Intelligent coating nondestructive testing method based on XRF spectrum and neural network

Abstract

The invention discloses an XRF spectrum and neural network-based intelligent nondestructive coating detection method, and relates to the field of intelligent coating parameter adjustment, and the method combines a dual-energy XRF spectrum and a physical constraint neural network (PINN), uses a graph neural network to carry out global correction, and accurately analyzes coating characteristics from a plurality of physical levels. According to the method, a plurality of physical constraint condition parameters can be extracted from complex spectral data, network training is optimized through physical model constraint, so that a detection result is more accurate, and meanwhile, relatively high adaptability and robustness are shown in dynamic detection.

Description

Technical Field

[0001] The present invention relates to the field of intelligent parameter adjustment of coatings, and particularly to an intelligent non-destructive testing method for coatings based on XRF spectroscopy and neural networks. Background Art

[0002] In the field of coating inspection, traditional techniques mainly use physical and chemical analysis means, such as X-ray fluorescence spectroscopy (XRF), optical microscopy, scanning electron microscopy (SEM), etc., to evaluate the thickness, elemental composition, and microstructure of coatings. As a non-destructive testing technique, XRF is widely used in the qualitative and quantitative analysis of coating elements. It analyzes the elemental composition and concentration in the coating by exciting the sample and detecting the emitted characteristic X-ray spectrum. Traditional XRF techniques have been proven to be highly accurate in multiple industries, especially suitable for non-destructive analysis of multi-layer coatings, metal surfaces, and complex samples. In addition to XRF techniques, techniques such as scanning electron microscopy (SEM) and laser profile scanning are also commonly used for detailed observation of coating structures, especially in high-resolution surface analysis.

[0003] The main deficiencies of the existing technology lie in its lack of precise physical constraints and adaptability to complex coating systems. Traditional XRF techniques have errors in dealing with coating thickness, elemental concentration, and substrate scattering effects, especially in the surface analysis of multi-layer coatings or different substrates, where the accuracy is significantly affected. In addition, coating analysis methods based on traditional algorithms often cannot consider spectral peak drift, coating thickness non-uniformity, and other physical effects, resulting in insufficient accuracy and reliability of measurement results. To overcome these deficiencies, the coating inspection technology urgently needs a more efficient, accurate, and self-adaptive analysis method. Summary of the Invention

[0004] To solve the problems existing in the prior art, the purpose of the present invention is to provide an intelligent non-destructive testing method for coatings based on XRF spectroscopy and neural networks. By combining physical prior knowledge and the global correction function of graph neural networks, the present invention can adapt to various complex coating characteristics and environmental factors, especially suitable for high-demand industrial applications.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is specifically as follows: An intelligent non-destructive testing method for coatings based on XRF spectroscopy and neural networks, comprising the following steps: S1. Through dual-energy XRF excitation, spectral data in the range of 1 keV - 50 keV are collected, and the spectral data are pre-processed by the Compton-Rayleigh ratio method and the diffusion probability model; S2. Based on the processed spectral data, construct a Physics-Informed Neural Network model, and use the improved physical constraint loss function to constrain the coating thickness estimation, element concentration calculation, substrate scattering correction, and spectral peak drift respectively. Train the Physics-Informed Neural Network model according to the improved physical constraint model loss function; S3. Input the new spectral data, locally correct the new spectral data through the trained Physics-Informed Neural Network model, and output the physical constraint condition parameters of the new spectral data. The physical constraint condition parameters include coating thickness parameters, element concentration parameters, substrate scattering parameters, and spectral peak offset parameters; S4. Use the output physical constraint condition parameters of the new spectral data as the node features of the element interaction graph, use the interaction relationship between each physical constraint condition parameter as the edge of the element interaction graph, define the element interaction graph, and define the edge weights according to the interaction relationship; globally correct the element interaction graph through a graph neural network, and output the physically constrained condition parameters after secondary correction.

