectrum data; extracting optical characteristics representing optical response of the coating, chemical characteristics representing chemical composition of the coating and electrochemical characteristics representing electrochemical behavior of the coating interface from the multi-modal detection data; inputting the optical characteristics, the chemical characteristics and the electrochemical characteristics into a pre-trained feature coupling model based on physical mechanism constraint; wherein the coating degradation physical mechanism is embedded into the training process of the feature coupling model as a constraint condition, and the multi-modal features are coupled and modeled; and outputting quantitative diagnosis parameters of the coating defect through the feature coupling model. The application realizes accurate and comprehensive coating diagnosis based on the coupling model.

Description

Technical Field

[0001] This application relates to the technical field of coating inspection, and in particular to a method and apparatus for diagnosing coating defects based on physical mechanism constraints. Background Technology

[0002] In the fields of major infrastructure and equipment such as ships, aerospace, and bridges, high-performance heavy-duty protective coatings are a key barrier protecting base materials from environmental factors such as corrosion, oxidation, and wear. With the continuous increase in equipment service life requirements and the popularization of preventive maintenance concepts, the industry has an increasingly urgent need for non-destructive and quantitative assessment technologies for coating health status, requiring a shift from passive repair to proactive prevention to reduce maintenance costs and improve equipment reliability.

[0003] Currently, coating defect detection mainly relies on single or simple combinations of techniques such as hyperspectral imaging, Fourier transform infrared spectroscopy, and electrochemical impedance spectroscopy. Among them, hyperspectral imaging can locate surface anomalies over a large area but cannot quantify key defect parameters; Fourier transform infrared spectroscopy can identify changes in molecular bonds but has low sampling efficiency and is difficult to establish macroscopic correlations; and electrochemical impedance spectroscopy can assess barrier performance but has low spatial resolution and lags in responding to early defects.

[0004] However, existing technologies typically simply superimpose the aforementioned multimodal detection methods, leaving the data acquired by different detection modes in an information silo state. They lack a multimodal feature coupling mechanism under physical constraints, making it impossible to establish a quantitative mapping relationship between the multidimensional characteristics of the coating. This results in highly subjective diagnostic results with poor repeatability, making it difficult to accurately quantify coating defects.

[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide a coating defect diagnosis method and device based on physical mechanism constraints, which aims to solve the technical problem that existing multimodal data are isolated and lack a feature coupling mechanism with physical mechanism constraints, making it impossible to accurately quantify coating defects.

[0007] To achieve the above objectives, this application proposes a coating defect diagnosis method based on physical mechanism constraints, the method comprising: In the same detection area of ​​the coating to be tested, multimodal detection data are acquired simultaneously. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data. Optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface are extracted from the multimodal detection data, respectively. The optical, chemical, and electrochemical features are input into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupled modeling of multimodal features; The feature coupling model outputs quantitative diagnostic parameters for coating defects, which include at least the coating debonding area, corrosion depth, and defect type.

[0008] In one embodiment, the step of simultaneously acquiring multimodal detection data in the same detection area of ​​the coating to be tested includes: Dark current correction and wavelet denoising are performed on the hyperspectral data to obtain a preprocessed hyperspectral data cube; The peak position, peak intensity, and peak area of ​​the Fourier transform infrared spectral data are extracted parametrically to obtain the preprocessed infrared spectral curve. Outlier removal and validity verification were performed on the electrochemical impedance spectroscopy data to obtain preprocessed complex impedance spectroscopy data.

[0009] In one embodiment, the step of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data includes: Principal component analysis was used to reduce the dimensionality of the hyperspectral data cube, and the spectral reflectance features and spatial features were selected using the maximum information coefficient as the optical features. Correlation analysis was used to analyze the intensity and ratio of the absorption peaks of characteristic functional groups in the infrared spectrum curves, and the chemical characteristics were output. The impedance complex spectrum data were fitted using an equivalent circuit fitting method to extract the coating resistance and double-layer capacitance as the electrochemical characteristics.

[0010] In one embodiment, the construction steps of the feature coupling model include spatiotemporal graph construction: The coating surface is divided into N spatial grid cells as nodes of the graph structure; The optical, chemical, and electrochemical features are concatenated to form the feature vector of each node at any given time. A dynamic adjacency matrix is ​​established based on spatial proximity and temporal continuity. The feature vector and the dynamic adjacency matrix are combined to output a dynamic spatiotemporal diagram of the coating state.

[0011] In one embodiment, the construction step of the feature coupling model further includes spatiotemporal graph convolutional network processing: In the spatial dimension, spectral graph convolution based on Chebyshev polynomial approximation is used to perform K-order neighborhood feature aggregation on the dynamic spatiotemporal graph to output a spatial feature map. In the time dimension, dilated causal convolution is introduced to process the spatial feature map, where the dilation factor increases exponentially with the network depth, and the spatiotemporal dependent features are output. Residual connections and gating mechanisms are applied to the spatiotemporal dependent features to output optimized spatiotemporal features.

[0012] In one embodiment, the construction step of the feature coupling model further includes multi-timescale analysis: Three parallel branches—high frequency, mid frequency, and low frequency—are constructed to receive optimized spatiotemporal features of corresponding sampling granularities, process them, and output multi-scale features. The feature pyramid module performs cross-scale interactive fusion of the multi-scale features to output full-spectrum evolutionary features.

[0013] In one embodiment, the construction step of the feature coupling model further includes physical constraint training: The physical mechanism of coating degradation is embedded in the network initialization process; the physical mechanism of coating degradation includes electrochemical kinetic equations and stress diffusion models; Construct a physical consistency loss function, and use the physical consistency loss function to constrain the prediction results obtained based on the full spectrum evolution characteristics, and output the trained feature coupling model based on physical mechanism constraints.

[0014] In one embodiment, after the step of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data, the method further includes: The spatial texture features and spectral response features of the optical features, the chemical features, and the electrochemical features are constructed into a four-dimensional feature vector; The four-dimensional feature vector is normalized, and the dynamically adjusted weights are determined based on the entropy weight method. The coating health index (CHI) is calculated by weighting and summing the normalized four-dimensional feature vectors according to the dynamically adjusted weights. When the coating health index (CHI) is lower than a preset threshold, the coating is determined to be in an early deterioration state and an early warning is triggered.

[0015] In one embodiment, the step of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data further includes the multi-coating layer interface: Savitzky-Golay smoothing and adaptive baseline correction are performed on the hyperspectral data to output denoised hyperspectral data; Pure spectral endmembers for each coating layer are extracted from the denoised hyperspectral data based on vertex component analysis; The spectral abundance of the pure spectral endmembers is inverted using the non-negative least squares method. Based on the physical constraint that the spectral abundance of each layer is non-negative and sums to 1, the spectral contribution ratio of each coating layer in each pixel is calculated, and the independent optical features of each coating layer are output. Based on the independent optical features, chemical features and electrochemical impedance features corresponding to each coating layer are extracted as layering features. The layering characteristics of each coating are input into the corresponding pre-trained feature coupling model based on physical mechanism constraints, and the defect type, health level or remaining protective performance of each coating is output.

[0016] Furthermore, to achieve the above objectives, this application also proposes a coating defect diagnosis device based on physical mechanism constraints, wherein the coating defect diagnosis device based on physical mechanism constraints includes: The data acquisition module is used to simultaneously acquire multimodal detection data in the same detection area of ​​the coating to be tested. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data. The feature extraction module is used to extract optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data, respectively. The feature processing module is used to input the optical features, chemical features, and electrochemical features into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupled modeling of multimodal features; The diagnostic output module is used to output quantitative diagnostic parameters of coating defects through the feature coupling model. The quantitative diagnostic parameters include at least the coating debonding area, corrosion depth, and defect type.