[0006] Further, the step S1 specifically includes the following sub-steps: S101. Irradiate the coating to be measured with a dual-energy XRF excitation source, and obtain its spectral data in the range of 1 keV - 50 keV; S102. Perform baseline correction on the obtained spectral data to eliminate background noise interference and remove artifact signals; S103. Perform diffusion noise reduction processing on the corrected spectral data through a diffusion probability model; S104. Locate the absorption peaks in the spectral data through a peak detection algorithm; S105. Extract the specific features of the preprocessed spectral data.

[0007] Further, the specific features of the spectral data at least include spectral intensity, wavelength, absorption peak position, and spectral line width.

[0008] Further, the step S2 specifically includes the following sub-steps: S201. Construct a Physics-Informed Neural Network model, and introduce physical constraint information into the hidden layer of the Physics-Informed Neural Network model; S202. Design a physical constraint loss function and train the Physics-Informed Neural Network model through the gradient descent optimization algorithm to converge the loss function; S203. Conduct generalization testing on the trained Physics-Informed Neural Network model; Among them, introducing physical constraint information in the hidden layer of the Physics-Informed Neural Network model specifically means: adding a physical model layer in the hidden layer, and respectively embedding physical prior knowledge in the physical model layer, and constraining the features of the hidden layer through the embedded physical prior knowledge.

[0009] Furthermore, the physical prior knowledge includes Beer-Lambert's law, fluorescence yield model, scattering-absorption mechanism, and electron migration theory.

[0010] Furthermore, the improved physical constraint loss function specifically includes a total loss function and four sub-loss functions. The total loss function is the weighted sum of the four sub-loss functions. Among them, the four sub-loss functions are: coating thickness loss function, element concentration loss function, substrate scattering loss function, and spectral peak drift loss function; the coating thickness loss function constrains the coating thickness according to Beer-Lambert's law, the element concentration loss function constrains the element concentration according to the fluorescence yield model, the substrate scattering loss function constrains the substrate scattering according to the scattering-absorption mechanism, and the spectral peak drift loss function constrains the spectral peak drift according to the electron migration theory.

[0011] Furthermore, the four sub-loss functions are specifically expressed as: Coating thickness loss function: ; Among them, the represents the coating thickness loss function, the represents the XRF signal intensity of the i-th pixel measured, the represents the intensity of the incident X-ray, the represents the linear absorption coefficient of the element, the represents the coating thickness of the i-th layer, the represents the constant of the coating, and the represents the index of the number of measurement points; Element concentration loss function: ; Among them, the represents the element concentration loss function, the represents the fluorescence signal intensity of the measured \(i\)-th element, and the represents the fluorescence signal intensity of the \(i\)-th element calculated theoretically; Substrate scattering loss function: ; wherein, the represents the substrate scattering loss function, the represents the scattering correction coefficient, the represents the incident spectral intensity, the represents the linear absorption coefficient of the substrate, and the represents the substrate thickness; Spectral peak drift loss function: ; wherein, the represents the spectral peak drift loss function, the represents the measured spectral peak offset, and the represents the theoretical spectral peak offset calculated according to the model.

[0012] Further, the step S3 specifically includes the following sub-steps: S301. According to the specific characteristics of the input spectral data and in combination with the Physics-Informed Neural Network model, preliminarily estimate the physical constraint condition parameters; S302. According to the physical constraint loss function, constrain the preliminarily estimated physical constraint condition parameters; S303. Use the physical constraint condition parameters constrained by the physical constraint loss function as the output of the Physics-Informed Neural Network model.

[0013] Further, the specific process of the step S301 is as follows: estimate the element concentration in the physical constraint condition parameters through the spectral intensity characteristics of the spectral data, estimate the spectral peak drift in the physical constraint condition parameters through the wavelength characteristics of the spectral data, estimate the substrate scattering correction in the physical constraint condition parameters through the absorption peak position characteristics of the spectral data, and estimate the coating thickness in the physical constraint condition parameters through the spectral line width characteristics of the spectral data.

[0014] Further, in the step S4, globally correct the element interaction graph through a graph neural network, and output the physically constrained condition parameters after secondary correction. The specific process steps include: S401. Through the message passing mechanism in the GNN, node features receive information from neighboring nodes and update the node features according to the edge weights; S402. The node features are updated and iterated by aggregating global information; S403. Through iteration, the estimation results of the physical constraint condition parameters are corrected at the global level; S404. Output the physical constraint condition parameters after the last iteration update.

[0015] The beneficial effects of the present invention are: By combining the dual-energy XRF spectroscopy technology with the Physics-Informed Neural Network, the present invention uses physical prior knowledge to constrain network training, overcomes the limitations of traditional detection methods in coating thickness, element concentration, and substrate scattering correction, realizes intelligent multi-dimensional and non-destructive detection, can accurately process complex interferences in spectral data simultaneously, improves the accuracy and robustness of coating analysis, and has high real-time performance and adaptability. Description of the Drawings

[0016] Figure 1 It is a flowchart of the method according to the embodiment of the present invention. Detailed Embodiments

[0017] The embodiments of the present invention will be described in detail below with reference to the drawings.

[0018] Embodiment 1 As Figure 1 shown, an intelligent non-destructive detection method for coatings based on XRF spectroscopy and neural network includes the following steps: S1. Through dual-energy XRF excitation, spectral data in the range of 1 keV - 50 keV is collected, and the spectral data is pre-processed by the Compton-Rayleigh ratio method and the diffusion probability model; S2. According to the processed spectral data, a Physics-Informed Neural Network model is constructed, and the coating thickness estimation, element concentration calculation, substrate scattering correction, and spectral peak drift are respectively constrained by an improved physical constraint loss function, and the Physics-Informed Neural Network model is trained according to the improved physical constraint model loss function; S3. Input the new spectral data, and perform local correction on the new spectral data through the trained Physics-Informed Neural Network model, and output the physical constraint condition parameters of the new spectral data. The physical constraint condition parameters include coating thickness parameters, element concentration parameters, substrate scattering parameters, and spectral peak offset parameters; S4. Take the output physical constraint condition parameters of the new spectral data as the node features of the element interaction graph, take the interaction relationships of each physical constraint condition parameter as the edges of the element interaction graph, define the element interaction graph, and define the edge weights according to the interaction relationships; perform global correction on the element interaction graph through a graph neural network, and output the physically constraint condition parameters after secondary correction.

[0019] Further, the step S1 specifically includes the following sub-steps: S101. Irradiate the coating to be measured with a dual-energy XRF excitation source, and obtain its spectral data in the range of 1 keV - 50 keV; S102. Perform baseline correction on the obtained spectral data to eliminate background noise interference and remove artifact signals; S103. Perform diffusion noise reduction processing on the corrected spectral data through a diffusion probability model; S104. Locate the absorption peaks in the spectral data through a peak detection algorithm; S105. Extract the specific features of the pre-processed spectral data.

[0020] Specifically, dual-energy X-ray fluorescence (XRF) technology irradiates a sample with two X-rays of different energies, exciting the elements in the sample to generate specific fluorescence signals. When X-rays irradiate a substance, the electrons in the atoms are excited, and fluorescence is released when the electrons transition. By measuring the intensity and wavelength of the fluorescence signal, the concentration and distribution of elements can be deduced. In addition, baseline correction is usually achieved by fitting the background signal and subtracting it from the total signal, and noise removal techniques (such as Gaussian smoothing, wavelet transform) can filter out high-frequency noise, making the effective signal clearer.

[0021] Further, the diffusion probability model is a noise reduction method based on the physical diffusion process, which smooths the signal by simulating the diffusion phenomenon during the signal propagation process and reduces the influence of noise. Exemplarily, the diffusion probability model is constructed based on the diffusion equation, and its specific description is: ; where, C represents concentration, t represents time, and D represents diffusion coefficient, Denotes the Laplace operator, which is used to describe the diffusion representation of a signal in space. Corresponding to the XRF spectrum, the diffusion equation represents the attenuation of the signal in the material due to scattering or absorption processes. In the above embodiments, the diffusion probability model is combined with physical constraints, and by embedding the physical constraints into the diffusion model, the stability and prediction accuracy of the model are improved. Specifically, physical models such as the Beer-Lambert law are used to constrain the absorption and scattering behavior of the signal to ensure that the model conforms to the actual physical process.