[0017] This application proposes a method and apparatus for diagnosing coating defects based on physical mechanism constraints. The method includes: simultaneously acquiring multimodal detection data in the same detection area of ​​the coating to be tested, wherein the multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data; extracting optical features characterizing the coating's optical response, chemical features characterizing the coating's chemical composition, and electrochemical features characterizing the coating's interfacial electrochemical behavior from the multimodal detection data; inputting the optical features, chemical features, and electrochemical features into a pre-trained feature coupling model based on physical mechanism constraints; wherein the physical mechanism of coating degradation is embedded as a constraint condition in the training process of the feature coupling model to couple and model the multimodal features; and outputting quantitative diagnostic parameters of coating defects through the feature coupling model, wherein the quantitative diagnostic parameters include at least the coating debonding area, corrosion depth, and defect type. This application employs a technique that simultaneously acquires multimodal data from hyperspectral, Fourier transform infrared, and electrochemical impedance spectroscopy, and extracts optical, chemical, and electrochemical multidimensional features. This addresses the issues of incomplete information from single detection technologies and the isolation of multi-source data, enabling multidimensional collaborative perception of coating conditions. Furthermore, it utilizes a technique that embeds the physical mechanisms of coating degradation as constraints into the training process to construct a feature coupling model. This addresses the problems of existing technologies lacking physical mechanism constraints and relying solely on data-driven fusion, resulting in strong subjective diagnosis and poor repeatability. This achieves deep coupling of multimodal features and model interpretability, ensuring the accuracy of feature correlation mapping. Finally, it employs a technique that uses the trained feature coupling model to directly output parameters such as debonding area, corrosion depth, and defect type. This addresses the problem of existing technologies being unable to quantify key defect parameters, enabling precise quantitative diagnosis of coating defects. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the coating defect diagnosis method based on physical mechanism constraints of this application; Figure 3This is a flowchart illustrating Embodiment 3 of the coating defect diagnosis method based on physical mechanism constraints of this application; Figure 4 This is a flowchart illustrating Embodiment 4 of the coating defect diagnosis method based on physical mechanism constraints in this application. Figure 5 This is a flowchart illustrating Embodiment 5 of the coating defect diagnosis method based on physical mechanism constraints in this application; Figure 6 This is a flowchart illustrating Embodiment Six of the coating defect diagnosis method based on physical mechanism constraints in this application; Figure 7 A schematic diagram of the architecture of a feature coupling model provided for the coating defect diagnosis method based on physical mechanism constraints in this application; Figure 8 This is a schematic diagram of the module structure of the coating defect diagnosis device based on physical mechanism constraints according to an embodiment of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] Due to the isolation of existing multimodal data and the lack of a feature coupling mechanism constrained by physical mechanisms, there is a technical problem that makes it impossible to accurately quantify coating defects.

[0025] This application provides a solution that employs a technique of simultaneously acquiring multimodal data from hyperspectral, Fourier transform infrared, and electrochemical impedance spectroscopy, and extracting optical, chemical, and electrochemical multidimensional features. This addresses the issues of incomplete information from single detection technologies and the isolation of multi-source data, enabling multidimensional collaborative perception of coating conditions. Furthermore, it utilizes a technique that embeds the physical mechanisms of coating degradation as constraints into the training process to construct a feature coupling model. This addresses the problems of existing technologies lacking physical mechanism constraints and relying solely on data-driven fusion, resulting in strong subjective diagnosis and poor repeatability. This achieves deep coupling of multimodal features and model interpretability, ensuring the accuracy of feature correlation mapping. Finally, it employs a technique where the trained feature coupling model directly outputs parameters such as debonding area, corrosion depth, and defect type, solving the problem of existing technologies being unable to quantify key defect parameters and enabling precise quantitative diagnosis of coating defects.

[0026] Based on this, embodiments of this application provide a coating defect diagnosis method based on physical mechanism constraints, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the coating defect diagnosis method based on physical mechanism constraints of this application.

[0027] In this embodiment, the coating defect diagnosis method based on physical mechanism constraints includes steps S10~S40: Step S10: In the same detection area of ​​the coating to be tested, multimodal detection data are acquired simultaneously. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data.

[0028] It should be noted that, in this embodiment, the coating under test refers to the protective coating on the surface of various infrastructure or equipment requiring deterioration defect diagnosis, such as heavy-duty protective coatings on the surfaces of ships, aerospace vehicles, and bridges. The same detection area refers to a locally overlapping area on the coating surface used for multimodal data acquisition. Synchronous acquisition refers to the parallel or alternating acquisition of different modal data within the same detection area within the same time period or according to a unified time reference, to ensure the alignment and consistency of each modal data in both time and spatial dimensions. Multimodal detection data refers to a set of detection data containing information about different dimensions of the coating, acquired through different physical principles or sensing technologies. Hyperspectral data refers to "space-spectral" cubic data that considers both the optical response characteristics in the spectral dimension and the defect distribution characteristics in the spatial dimension. Fourier transform infrared spectroscopy data refers to "wavenumber-absorbance" curve data reflecting changes in the coating's molecular bonds and the degree of chemical aging. Electrochemical impedance spectroscopy data refers to "frequency-impedance" complex spectrum data reflecting the coating's bulk protective performance and interfacial electrochemical behavior.

[0029] This embodiment enables multi-scale and multi-dimensional collaborative perception of the coating state by simultaneously acquiring multi-modal detection data in the same detection area of ​​the coating to be tested. This breaks the limitation of a single detection technology that can only acquire partial information, thus effectively solving the problem of isolated information and difficulty in establishing macro-to-micro correlations in a single detection technology.

[0030] In one possible implementation, the multimodal detection data may also include thermal imaging data or ultrasonic detection data to further enrich the physical structure information of the coating.

[0031] Step S20: Extract optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data.

[0032] It should be noted that, in this embodiment, the coating optical response refers to the changes in the optical properties of the coating, such as reflection and absorption, in response to incident light. Optical characteristics refer to parameters extracted from hyperspectral data that reflect the surface optical anomalies and spatial distribution of the coating, such as spectral reflectance characteristics and spatial characteristics. The coating chemical composition refers to the resins, fillers, and other substances constituting the coating, and their molecular bond structures. Chemical characteristics refer to parameters extracted from Fourier transform infrared spectroscopy data that reflect the degree of chemical aging and degradation of the coating, such as the intensity and ratio of characteristic functional group absorption peaks. The coating interfacial electrochemical behavior refers to electrochemical processes such as charge transfer and corrosion reactions occurring at the interface between the coating and the substrate. Electrochemical characteristics refer to parameters extracted from electrochemical impedance spectroscopy data that reflect the barrier performance and interfacial reactions of the coating, such as coating resistance and double-layer capacitance.

[0033] This embodiment aims to reduce the dimensionality of massive, heterogeneous raw data and transform it into refined feature vectors with clear physicochemical significance, effectively eliminating redundancy and noise interference in the raw data, and realizing the extraction of deep information from "spectral appearance" to "chemical mechanism" and then to "performance failure".

[0034] Step S30: Input the optical features, chemical features and electrochemical features into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupling modeling of multimodal features.

[0035] It should be noted that in this embodiment, the pre-trained model refers to a model whose parameters have been optimized using historical sample data and a specific training strategy, and has reached convergence, making it directly usable for inference and prediction. The feature coupling model based on physical mechanism constraints refers to a neural network model that integrates multiphysics equations and a deep learning architecture, enabling it to follow physical laws during feature fusion and prediction. The physical mechanism of coating degradation refers to the objective physical laws governing coating degradation in service environments, such as electrochemical kinetic equations and stress diffusion models. Constraints refer to penalty terms or guiding principles used to limit the model's prediction results to conform to specific physical laws or boundary conditions. Embedding refers to introducing physical equations as prior knowledge into the network initialization or loss function calculation process. Multimodal features refer to the collective term for extracted optical, chemical, and electrochemical features. This embodiment solves the problems of black-box nature, weak generalization ability, and difficulty in achieving cross-modal correlation under isolated data in purely data-driven models, realizing deep coupling and high-precision mapping of multimodal features under the guidance of physical laws.

[0036] In one possible implementation, the physical mechanism of coating degradation may also include coating thermodynamic degradation equations or diffusion penetration models to adapt to coating degradation mechanisms under different environments.

[0037] In one specific implementation, the system concatenates the extracted optical, chemical, and electrochemical features to form graph node feature vectors and constructs a dynamic spatiotemporal graph of the coating surface. This graph is then input into a feature coupling model based on the ST-GCN architecture, PhysNet-ST (a spatiotemporal graph neural network model incorporating physical rules). During the training process of this model, the electrochemical kinetic equations and stress diffusion models are embedded into the network initialization process. A physical consistency loss function is constructed and used to constrain the model's prediction results during training, thereby outputting a well-trained feature coupling model that can accurately capture the deep physical correlations between multimodal features during inference.

[0038] Step S40: Output quantitative diagnostic parameters of coating defects through the feature coupling model. The quantitative diagnostic parameters include at least the coating debonding area, corrosion depth, and defect type.

[0039] It should be noted that, in this embodiment, coating defects refer to abnormalities or damage in the physical structure, chemical composition, or protective performance of the coating, such as debonding, corrosion, and blistering. Quantitative diagnostic parameters refer to diagnostic indicators that can accurately describe the severity, size, or specific category of defects with concrete numerical values. Coating debonding area refers to the area where the coating loses adhesion to the substrate and peels off. Corrosion depth refers to the depth to which the substrate or coating is eroded by chemical or electrochemical processes. Defect type refers to the specific manifestations of coating deterioration damage, such as micro-debonding, resin oxidation, and blistering.

[0040] This embodiment outputs quantitative diagnostic parameters for coating defects through a feature coupling model, achieving a leap from qualitative assessment to precise quantitative diagnosis. It solves the problem that existing technologies cannot quantify key parameters such as defect size and depth, providing direct, reliable, and quantifiable data support for predictive maintenance and greatly improving the scientific nature and timeliness of coating maintenance decisions.

[0041] In one possible implementation, quantitative diagnostic parameters may also include the remaining protective life of the coating or the coating health level to further guide the formulation of maintenance cycles.