[0022] Furthermore, the specific features of the spectral data at least include spectral intensity, wavelength, absorption peak position, and spectral line width.

[0023] Specifically, spectral intensity is one of the most direct features in spectral data. Each element emits fluorescent radiation of different intensities in a specific energy band, and the concentration of the element is proportional to its spectral intensity in this band. Therefore, the estimation of the element concentration directly depends on the spectral intensity of the element. Furthermore, different elements have specific characteristic peaks, which are related to the transitions of inner-layer electrons. The change or drift of the spectral peak reflects the change of the material, such as the spectral changes caused by environmental factors such as temperature and stress. By analyzing the change of the wavelength, the drift of the spectral peak can be estimated. Further, the position of the absorption peak is related to the characteristic spectral line position of the element in the sample. When the sample substrate or the spectral environment changes, the position of the absorption peak may shift. By accurately positioning the absorption peak, the substrate scattering effect can be corrected. Further, the spectral line width is closely related to the element concentration, coating thickness, and physical state of the sample. The change of the coating thickness will cause the change of the spectral line width. A thicker coating results in spectral line broadening, so the coating thickness can be estimated by the spectral line width.

[0024] Furthermore, the step S2 specifically includes the following sub-steps: S201. Construct a Physics-Informed Neural Network model and introduce physical constraint information into the hidden layer of the Physics-Informed Neural Network model; S202. Design a physical constraint loss function and train the Physics-Informed Neural Network model through the gradient descent optimization algorithm to make the loss function converge; S203. Conduct a generalization test on the trained Physics-Informed Neural Network model; Among them, introducing physical constraint information into the hidden layer of the Physics-Informed Neural Network model specifically means: adding a physical model layer to the hidden layer, and respectively embedding physical prior knowledge in the physical model layer, and constraining the features of the hidden layer through the embedded physical prior knowledge.

[0025] Specifically, the Physics-Informed Neural Network model solves the differential equation problem containing physical constraints through a neural network. During the training process, the neural network not only learns according to the data, but also guides the learning direction of the network through physical equations. The output of the network not only conforms to the trend of the data, but also follows the known physical laws. For the above embodiments, the Physics-Informed Neural Network model includes: Input layer: Accept spectral data (spectral data in the range of 1 keV - 50 keV). After the data is preliminarily processed (such as preprocessing by the Compton-Rayleigh ratio method and the diffusion probability model), it is used as input data.

[0026] Hidden layer: In the hidden layer, physical prior knowledge is embedded. Through the physical prior knowledge, the neural network is forced to conform to specific physical laws during the learning process.

[0027] Output layer: The finally output are physical constraint condition parameters, including coating thickness, element concentration, substrate scattering correction, and spectral peak drift.

[0028] Furthermore, in each neuron of the hidden layer, the output of the network is constrained with physical quantities based on physical equations. Each neuron in the hidden layer will interact with the corresponding physical model layer to ensure that the output conforms to physical laws. Exemplarily, the relationships between coating thickness and element concentration and spectral data are explicitly modeled and passed to each layer in the network. In addition, the physical constraint information adjusts the activation functions of each layer in the network in a non-linear manner, so that the features output by each layer can meet the physical constraint requirements.

[0029] Furthermore, the hidden layer adopts a standard multi-layer perceptron (MLP) structure.

[0030] Furthermore, the physical prior knowledge includes Beer-Lambert's law, fluorescence yield model, scattering-absorption mechanism, and electron migration theory.

[0031] Furthermore, the improved physical constraint loss function specifically includes a total loss function and four sub-loss functions. The total loss function is the weighted sum of the four sub-loss functions. Among them, the four sub-loss functions are respectively: coating thickness loss function, element concentration loss function, substrate scattering loss function, and spectral peak drift loss function. The coating thickness loss function constrains the coating thickness according to the Beer-Lambert law. The element concentration loss function constrains the element concentration according to the fluorescence yield model. The substrate scattering loss function constrains the substrate scattering according to the scattering-absorption mechanism. The spectral peak drift loss function constrains the spectral peak drift according to the electron migration theory.

[0032] Furthermore, the four sub-loss functions are specifically expressed as: Coating thickness loss function: ; Among them, the represents the coating thickness loss function, the represents the XRF signal intensity of the measured i-th pixel, the represents the intensity of the incident X-ray, the represents the linear absorption coefficient of the element, the represents the coating thickness of the i-th layer, the represents the constant of the coating, and the represents the index of the number of measurement points; Element concentration loss function: ; Among them, the represents the element concentration loss function, the represents the fluorescence signal intensity of the measured i-th element, and the represents the fluorescence signal intensity of the theoretically calculated i-th element; Substrate scattering loss function: ; Among them, the represents the substrate scattering loss function, the represents the scattering correction coefficient, and the represents the incident spectral intensity. The represents the linear absorption coefficient of the substrate, and the represents the substrate thickness; Spectral peak drift loss function: ; Among them, the represents the spectral peak drift loss function, and the denotes the measured spectral peak offset, and the denotes the theoretical spectral peak offset calculated according to the model.

[0033] Furthermore, the step S3 specifically includes the following sub-steps: S301. Based on the specific characteristics of the input spectral data and in combination with the Physics-Informed Neural Network model, preliminarily estimate the physical constraint condition parameters; S302. Constrain the preliminarily estimated physical constraint condition parameters according to the physical constraint loss function; S303. Use the physical constraint condition parameters constrained by the physical constraint loss function as the output of the Physics-Informed Neural Network model.

[0034] Furthermore, the specific process of the step S301 is as follows: estimate the element concentration in the physical constraint condition parameters through the spectral intensity characteristics of the spectral data, estimate the spectral peak drift in the physical constraint condition parameters through the wavelength characteristics of the spectral data, estimate the substrate scattering correction in the physical constraint condition parameters through the absorption peak position characteristics of the spectral data, and estimate the coating thickness in the physical constraint condition parameters through the spectral line width characteristics of the spectral data.

[0035] Furthermore, in the step S4, globally correct the element interaction graph through the graph neural network, and output the physically constrained condition parameters after secondary correction. The specific process steps include: S401. Through the message passing mechanism in the GNN, the node features receive information from neighboring nodes and update the node features according to the edge weights; S402. The node features are updated and iterated by aggregating global information; S403. Through iteration, correct the estimation results of the physical constraint condition parameters at the global level; S404. Output the physically constrained condition parameters after the last iteration update.

[0036] Specifically, for the above embodiments, a graph neural network (GNN) is used to further optimize and correct the physical constraint parameters output by the Physics-Informed Neural Network (PINN) model. The GNN adjusts each physical constraint parameter further by propagating information between nodes, enabling the network to more accurately reflect physical phenomena. The output of the GNN will be a set of physically constrained parameters that have been corrected twice. Specifically, the physical constraint parameters (coating thickness, element concentration, substrate scattering correction, spectral peak shift) are calculated and output in the output layer of the PINN model. These output physical constraint parameters will be used as the node features of the graph neural network. The task of the graph neural network is to further correct these output parameters through the element interaction graph.

[0037] Furthermore, the working principle of the GNN is as follows: Message passing mechanism: In a graph neural network, nodes update their states through information interaction with their neighboring nodes. Through this process, each node receives the feature information of its neighbor nodes, thereby incorporating more context information into its own features. This information propagation is weighted according to the edge weights, thus adjusting the state of the nodes.

[0038] Node feature update: After each round of message passing, the features of each node are updated based on the features of its neighbor nodes. Through multiple rounds of iteration, the GNN can capture the global information in the graph, thereby effectively correcting the estimated values of each physical constraint parameter.