[0042] In one specific implementation, the system uses a trained feature coupling model to infer the feature data of the coating under test, directly outputting that the debonding area of ​​the coating in that region is 2.5 square centimeters, the corrosion depth is 0.3 millimeters, and the defect type is determined to be "interfacial micro-debonding accompanied by resin oxidation". These precise quantitative diagnostic parameters are transmitted to the terminal display device, providing maintenance personnel with accurate maintenance basis.

[0043] This application simultaneously acquires and extracts optical, chemical, and electrochemical features from multimodal coating detection data, inputting these features into a feature coupling model trained with constraints on the physical mechanisms of coating degradation, thereby outputting quantitative diagnostic parameters for coating defects. This approach breaks down information silos in multimodal data and overcomes the shortcomings of purely data-driven fusion, which suffers from strong subjectivity and poor interpretability. It achieves deep feature coupling from multi-source manifestations to physical mechanisms, enabling accurate and objective quantitative diagnosis of coating defects.

[0044] Furthermore, the step of simultaneously acquiring multimodal detection data in the same detection area of ​​the coating to be tested includes steps A201~A203: Step A201: Perform dark current correction and wavelet denoising on the hyperspectral data to obtain a preprocessed hyperspectral data cube; Step A202: Parametric extraction of peak position, peak intensity and peak area is performed on the Fourier transform infrared spectral data to obtain the preprocessed infrared spectral curve; Step A203: Outlier removal and validity verification are performed on the electrochemical impedance spectroscopy data to obtain preprocessed complex impedance spectroscopy data.

[0045] It should be noted that in this embodiment, dark current correction refers to the process of eliminating interference from the substrate current signal generated by the hyperspectral sensor under no-light conditions; wavelet denoising refers to an algorithm that uses wavelet transform to decompose the signal into different frequency bands and filter out high-frequency noise; the preprocessed hyperspectral data cube refers to a clean three-dimensional data array that has been freed from optical noise and current interference. Parameterized extraction of peak position, peak intensity, and peak area refers to the operation of identifying characteristic functional group absorption peaks directly related to the chemical aging of the coating from the infrared spectral data and calculating their position, height, and integral area; the preprocessed infrared spectral curve refers to simplified curve data containing key parameters for quantifying the degree of resin oxidation and degradation. Outlier removal refers to removing measurement noise points that deviate from the normal distribution due to disturbances in the test environment; validity verification refers to the verification process that confirms the impedance data conforms to causality and stability; the preprocessed impedance complex spectrum data refers to reliable electrochemical frequency response data that has been cleaned and verified by physical laws.

[0046] This application aims to eliminate equipment noise, environmental interference, and invalid redundant data caused by various sensors by performing preprocessing on the three modal data according to their physical characteristics. It transforms the original high-dimensional complex signals into standardized data that can focus on the key degradation information of the coating, effectively avoiding the interference terms in the original data from misleading the diagnostic model.

[0047] In one possible implementation, the execution order of dark current correction and wavelet denoising can be interchanged depending on the sensor type of the hyperspectral imager, or a deep learning denoising network can be used instead of wavelet denoising to preserve weaker spectral details. In another possible implementation, the parameterized extraction process can also be combined with a second-derivative spectral algorithm to accurately locate overlapping characteristic absorption peaks. In yet another possible implementation, outlier removal can be achieved using a statistical Laida criterion or a density-based clustering algorithm, and validity verification can be performed by using the Kramers-Kronig transform to check the causal compliance of the impedance data.

[0048] Further, the steps of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data include A301~A303: Step A301: Principal component analysis is used to reduce the dimensionality of the hyperspectral data cube, and the spectral reflectance features and spatial features are selected using the maximum information coefficient as the optical features.

[0049] It should be noted that, in this embodiment, principal component analysis refers to a dimensionality reduction algorithm that uses orthogonal transformation to convert potentially correlated high-dimensional variables into linearly uncorrelated low-dimensional variables, thereby extracting the main feature information of the data; the maximum information coefficient refers to a statistical index used to measure the strength of nonlinear and linear correlation between two variables; spectral reflectance characteristics and spatial characteristics refer to parameters that reflect the reflectivity of the coating surface to different wavelengths of light and the morphology and distribution of defects on a two-dimensional plane, respectively; optical characteristics refer to the set of features characterizing the optical response of the coating.

[0050] This embodiment aims to eliminate redundant information among massive spectral bands and significantly reduce the complexity of subsequent calculations. Then, the maximum information coefficient is used to evaluate the correlation between each component and the coating degradation state after dimensionality reduction, thereby accurately selecting the most representative spectral reflectance and spatial features as optical characteristics. Simultaneously, using the maximum information coefficient for screening effectively captures the complex nonlinear relationship between spectral variables and coating degradation, avoiding the omission of crucial and effective information that may occur with traditional linear screening methods.

[0051] Step A302: Analyze the intensity and ratio of the absorption peaks of characteristic functional groups in the infrared spectrum using correlation analysis, and output the chemical characteristics.

[0052] It should be noted that, in this embodiment, the intensity and ratio of the absorption peaks of characteristic functional groups refer to the peak height or area and relative proportion of the absorption of specific chemical bond vibrations in the infrared spectrum, which are used to quantify the degree of resin oxidation and degradation; correlation analysis refers to the mathematical method of calculating the strength and direction of statistical correlation between variables; chemical characteristics refer to the set of characteristics that characterize the evolution of the chemical composition of the coating.

[0053] This embodiment utilizes correlation analysis to process the absorption peak intensities and ratios of characteristic functional groups in the infrared spectral curve, aiming to uncover the deep statistical relationship between changes in functional group parameters and the chemical aging mechanism of the coating. By quantifying the evolution of molecular bonds during resin oxidation and degradation, the system can output chemical characteristics that accurately reflect the deterioration of the coating's chemical composition. This embodiment overcomes the limitation of Fourier transform infrared spectroscopy, as a point measurement technique, in directly establishing macroscopic correlations by transforming complex microscopic spectral signals into explicit chemical evolution indicators, achieving a precise characterization from "spectral appearance" to "chemical mechanism."

[0054] Step A303: The impedance complex spectrum data is fitted using the equivalent circuit fitting method to extract the coating resistance and double-layer capacitance as the electrochemical characteristics.

[0055] It should be noted that, in this embodiment, the equivalent circuit fitting method refers to using an equivalent circuit model composed of electronic components such as resistors and capacitors to simulate the electrochemical interface behavior of the coating, and adjusting the component parameters through an algorithm to make the model output approximate the measured impedance data; the coating resistance and double-layer capacitance refer to the key electrical parameters in the equivalent circuit that characterize the coating's ability to block the penetration of corrosive media and the charge accumulation and transfer characteristics at the interface between the coating and the substrate, respectively; the electrochemical characteristics refer to the set of features that characterize the electrochemical behavior of the coating interface.

[0056] This embodiment utilizes the equivalent circuit fitting method to fit the impedance complex spectrum data, aiming to decouple the high-dimensional and complex frequency-impedance response into circuit element parameters with clear physical meaning, thereby extracting the coating resistance and double-layer capacitance as electrochemical features. This effectively avoids the computational burden and information ambiguity caused by directly using the high-dimensional original impedance data, and transforms the macroscopic electrochemical frequency domain response into a quantitative indicator that intuitively characterizes the performance degradation of the coating barrier and the activity of interfacial electrochemical reactions.

[0057] Furthermore, referring to Figure 2 The second embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application provides a flowchart, based on the above. Figure 2 The illustrated embodiment further refines the step of "constructing the feature coupling model, including spatiotemporal graph construction" into steps A401-A403: Step A401: Divide the coating surface into N spatial grid cells as nodes of the graph structure.

[0058] It should be noted that in this embodiment, the coating surface refers to the external area of ​​the coating to be tested exposed to the environment; N refers to the classification determined according to the coating area size and detection accuracy requirements; spatial grid cell refers to the two-dimensional local area obtained after discretizing the coating surface into a grid; graph structure refers to a mathematical model composed of nodes and edges used to characterize the topological relationship between objects; node refers to the basic element in the graph structure, which is used here to represent a spatial grid cell.

[0059] This embodiment discretizes the continuous physical space of the coating into a computable network topology, providing a spatial data organization foundation for constructing a spatiotemporal graph convolutional network. This meshing process effectively preserves the spatial distribution characteristics and location information of coating defects, enabling the model to accurately capture the spatial propagation and evolution of defects, creating conditions for the spatial aggregation of multimodal features, and realizing the spatial structured representation of the coating state.

[0060] In one possible implementation, the N spatial grid cells are divided using a uniform rectangular grid. Alternatively, a non-uniform grid can be used based on the curvature variation of the coating surface or the distribution of suspected defect areas. In areas with dense suspected defects, the grid can be refined to improve the accuracy of local feature representation.

[0061] Step A402: The optical features, chemical features and electrochemical features are spliced ​​together to form the feature vector of each node at any time.

[0062] It should be noted that, in this embodiment, splicing refers to the operation of concatenating and merging feature vectors of different dimensions into a higher-dimensional vector according to a specific dimensional order; each node refers to the element corresponding to each spatial grid unit in the graph structure; any moment refers to any discrete time sampling point in the time series monitoring process; and the feature vector refers to a mathematical vector composed of multiple feature values ​​arranged in a certain order, used to comprehensively represent the state of the node.

[0063] This embodiment aims to break down information silos between different detection modalities and unify heterogeneous multimodal information into a single mathematical representation. Through this feature splicing operation, the entire chain of information, from "spectral appearance" to "chemical mechanism" and then to "performance failure," is integrated. This ensures that each node not only contains apparent optical anomalies but also deep chemical evolution and electrochemical degradation information, greatly enhancing the feature coupling model's comprehensive ability to express the coating degradation state.

[0064] In one possible implementation, the optical, chemical, and electrochemical features can be normalized separately before feature stitching to eliminate the impact of differences in the dimensions and numerical scales of different modal features on model training convergence.

[0065] Step A403: Establish a dynamic adjacency matrix based on spatial proximity and temporal continuity, combine the feature vector and the dynamic adjacency matrix, and output a dynamic spatiotemporal diagram of the coating state.

[0066] It should be noted that, in this embodiment, spatial proximity refers to the degree to which spatial grid cells are geographically adjacent or close on the surface of the physical coating; temporal continuity refers to the coherence and dependency of the state changes of the same spatial grid cell between adjacent time sampling points; dynamic adjacency matrix refers to a mathematical matrix used to describe the connection relationships and weights between nodes in the graph structure, and this matrix is ​​dynamically updated with time steps; eigenvector refers to the multimodal state vector obtained by splicing; coating state refers to the comprehensive health or deterioration performance of the coating in terms of physical structure, chemical composition and electrochemical performance; dynamic spatiotemporal graph refers to a graph data structure that integrates temporal dimension features, spatial dimension features and dynamic connection relationships.

[0067] This embodiment aims to construct a mathematical graph model that reflects both the spatial distribution and correlation of multimodal features and their evolution over time. Spatial proximity allows the propagation path of defects between adjacent grids to be traced, while temporal continuity ensures that the model can capture the temporal dependencies from rapid short-term changes to long-term trend evolution. The introduction of a dynamic adjacency matrix enables the model to adaptively adjust the influence weights between nodes. The final output dynamic spatiotemporal graph achieves a four-dimensional representation of the coating degradation state, significantly improving the feature coupling model's ability to make high-precision spatiotemporal predictions of coating degradation processes under complex environmental conditions.

[0068] In one possible implementation, the construction of the dynamic adjacency matrix can be based on calculating the similarity of spatial distances between nodes using a Gaussian kernel function, and combined with a time decay factor to enhance the influence of recent state on the current state, thereby more realistically reflecting the physical diffusion process of coating degradation.

[0069] In one specific implementation, the system constructs spatial adjacency edges based on the physical border relationships of spatial grid cells and constructs temporal edges based on the feature changes of adjacent time steps, generating a dynamic adjacency matrix that updates over time. Subsequently, the feature vectors of each node at each time step are combined with the dynamic adjacency matrix to output a dynamic spatiotemporal graph containing the multimodal features and dynamic connection weights of 100 nodes at multiple consecutive time steps, which is then used by the feature coupling model for subsequent spatiotemporal graph convolution processing.

[0070] This application discretizes the coating surface into nodes of a graph structure, concatenates multimodal features into feature vectors for each node, and combines this with a dynamic adjacency matrix constructed based on spatial proximity and temporal continuity to output a dynamic spatiotemporal graph of the coating state. This scheme preserves the spatial distribution and propagation characteristics of defects through spatial discretization, breaks down information silos in multimodal data through feature concatenation and achieves unified fusion of multidimensional information, and accurately characterizes the spatial correlation and temporal evolution of defects using the dynamic adjacency matrix. Ultimately, it achieves a four-dimensional representation of the coating degradation state, laying a solid foundation for subsequent spatiotemporal graph convolution to extract high-order dependency features and for high-precision spatiotemporal prediction.

[0071] Furthermore, referring to Figure 3 The third embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application provides a flowchart, based on the above. Figure 3 The illustrated embodiment further refines the step of "the construction step of the feature coupling model also includes spatiotemporal graph convolutional network processing", including steps A501 to A503: Step A501: In the spatial dimension, spectral graph convolution based on Chebyshev polynomial approximation is used to perform K-order neighborhood feature aggregation on the dynamic spatiotemporal graph to output a spatial feature map.

[0072] It should be noted that, in this embodiment, spatial dimension refers to the topological positional relationship and lateral association between various spatial grid units on the coating surface; Chebyshev polynomial approximation refers to the graph convolution approximation calculation method that uses Chebyshev polynomials to approximate the graph Laplacian operator spectral domain filter, thereby avoiding complex eigenvalue decomposition; spectral graph convolution refers to the convolution operation that filters the graph signal in the frequency domain of the graph Fourier transform; dynamic spatiotemporal graph refers to a graph data structure that integrates time dimension features, spatial dimension features, and dynamic connection relationships; K-order neighborhood feature aggregation refers to the process of collecting and fusing the feature information of all neighboring nodes within K hops of the center node in the graph structure; and spatial feature graph refers to the feature representation that contains local and global spatial association information of the coating obtained after spatial dimension convolution calculation.

[0073] This embodiment significantly reduces the computational complexity of spectral convolution by using Chebyshev polynomial approximation, while allowing the model to perform aggregation operations in the K-order neighborhood, thereby capturing a wider range of spatial correlations. This enables the effective extraction of the spatial propagation patterns and topological associations of coating defects between adjacent grids, ultimately outputting a spatial feature map that can accurately characterize the spatial distribution features of the coating state.

[0074] In one possible implementation, the order K of the K-order neighborhood can be set to 2 or 3 to strike a balance between capturing a sufficiently wide range of spatial correlations and controlling the computational overhead of the model.

[0075] Step A502: In the time dimension, dilated causal convolution is introduced to process the spatial feature map, wherein the dilation factor increases exponentially with the network depth, and spatiotemporal dependent features are output.

[0076] It should be noted that, in this embodiment, the time dimension refers to the longitudinal temporal relationship of the coating state evolving over time; dilated causal convolution refers to a convolution operation that introduces a dilation rate into standard convolution to expand the receptive field and strictly requires that the output at the current moment depends only on the input at the current and historical moments; spatial feature map refers to the feature representation obtained after spatial dimension convolution calculation, which contains local and global spatial correlation information of the coating; dilation factor refers to the parameter that controls the sampling interval or jump step size of the convolution kernel on the time series; network depth refers to the number of hidden layers in the neural network; exponential growth refers to the way in which the value of the dilation factor increases with the number of network layers according to an exponential function with a constant base; spatiotemporal dependency feature refers to a comprehensive feature representation that simultaneously encodes spatial topological correlation and temporal evolution dependency.

[0077] This embodiment utilizes dilated causal convolution to ensure strict temporal causality (i.e., predictions rely only on historical states without revealing future information). By exponentially expanding the temporal receptive field through an expansion factor that grows exponentially with network depth, the model can capture multi-level temporal dependencies from short-term rapid changes to long-term trend evolution without increasing network parameters. This transforms the spatial feature map into spatiotemporal dependency features that profoundly reflect the dynamic temporal laws of coating degradation, effectively improving the model's ability to model long-term degradation trends.

[0078] In one possible implementation, the initial value of the expansion factor can be set to 1, and it increases with the network depth according to the exponential sequence of 1, 2, 4, 8, so as to achieve hierarchical extraction of time dependencies of different granularities.

[0079] Step A503: Apply residual connections and gating mechanisms to the spatiotemporal dependent features to output optimized spatiotemporal features.

[0080] It should be noted that, in this embodiment, residual connection refers to the architectural design that adds the direct skip connections of the network layer input to the network layer output, which is used to alleviate the gradient vanishing or degradation problem in deep network training; gating mechanism refers to the structure that adaptively controls the filtering and forgetting process of information flow through learnable gating parameters to determine which feature information should be retained or suppressed; spatiotemporal dependency feature refers to the comprehensive feature representation that simultaneously encodes spatial topological association and temporal series evolution dependency; optimized spatiotemporal feature refers to high-quality spatiotemporal feature with stronger representation ability and stability after residual connection enhancement and gating feature filtering.

[0081] This embodiment aims to ensure that gradients can be effectively backpropagated in extremely deep spatiotemporal graph convolutional networks through residual connections, thereby solving the degradation problem in deep networks. At the same time, it uses a gating mechanism to adaptively filter and weight important spatiotemporal dependent features, suppressing redundant information and noise interference. The two work together to output optimized spatiotemporal features, thereby significantly improving the feature representation ability and training stability of the PhysNet-ST feature coupling model, ensuring high accuracy and high reliability of coating degradation diagnosis.

[0082] In one possible implementation, the gating mechanism can be implemented using a gating structure in a gated recurrent unit (GRU) or a long short-term memory network (LSTM), or by using a simple sigmoid gate to dynamically adjust the importance of features element-wise.