[0039] Global information integration: The GNN is not limited to local information propagation but can integrate the information of the entire graph through multi-level information transmission. Therefore, the model can more accurately reflect the global interactions and dependencies between physical constraint parameters.

[0040] Exemplarily, for the above embodiments, the correction process of the GNN is as follows: Coating thickness: The GNN further corrects the estimation of the coating thickness based on its interaction relationship with the physical constraints of element concentration and substrate scattering correction. For example, when the spectral data is very sensitive to changes in the coating thickness, the GNN strengthens the correction of this parameter; Element concentration: There is an association between the element concentration and other physical constraints, namely the coating thickness and substrate scattering correction. The GNN will adjust the estimated value of the concentration based on these relationships to enhance its accuracy; Substrate scattering correction: The GNN combines other physical constraints, namely the coating thickness and element concentration, to correct the substrate scattering correction parameter to ensure the global consistency of the physical model; Spectral peak drift: The GNN will correct the spectral peak drift. When there is a strong correlation between the spectral peak shift and other physical constraint conditions, the GNN captures the changes and makes adjustments.

[0041] Furthermore, the above embodiments include two corrections. One correction is a process of adjusting the preliminary physical constraint condition parameters through the Physics-Informed Neural Network model, which is completed during the training process of the Physics-Informed Neural Network model. The main goal is to ensure that the physical constraint conditions output by the model conform to the actual physical laws by embedding physical constraint information. The second correction is a process of further optimizing and correcting the physical constraint conditions using a graph neural network (GNN) after the Physics-Informed Neural Network model completes the first correction. The GNN mainly models the mutual relationship between the physical constraint condition parameters through a graph structure and adjusts the final value of each physical constraint condition through information transmission between nodes. Specifically, the first correction is a preliminary adjustment based on the constraint of the physical constraint loss function after the Physics-Informed Neural Network model outputs the preliminary physical constraint condition parameters. This ensures that the model conforms to the physical laws when outputting parameters, but due to its dependence on the training results of PINN, there are local errors.

[0042] The second correction is to further optimize the parameters after the first correction through a graph neural network. The second correction can fully consider the mutual dependence relationship between the physical constraint condition parameters, thereby achieving global optimization through information transmission and feature update, making the parameter correction more accurate.

[0043] Embodiment 2 As a preferred implementation of the above embodiment, a coating intelligent non-destructive testing system based on XRF spectroscopy and neural network is proposed, including: A data acquisition module for collecting spectral data in the range of 1 keV - 50 keV through dual-energy XRF excitation and preprocessing the spectral data through the Compton-Rayleigh ratio method and the diffusion probability model; a model construction module for constructing a Physics-Informed Neural Network model based on the processed spectral data and constraining the coating thickness estimation, element concentration calculation, substrate scattering correction, and spectral peak drift respectively through an improved physical constraint loss function, and training the Physics-Informed Neural Network model according to the improved physical constraint model loss function. The primary correction module is used to input new spectral data, locally correct the new spectral data through the trained Physics-Informed Neural Network model, and output the physical constraint condition parameters of the new spectral data. The physical constraint condition parameters include coating thickness parameters, element concentration parameters, substrate scattering parameters, and spectral peak shift parameters. The secondary correction module is used to use the physical constraint condition parameters of the output new spectral data as the node features of the element interaction graph, use the interaction relationships of the respective physical constraint condition parameters as the edges of the element interaction graph, define the element interaction graph, and define the edge weights according to the interaction relationships; globally correct the element interaction graph through a graph neural network, and output the physically constraint condition parameters after secondary correction.

[0044] Specifically, the improved physical constraint loss function specifically includes a total loss function and four sub-loss functions. The total loss function is the weighted sum of the four sub-loss functions. Among them, the four sub-loss functions are: coating thickness loss function, element concentration loss function, substrate scattering loss function, and spectral peak drift loss function; the coating thickness loss function constrains the coating thickness according to the Beer-Lambert law, the element concentration loss function constrains the element concentration according to the fluorescence yield model, the substrate scattering loss function constrains the substrate scattering according to the scattering-absorption mechanism, and the spectral peak drift loss function constrains the spectral peak drift according to the electron migration theory.