[0083] In one specific implementation, the system inputs a dynamic spatiotemporal graph containing multimodal information from 100 nodes into a spatiotemporal graph convolutional network processing module. First, in the spatial dimension, a second-order neighborhood feature aggregation is performed on the dynamic spatiotemporal graph using spectral graph convolution based on Chebyshev polynomial approximation to capture the defect propagation patterns in surrounding regions and output a spatial feature map. Next, in the temporal dimension, dilated causal convolution is introduced to perform temporal processing on the spatial feature map. The first layer has a dilation factor of 1 to capture short-term mutations, the second layer has a dilation factor of 2 to capture intermediate-term evolutions, and the third layer has a dilation factor of 4 to capture long-term trends, thereby outputting spatiotemporal dependency features that fuse multi-scale temporal dependencies. Finally, the system directly superimposes residual connections onto the spatiotemporal dependency features to preserve the original mapping and introduces a Sigmoid gating mechanism to adaptively filter irrelevant interference, ultimately outputting optimized spatiotemporal features with strong representational ability and stability for use in multi-timescale analysis and physical constraint training.

[0084] This application efficiently captures the extensive spatial propagation correlation of coating defects through K-order aggregation in the spatial dimension. By expanding causal convolution in the temporal dimension, it captures multi-level temporal dependencies from short-term mutations to long-term evolutions with an exponential receptive field while ensuring causality. Furthermore, it overcomes the degradation problem and noise interference of deep networks by combining residual connections and gating mechanisms, thereby significantly improving the model's ability to accurately extract, express, and train complex spatiotemporal dynamic features.

[0085] Furthermore, referring to Figure 4 The fourth embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application provides a flowchart, based on the above... Figure 4 The illustrated embodiment further refines the step of "constructing the feature coupling model also includes multi-timescale analysis," including steps A601-A602: Step A601: Construct three parallel branches for high frequency, medium frequency and low frequency, respectively, to receive the optimized spatiotemporal features of the corresponding sampling granularity, process them and output multi-scale features.

[0086] It should be noted that in this embodiment, the three parallel branches of high frequency, medium frequency and low frequency refer to three sub-network channels that run independently in the network and process data for different time frequency ranges respectively; sampling granularity refers to the size of the time interval or frequency coarseness of data acquisition; optimized spatiotemporal features refer to the features obtained after spatiotemporal graph convolution and gated residual processing; multi-scale features refer to the feature representations extracted from different time spans or frequency levels.

[0087] This embodiment aims to address the problem that a single time scale cannot simultaneously capture both short-term rapid mutations and long-term slow degradation of coatings. By splitting the feature processing at different time resolutions, the high-frequency branch focuses on capturing minute-level rapid mutations and transient anomalies, the mid-frequency branch focuses on day-level or week-level periodic fluctuations and stage-level evolution, and the low-frequency branch extracts long-term slow degradation trends at the month-level or year-level. This comprehensively extracts the evolution patterns of the coating at different time scales, outputting multi-scale features containing rich time-level information, ensuring that the model does not miss any key degradation information in any time band.

[0088] In one possible implementation, the division of high frequency, medium frequency and low frequency can be adaptively and dynamically adjusted according to the specific coating service environment and the sampling frequency of the monitoring equipment. For example, in cases where the high frequency fluctuation characteristics of the marine environment are obvious, the sampling rate weight of the high frequency branch can be increased.

[0089] Step A602: The multi-scale features are fused across scales using the feature pyramid module to output full-spectrum evolutionary features.

[0090] It should be noted that, in this embodiment, the feature pyramid module refers to a multi-scale feature processing architecture that aligns and fuses features at different levels through a top-down, bottom-up, or horizontal connection structure; multi-scale features refer to features at different time levels output by high-frequency, mid-frequency, and low-frequency branches; cross-scale interactive fusion refers to the process of combining high-level semantic features with low-level detailed features by performing information complementarity, feature mapping, and fusion calculation on features of different frequencies or scales; and full-spectrum evolutionary features refer to coating state evolution features that cover the complete time-frequency range from extremely short-term abrupt changes to extremely long-term trends and have undergone deep fusion.

[0091] This embodiment aims to break down information barriers between features at different time scales, enabling high-frequency detailed mutation information to corroborate and complement low-frequency global trend information. The feature pyramid module, through a bidirectional feature flow and fusion mechanism, effectively captures the full spectrum of evolutionary patterns, from minute-level mutations to interannual degradation, outputting full spectrum evolutionary features. This achieves a comprehensive and detailed characterization of the coating degradation process, significantly improving the PhysNet-ST model's ability to identify and predict complex dynamic degradation patterns.

[0092] In one possible implementation, cross-scale interactive fusion can introduce an attention mechanism to dynamically allocate the weights of features at different scales during the fusion process, so as to highlight the time-scale information that is most discriminative and physically significant for the current coating degradation state.

[0093] In one specific implementation, the system constructs three parallel branches for monitoring data of a bridge coating: high frequency (hourly sampling), medium frequency (daily sampling), and low frequency (monthly sampling). Each branch receives optimized spatiotemporal features of the corresponding sampling granularity and performs independent convolution processing, outputting high-frequency multi-scale features, medium-frequency multi-scale features, and low-frequency multi-scale features. Subsequently, the system upsamples the low-frequency global degradation trend features through the feature pyramid module and performs horizontal fusion with the medium-frequency features. The fusion result is then upsampled and fused with the high-frequency features. At the same time, the high-frequency abrupt change features are downsampled and passed to the medium-frequency and low-frequency branches to complete cross-scale interactive fusion. Finally, the system outputs a full-spectrum evolution feature containing the evolution from the initiation of surface microcracks at the hourly level to the overall corrosion resistance degradation law at the grade level, which can be used for physical constraint training.

[0094] This application achieves targeted extraction of features at different time granularities through parallel processing of multiple frequency branches, solving the problem that a single time scale cannot simultaneously account for both short-term rapid mutations and long-term slow degradation of coatings. At the same time, it breaks down the information barriers of multiple scales through cross-scale fusion of feature pyramids, achieving complementary advantages between high-frequency details and low-frequency trends, thereby accurately capturing the full spectrum of evolution from minute-level mutations to interannual degradation, and significantly improving the model's ability to identify and predict complex dynamic degradation patterns.

[0095] Furthermore, referring to Figure 5 The fifth embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application provides a flowchart, based on the above... Figure 5 The illustrated embodiment further refines the step of "constructing the feature coupling model also includes physical constraint training," including steps A701 to A702: Step A701: Embed the physical mechanism of coating degradation into the network initialization process; the physical mechanism of coating degradation includes electrochemical kinetic equations and stress diffusion models.

[0096] It should be noted that, in this embodiment, the physical mechanism of coating degradation refers to the objective physical laws and mathematical equations that govern the degradation of the coating in the service environment; the network initialization process refers to the process of initially assigning values ​​to parameters such as network weights and biases before the neural network begins training iterations; the electrochemical kinetic equation refers to the mathematical equation describing the relationship between the corrosion electrochemical reaction rate at the interface between the coating and the metal substrate and variables such as substance concentration and potential; and the stress diffusion model refers to the mathematical model describing the time and space expansion of defects such as microcracks and debonding caused by environmental stress inside the coating or at the interface.

[0097] This embodiment aims to inject explicit physical prior knowledge into the neural network in the early stage of model training, so that the initial parameters of the network are in the solution space that conforms to the physical logic of the coating degradation. This effectively avoids the local optima and black box problem that pure data-driven models are prone to fall into physical fallacies, greatly accelerates the convergence speed of the model and significantly improves its generalization ability in small samples or unseen conditions.

[0098] In one possible implementation, embedding the physical mechanism of coating degradation into the network initialization process can be achieved by converting the discretization results of the partial differential equations of the electrochemical kinetic equations and stress diffusion model into constraints for the initial weights of the network, or by using numerical simulation data of the physical equations to pre-train the network to determine the initial parameter distribution.

[0099] Step A702: Construct a physical consistency loss function, use the physical consistency loss function to constrain the prediction results obtained based on the full spectrum evolution characteristics, and output the trained feature coupling model based on physical mechanism constraints.

[0100] It should be noted that, in this embodiment, the physical consistency loss function refers to a comprehensive error evaluation function used to measure the data deviation between the neural network's predicted output and the true label, as well as the degree to which the predicted result deviates from the constraints of known physical equations; the full spectrum evolution characteristics refer to the coating state evolution characteristics that cover the complete time-frequency range from extremely short-term abrupt changes to extremely long-term trends and have been deeply fused; the prediction result refers to the output value of the coating's future state or defect parameters inferred by the feature coupling model based on the input features; constraint training refers to the process of using specific constraints or penalty mechanisms to guide the model towards iterative optimization in the direction of satisfying specific laws during the model parameter optimization and update process; the trained feature coupling model based on physical mechanism constraints refers to a deep learning network model whose parameters have been optimized and whose prediction behavior both fits the observed data and strictly follows the physical laws of coating degradation.

[0101] Based on data-driven learning, this application forces the model's predicted output to obey physical laws such as electrochemical kinetics and stress diffusion by using a penalty term in the residuals of the physical equations. This enables the model to not only have high-precision fitting ability but also strong physical interpretability and credibility, effectively overcoming the shortcomings of traditional black-box models, such as poor extrapolation ability and lack of physical meaning. The final output is a well-trained feature coupling model based on physical mechanism constraints, which enables high-fidelity and strong generalization spatiotemporal prediction of coating degradation process.

[0102] In one possible implementation, the physical consistency loss function can be constructed in the form of a weighted summation, including a mean squared error data loss term to measure the difference between the predicted and the true values, and a physical constraint loss term to measure the residuals generated after the predicted results are substituted into the electrochemical kinetic equation and the stress diffusion model. The weights of the two losses are dynamically adjusted to balance the data fit and physical consistency.

[0103] In one specific implementation, when constructing the feature coupling model for coating defect diagnosis, the system first discretizes the partial differential equations of the two physical mechanisms of coating degradation—the electrochemical kinetic equation and the stress diffusion model—and maps their prior solution space to the distribution constraints of the initial weights of the PhysNet-ST spatiotemporal graph convolutional network, completing the network initialization process. Subsequently, during the model training phase, the system constructs a physical consistency loss function that includes data loss terms and physical residual loss terms. The full spectrum evolution features are input into the network to obtain the predicted results of the coating debonding area and corrosion depth. The predicted results are substituted into the electrochemical kinetic equation and the stress diffusion model to calculate the physical residuals. The physical consistency loss function is used to perform joint backpropagation optimization on the deviation between the predicted results and the true labels, as well as the physical residuals. When the loss function converges, the trained feature coupling model based on physical mechanism constraints is output. This model can not only accurately predict the coating state, but its internal feature evolution also strictly follows physical laws.

[0104] This application embeds the physical mechanism of coating degradation, including electrochemical kinetic equations and stress diffusion models, into the network initialization process. It then uses a constructed physical consistency loss function to constrain the prediction results based on the full spectrum of evolutionary characteristics, outputting a well-trained feature-coupled model. This approach injects prior knowledge into the model through physical mechanism embedding initialization, avoiding the black-box problem and local optima of purely data-driven models, accelerating model convergence and improving generalization ability. Simultaneously, by constraining training with the physical consistency loss function, it forces the prediction output to conform to objective physical laws, significantly enhancing the model's interpretability and credibility. This overcomes the poor extrapolation ability of traditional models, achieving high-fidelity and highly generalized spatiotemporal prediction of coating degradation.

[0105] Furthermore, referring to Figure 6The sixth embodiment of the coating defect diagnosis method based on physical mechanism constraints in this application provides a flowchart, based on the above... Figure 6 The illustrated embodiment further refines the step of "extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data," including steps A801 ​​to A804: Step A801: Construct a four-dimensional feature vector from the spatial texture features and spectral response features of the optical features, the chemical features, and the electrochemical features.

[0106] It should be noted that, in this embodiment, spatial texture features refer to structural information reflecting the microstructure, roughness, and defect distribution of the coating surface; spectral response features refer to the distribution characteristics of the coating's reflection or absorption intensity of different wavelengths of light; chemical features refer to parameters extracted from Fourier transform infrared spectroscopy data that characterize the coating's chemical composition and molecular bond changes; electrochemical features refer to parameters extracted from electrochemical impedance spectroscopy data that characterize the coating's interfacial electrochemical behavior and corrosion resistance; and the four-dimensional feature vector refers to a mathematical vector composed of the above four independent and complementary feature dimensions arranged in a specific order.

[0107] This embodiment aims to map the originally heterogeneous multimodal detection data into a unified structured mathematical space, breaking down the data silos between different modes such as optical, chemical, and electrical. This allows optical information characterizing the surface deformation of the coating and chemical and electrochemical information characterizing the deep mechanism to complement and synergize within the same vector, effectively avoiding the problem of incomplete characterization of early degradation by single-modal features.

[0108] In one possible implementation, the order of the dimensions in the four-dimensional feature vector can be flexibly adjusted according to the actual output order of the feature extraction network, or the spatial texture features and spectral response features can be dimensionality reduced before construction to balance the dimensionality distribution of the four-dimensional feature vector.

[0109] Step A802: Normalize the four-dimensional feature vector and determine the dynamically adjusted weights based on the entropy weight method.

[0110] It should be noted that in this embodiment, the entropy weight method refers to an objective weighting method based on information entropy, which is an algorithm that determines the weight of each indicator by calculating the degree of variation of the data; the dynamically adjusted weight refers to the weight coefficient that is adaptively calculated and updated according to the actual distribution characteristics of the input data at different times or under different working conditions, rather than a fixed constant.

[0111] This embodiment eliminates the incomparability between four-dimensional features due to differences in physical dimensions and numerical scales by normalization, ensuring that each feature is on an equal footing in subsequent calculations. At the same time, it uses the entropy weight method to objectively measure the degree of variation of each dimension feature. The greater the degree of variation of a certain dimension data, the greater the amount of information it provides, and the higher the weight it is assigned. This achieves an objective and dynamic allocation of weights. This processing method can accurately match the physical fact that the dominant factors of the coating change at different stages of degradation (e.g., drastic electrochemical changes in the early stage and significant texture changes in the later stage), avoids the bias caused by subjective weighting, and significantly improves the objectivity and accuracy of health assessment.

[0112] In one possible implementation, the normalization process can employ the min-max normalization method or the Z-score normalization method to adapt to feature data with different distribution characteristics.

[0113] Step A803: Calculate the coating health index (CHI) by weighted summation of the normalized four-dimensional feature vectors according to the dynamically adjusted weights.

[0114] It should be noted that, in this embodiment, the Coating Health Index (CHI) refers to a comprehensive scalar indicator used to quantitatively characterize the overall health status and corrosion resistance degradation of the coating. This embodiment achieves a mapping from multi-dimensional heterogeneous data to a one-dimensional intuitive state by fusing and reducing complex multi-dimensional feature information into a single, intuitive scalar indicator. The CHI comprehensively reflects the overall degradation of the coating in terms of appearance, spectral characteristics, chemical composition, and interfacial electrochemical behavior. This allows engineers to quickly and accurately grasp the current service health level of the coating directly from this value without analyzing complex multi-modal data, greatly improving the efficiency and operability of condition assessment.

[0115] In one possible implementation, the coating health index (CHI) can be set to a range of 0 to 1, where 1 represents a completely healthy initial state and 0 represents a completely failed extreme state, and the lower the value, the more severe the degradation.

[0116] Step A804: When the coating health index (CHI) is lower than a preset threshold, the coating is determined to be in an early deterioration state and an early warning is triggered.

[0117] It should be noted that in this embodiment, the preset threshold refers to the critical judgment value of the coating health index set in advance according to the coating service standards, historical degradation data statistics or safe operation requirements; the early degradation state refers to the initial decay stage in which the coating has not yet experienced macroscopic large-area peeling or severe substrate corrosion, but irreversible chemical degradation, interfacial micro-cell formation or microcrack initiation has occurred at the micro level.

[0118] This embodiment aims to establish a quantitative failure judgment mechanism and closed-loop feedback response to accurately capture the key nodes of coating degradation from quantitative to qualitative change. By setting scientifically reasonable preset thresholds, the system can promptly identify and trigger early warnings when the coating is in an early deterioration state, avoiding the safety hazards and high costs of traditional post-treatment maintenance, and realizing condition-based maintenance and predictive maintenance, thereby significantly improving the safety of equipment operation and the economic benefits throughout its entire life cycle.

[0119] In one possible implementation, the preset threshold can be set differently according to different coating types and service environments, and the way to trigger the warning can include various forms such as popping up a red alarm interface on the monitoring terminal, sending a prompt text message, or pushing a maintenance work order.

[0120] In one specific implementation, spatial features, spectral features, chemical features, and electrochemical features are constructed into a four-dimensional feature vector: Among them, spatial texture features This mainly includes the area and morphological distribution of coating defects; and spectral response characteristics. Includes coating reflectance variation, absorption peak intensity; chemical characteristics Including the intensity ratio of characteristic functional group absorption peaks and the rate of change of functional group activity; electrochemical characteristics This includes coating resistance, double-layer capacitance, and diffusion parameters; based on normalized eigenvalues ​​and entropy weighting, a weighted summation coating health index (CHI) formula is constructed to quantitatively characterize the overall health status of the coating. ,in, The weights are dynamically adjusted based on physical sensitivity and degradation stage. These are the normalized feature values. When the coating health index is lower than a preset threshold, the coating is determined to be in an early deterioration state, and a deterioration warning is triggered, thereby enabling early identification of latent and nascent defects in the coating.

[0121] This application achieves unified representation and complementary fusion of heterogeneous multimodal data by constructing a four-dimensional feature vector. It uses the entropy weight method to dynamically assign weights, objectively reflecting the change of dominant factors in different deterioration stages of the coating. This eliminates the interference of subjective weighting and dimensional differences, and reduces complex multidimensional information to intuitive quantitative indicators, thereby achieving accurate assessment of the overall health status of the coating and timely early warning of early deterioration.

[0122] Furthermore, the step of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data also includes further refinement of the multi-coating layer interface, including steps A901~A905: Step A901: Perform Savitzky-Golay smoothing and adaptive baseline correction on the hyperspectral data to output denoised hyperspectral data.

[0123] It should be noted that in this embodiment, Savitzky-Golay smoothing refers to a filtering method based on the least squares principle, which uses polynomial fitting within a sliding window to filter out high-frequency noise while preserving spectral peaks and shape features; adaptive baseline correction refers to a preprocessing operation that dynamically adjusts parameters to deduct baseline shifts caused by scattering or fluorescence based on the actual background drift trend of the spectral signal; and denoised hyperspectral data refers to high signal-to-noise ratio spectral data that truly reflects the spectral response of the coating after noise filtering and baseline correction.

[0124] This embodiment aims to eliminate random noise and baseline drift interference caused by equipment vibration, changes in ambient light, and scattering from the coating surface during the acquisition process. By using Savitzky-Golay smoothing to filter out high-frequency spikes while preserving the morphological characteristics of the spectral curve to the greatest extent, and combining adaptive baseline correction to accurately remove background interference, the signal-to-noise ratio and authenticity of the data are significantly improved.

[0125] In one possible implementation, adaptive baseline correction can be achieved using asymmetric least squares or a dynamic baseline estimation method based on polynomial fitting to adapt to baseline drift of varying complexity.

[0126] Step A902: Extract the pure spectral endmembers of each coating layer from the denoised hyperspectral data based on vertex component analysis.

[0127] It should be noted that, in this embodiment, vertex component analysis refers to an unsupervised endmember extraction algorithm based on convex geometry theory, which extracts pure pixel spectra by finding the vertices of a single shape in a high-dimensional data space; each coating level refers to the physical layering structure of a multi-layer coating system (such as topcoat, intermediate coating, primer, etc.) consisting of the surface to the interior; and pure spectral endmembers refer to the characteristic spectral curves in the mixed spectrum that represent the inherent spectral response of a single pure substance (i.e., a specific coating level).

[0128] This embodiment aims to address the challenge of highly mixed spectral signals in multilayer coating systems. By utilizing the characteristic of vertex component analysis to locate extreme points in the space of convex geometry, the superimposed mixed spectra are decoupled, and pure spectral endmembers representing each coating layer such as topcoat, intermediate coat, and primer are accurately extracted. This enables effective identification and physical separation of the multilayer coating interface, providing benchmark features for subsequent quantification of the optical contribution of each layer.

[0129] In one possible implementation, when performing vertex composition analysis, the maximum distance projection method can be used to find the vertices of the simple shape one by one, or an iterative optimization method based on the maximum volume method can be used to extract pure spectral endmembers, so as to adapt to the endmember extraction requirements under different coating thicknesses and mixing degrees.

[0130] Step A903: The spectral abundance of the pure spectral endmember is inverted using the non-negative least squares method. Based on the physical constraint that the spectral abundance of each layer is non-negative and sums to 1, the spectral contribution ratio of each coating layer in each pixel is calculated, and the independent optical features of each coating layer are output.

[0131] It should be noted that in this embodiment, the nonnegative least squares method refers to an optimization algorithm that adds nonnegative constraints to variables when solving linear equations to minimize the sum of squared residuals; spectral abundance inversion refers to the reverse process of calculating the proportion or coverage of each endmember in the mixed pixel based on the known endmember spectrum; spectral abundance refers to the proportion or contribution of a certain pure endmember in the mixed pixel signal; each pixel refers to the smallest imaging unit under the spatial resolution of a hyperspectral image; spectral contribution ratio refers to the percentage of the spectral signal of a specific coating level in the total spectral signal of that pixel; and the independent optical characteristics of each coating level refer to the feature data that reflects the optical characteristics of a single coating level after removing interference from other coatings.

[0132] This embodiment accurately demixes the mixed spectrum at the pixel level. It uses the physical constraint of non-negative summation to ensure that the inversion results conform to the objective facts of the indestructibility of coating materials and area coverage. This accurately quantifies the spatial distribution and optical contribution of each coating layer, successfully realizing the transformation of multilayer coating features from mixed and intertwined to independent separation, and outputting the independent optical features of each coating layer.

[0133] In one possible implementation, the algorithm for realizing spectral abundance inversion under physical constraints can be the fully constrained least squares method, or Lagrange multipliers can be introduced on the basis of non-negative least squares to force the constraint condition that the abundance sum is 1 to be satisfied.

[0134] Step A904: Based on the independent optical features, extract the chemical features and electrochemical impedance features that correspond one-to-one with each coating layer as layering features.

[0135] It should be noted that, in this embodiment, each coating level refers to the physical layered structure of a multi-layer coating system consisting of the surface and the interior; chemical characteristics refer to parameters characterizing the chemical composition and molecular bond changes of a specific coating level; electrochemical impedance characteristics refer to impedance parameters characterizing the electrochemical corrosion behavior of a specific coating level at the interface with the metal substrate or between layers; and layering characteristics refer to a combination of features that strictly correspond to a specific coating physical level and contain optical, chemical, and electrical multimodal information.

[0136] This embodiment uses the spatial distribution and abundance information contained in independent optical features as a guide to accurately decouple the chemical features and electrochemical impedance features in multimodal detection data and associate them with the corresponding physical coatings. This avoids misjudgment caused by signal crosstalk between different layers, thereby forming exclusive photo-chemical-electrochemical multimodal layered features for each of the topcoat, intermediate coat, and primer, and realizing the accurate positioning and isolation of feature information in complex multi-layer systems.

[0137] In one possible implementation, the extraction of chemical and electrochemical impedance features corresponding one-to-one with each coating layer can be achieved through spatial registration and mapping extraction based on an abundance distribution mask of independent optical features, or by obtaining layered features at a specific depth through depth resolution scanning combined with tomographic stripping techniques.

[0138] Step A905: Input the layering features of each coating into the corresponding pre-trained feature coupling model based on physical mechanism constraints, and output the defect type, health level or remaining protective performance of each coating.

[0139] It should be noted that, in this embodiment, the pre-trained feature coupling model based on physical mechanism constraints refers to a deep learning network that has completed parameter optimization, incorporated prior knowledge of coating degradation physics, and is capable of multimodal feature coupling inference; the defect type refers to the specific morphological category of coating degradation; the health level refers to the qualitative or semi-quantitative level of the current integrity of the coating; and the remaining protective performance refers to the quantitative prediction index of the coating's ability to continue to provide effective anti-corrosion protection within a specific time in the future.

[0140] This embodiment addresses the significant differences in degradation mechanisms and rates among different layers in a multilayer coating system. By employing a dedicated feature coupling model for each layer to perform parallel and independent diagnostic reasoning, it achieves a leap from overall fuzzy assessment to precise layer-by-layer localization. This enables the system to clearly distinguish whether it is topcoat chalking, intermediate coat blistering, or primer peeling, and provides quantitative assessment results for each. This significantly improves the precision, physical interpretability, and engineering guidance value of coating condition assessment.

[0141] In one possible implementation, the pre-trained feature coupling model based on physical mechanism constraints corresponding to each coating can adopt a fully parallel architecture with independent network parameters, or a hybrid architecture with shared bottom-level feature extraction networks and hierarchical branching of top-level decision networks to balance the computational efficiency of the model with the particularity of each level.

[0142] In one possible implementation, refer to Figure 7In the overall technical architecture of the PhysNet-ST feature coupling model in this embodiment, its core logic is to achieve accurate assessment and prediction of coating degradation state through the collaborative efforts of three major modules. First, the data preparation stage completes the coating marking mesh division, multimodal feature extraction, and spatiotemporal data sequence organization, providing structured input for subsequent processing. Next, the spatiotemporal graph construction module transforms the coating region into a graph structure that can represent spatial relationships through node definition and feature encoding, dynamic graph structure generation, adjacency matrix learning, multi-layer graph structure construction, and inter-layer connection edge definition, laying the foundation for spatiotemporal feature extraction. The core GCN processing module sequentially executes spatial graph convolution (capturing spatial propagation correlations in the coating), temporal convolutional layers (handling temporal dependencies), gating mechanisms and residual connections (optimizing feature output), multi-scale temporal feature fusion (integrating information from different time frequencies), application of physical constraints (embedding the physical mechanisms of coating degradation), physical constraint loss calculation, and model parameter optimization, achieving deep extraction of spatiotemporal features and fusion of physical mechanisms. The multi-timescale analysis module processes features at corresponding sampling granularities through high-frequency, mid-frequency, and low-frequency data branches, and outputs full-spectrum evolution features after feature pyramid fusion, comprehensively capturing the multi-level temporal patterns of coating evolution from short-term abrupt changes to long-term evolution. Finally, the output and visualization module achieves degradation trajectory prediction, remaining lifetime probability prediction, critical path identification, spatiotemporal evolution visualization, phase transition point detection, and degradation heatmap generation, presenting the coating degradation state with multi-dimensional results. The entire process integrates multimodal feature extraction, spatiotemporal graph construction, ST-GCN processing with physical constraints, and multi-timescale analysis, forming a complete technical solution from data input to result output, supporting accurate diagnosis and prediction of early coating degradation states.

[0143] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the coating defect diagnosis method based on physical mechanism constraints of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0144] This application also provides a coating defect diagnosis device based on physical mechanism constraints, please refer to... Figure 8 The coating defect diagnosis device based on physical mechanism constraints includes: The data acquisition module 10 is used to simultaneously acquire multimodal detection data in the same detection area of ​​the coating to be tested. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data and electrochemical impedance spectroscopy data. The feature extraction module 20 is used to extract optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data, respectively. The feature processing module 30 is used to input the optical features, chemical features and electrochemical features into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupled modeling of multimodal features; The diagnostic output module 40 is used to output quantitative diagnostic parameters of coating defects through the feature coupling model. The quantitative diagnostic parameters include at least the coating debonding area, corrosion depth, and defect type.

[0145] The coating defect diagnosis device based on physical mechanism constraints provided in this application, employing the coating defect diagnosis method based on physical mechanism constraints in the above embodiments, can solve the technical problem that existing multimodal data is isolated, lacks a feature coupling mechanism based on physical mechanism constraints, and cannot perform accurate quantitative diagnosis of coating defects. Compared with the prior art, the beneficial effects of the coating defect diagnosis device based on physical mechanism constraints provided in this application are the same as those of the coating defect diagnosis method based on physical mechanism constraints provided in the above embodiments, and other technical features in the coating defect diagnosis device based on physical mechanism constraints are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0146] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A coating defect diagnosis method based on physical mechanism constraints, characterized in that, The coating defect diagnosis method based on physical mechanism constraints includes: In the same detection area of ​​the coating to be tested, multimodal detection data are acquired simultaneously. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data. Dark current correction and wavelet denoising are performed on the hyperspectral data to obtain a preprocessed hyperspectral data cube; parameterized extraction of peak position, peak intensity and peak area is performed on the Fourier transform infrared spectral data to obtain a preprocessed infrared spectral curve; outlier removal and validity verification are performed on the electrochemical impedance spectral data to obtain preprocessed complex impedance spectral data. The process involves extracting optical features characterizing the coating's optical response, chemical features characterizing the coating's chemical composition, and electrochemical features characterizing the coating's interfacial electrochemical behavior from the multimodal detection data. Specifically, this includes: reducing the dimensionality of the hyperspectral data cube using principal component analysis and selecting spectral reflectance and spatial features using the maximum information coefficient as the optical features; analyzing the intensity and ratio of characteristic functional group absorption peaks in the infrared spectrum using correlation analysis to output the chemical features; and fitting the impedance complex spectrum data using an equivalent circuit fitting method to extract the coating resistance and double-layer capacitance as the electrochemical features. The optical, chemical, and electrochemical features are input into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupled modeling of multimodal features; The feature coupling model outputs quantitative diagnostic parameters for coating defects, which include at least the coating debonding area, corrosion depth, and defect type.

2. The coating defect diagnosis method based on physical mechanism constraints as described in claim 1, characterized in that, The construction steps of the feature coupling model include spatiotemporal graph construction: The coating surface is divided into N spatial grid cells as nodes of the graph structure; The optical, chemical, and electrochemical features are concatenated to form the feature vector of each node at any given time. A dynamic adjacency matrix is ​​established based on spatial proximity and temporal continuity. The feature vector and the dynamic adjacency matrix are combined to output a dynamic spatiotemporal diagram of the coating state.

3. The coating defect diagnosis method based on physical mechanism constraints as described in claim 2, characterized in that, The construction steps of the feature coupling model also include spatiotemporal graph convolutional network processing: In the spatial dimension, spectral graph convolution based on Chebyshev polynomial approximation is used to perform K-order neighborhood feature aggregation on the dynamic spatiotemporal graph to output a spatial feature map. In the time dimension, dilated causal convolution is introduced to process the spatial feature map, where the dilation factor increases exponentially with the network depth, and the spatiotemporal dependent features are output. Residual connections and gating mechanisms are applied to the spatiotemporal dependent features to output optimized spatiotemporal features.

4. The coating defect diagnosis method based on physical mechanism constraints as described in claim 3, characterized in that, The construction steps of the feature coupling model also include multi-timescale analysis: Three parallel branches—high frequency, mid frequency, and low frequency—are constructed to receive optimized spatiotemporal features of corresponding sampling granularities, process them, and output multi-scale features. The feature pyramid module performs cross-scale interactive fusion of the multi-scale features to output full-spectrum evolutionary features.

5. The coating defect diagnosis method based on physical mechanism constraints as described in claim 4, characterized in that, The construction steps of the feature coupling model also include physical constraint training: The physical mechanism of coating degradation is embedded in the network initialization process; the physical mechanism of coating degradation includes electrochemical kinetic equations and stress diffusion models; Construct a physical consistency loss function, and use the physical consistency loss function to constrain the prediction results obtained based on the full spectrum evolution characteristics, and output the trained feature coupling model based on physical mechanism constraints.

6. The coating defect diagnosis method based on physical mechanism constraints as described in claim 5, characterized in that, The step of extracting optical features characterizing the coating's optical response, chemical features characterizing the coating's chemical composition, and electrochemical features characterizing the coating's interfacial electrochemical behavior from the multimodal detection data further includes: The spatial texture features and spectral response features of the optical features, the chemical features, and the electrochemical features are constructed into a four-dimensional feature vector; The four-dimensional feature vector is normalized, and the dynamically adjusted weights are determined based on the entropy weight method. The coating health index (CHI) is calculated by weighting and summing the normalized four-dimensional feature vectors according to the dynamically adjusted weights. When the coating health index (CHI) is lower than a preset threshold, the coating is determined to be in an early deterioration state and an early warning is triggered.

7. The coating defect diagnosis method based on physical mechanism constraints as described in claim 1, characterized in that, The step of extracting optical features characterizing the optical response of the coating, chemical features characterizing the chemical composition of the coating, and electrochemical features characterizing the electrochemical behavior of the coating interface from the multimodal detection data also includes the multi-coating layer interface: Savitzky-Golay smoothing and adaptive baseline correction are performed on the hyperspectral data to output denoised hyperspectral data; Pure spectral endmembers for each coating layer are extracted from the denoised hyperspectral data based on vertex component analysis; The spectral abundance of the pure spectral endmembers is inverted using the non-negative least squares method. Based on the physical constraint that the spectral abundance of each layer is non-negative and sums to 1, the spectral contribution ratio of each coating layer in each pixel is calculated, and the independent optical features of each coating layer are output. Based on the independent optical features, chemical features and electrochemical impedance features corresponding to each coating layer are extracted as layering features. The layering characteristics of each coating are input into the corresponding pre-trained feature coupling model based on physical mechanism constraints, and the defect type, health level or remaining protective performance of each coating is output.

8. A coating defect diagnosis device based on physical mechanism constraints, characterized in that, The coating defect diagnosis device based on physical mechanism constraints includes: The data acquisition module is used to simultaneously acquire multimodal detection data in the same detection area of ​​the coating to be tested. The multimodal detection data includes at least hyperspectral data, Fourier transform infrared spectral data, and electrochemical impedance spectroscopy data. The data preprocessing module is used to perform dark current correction and wavelet denoising on the hyperspectral data to obtain a preprocessed hyperspectral data cube; to perform parameterized extraction of peak position, peak intensity and peak area on the Fourier transform infrared spectral data to obtain a preprocessed infrared spectral curve; and to perform outlier removal and validity verification on the electrochemical impedance spectral data to obtain preprocessed complex impedance spectral data. The feature extraction module is used to extract optical features characterizing the coating's optical response, chemical features characterizing the coating's chemical composition, and electrochemical features characterizing the coating's interfacial electrochemical behavior from the multimodal detection data. Specifically, this includes: reducing the dimensionality of the hyperspectral data cube using principal component analysis and selecting spectral reflectance and spatial features using the maximum information coefficient as the optical features; analyzing the absorption peak intensities and ratios of characteristic functional groups in the infrared spectral curves using correlation analysis to output the chemical features; and fitting the impedance complex spectrum data using an equivalent circuit fitting method to extract the coating resistance and double-layer capacitance as the electrochemical features. The feature processing module is used to input the optical features, chemical features, and electrochemical features into a pre-trained feature coupling model based on physical mechanism constraints; wherein, the physical mechanism of coating degradation is embedded as a constraint condition into the training process of the feature coupling model to perform coupled modeling of multimodal features; The diagnostic output module is used to output quantitative diagnostic parameters of coating defects through the feature coupling model. The quantitative diagnostic parameters include at least the coating debonding area, corrosion depth, and defect type.

Images