[0045] Furthermore, the four sub-loss functions are specifically expressed as: Coating thickness loss function: ; Among them, the represents the coating thickness loss function, the represents the XRF signal intensity of the i-th pixel measured, the represents the intensity of the incident X-ray, the represents the linear absorption coefficient of the element, the represents the coating thickness of the i-th layer, the represents the constant of the coating, and the represents the index of the number of measurement points; Element concentration loss function: ; Among them, the represents the element concentration loss function, the represents the fluorescence signal intensity of the i-th element measured, and the Represents the fluorescence signal intensity of the i-th element calculated theoretically; Substrate scattering loss function: ; Wherein, the Represents the substrate scattering loss function, the Represents the scattering correction coefficient, the Represents the incident spectral intensity, the Represents the linear absorption coefficient of the substrate, the Represents the substrate thickness; Spectral peak drift loss function: ; Wherein, the Represents the spectral peak drift loss function, the Represents the measured spectral peak offset, the Represents the theoretical spectral peak offset calculated according to the model.

[0046] The above embodiments only represent the specific implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. An intelligent nondestructive testing method for coatings based on XRF spectroscopy and neural network, characterized in that: The following steps are involved: S1. The spectral data in the range of 1 keV - 50 keV are collected by dual-energy XRF excitation, and the spectral data are pre-processed by the Compton-Rayleigh ratio method and the diffusion probability model; S2. Based on the processed spectral data, a Physics-Informed Neural Network model is constructed, and the coating thickness estimation, element concentration calculation, substrate scattering correction and spectral peak drift are constrained respectively through the improved physical constraint loss function, and the Physics-Informed Neural Network model is trained according to the improved physical constraint model loss function; S3. Input new spectral data, locally correct the new spectral data through the trained Physics-Informed Neural Network model, and output physical constraint parameters of the new spectral data, wherein the physical constraint parameters include coating thickness parameters, element concentration parameters, substrate scattering parameters, and spectral peak shift parameters; S4. The physical constraint parameters of the output new spectral data are used as the node features of the element interaction graph, and the interaction relationship of each physical constraint parameter is used as the edge of the element interaction graph. The element interaction graph is defined, and the edge weights are defined according to the interaction relationship; the element interaction graph is globally corrected through the graph neural network, and the physical constraint parameters after secondary correction are output.

2. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 1, characterized in that: The step S1 specifically includes the following sub-steps: S101. Irradiating the coating to be tested by a dual-energy XRF excitation source and obtaining its spectral data in the range of 1 keV - 50 keV; S102. Performing baseline correction on the acquired spectral data, eliminating background noise interference, and removing artifact signals; S103. Performing diffusion noise reduction processing on the corrected spectral data through a diffusion probability model; S104. Locating an absorption peak in the spectral data by a peak detection algorithm; S105. Extract specific features of the pre-processed spectral data.

3. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 2, characterized in that: The specific characteristics of the spectral data include at least spectral intensity, wavelength, absorption peak position and spectral line width.

4. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 1, characterized in that: The step S2 specifically includes the following sub-steps: S201. Construct a Physics-Informed Neural Network model and introduce physical constraint information into the hidden layer of the Physics-Informed Neural Network model; S202. Design a physical constraint loss function and train the Physics-Informed Neural Network model through a gradient descent optimization algorithm to make the loss function converge; S203. Perform generalization test on the trained Physics-Informed Neural Network model; Specifically, the physical constraint information is introduced into the hidden layer of the Physics-Informed Neural Network model as follows: a physical model layer is added to the hidden layer, and physical prior knowledge is embedded in the physical model layer, and the characteristics of the hidden layer are constrained by the embedded physical prior knowledge.

5. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 4, characterized in that: The physical prior knowledge includes Beer-Lambert law, fluorescence yield model, scattering-absorption mechanism and electron transfer theory.

6. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 4, characterized in that: The improved physical constraint loss function specifically includes a total loss function and four sub-loss functions, the total loss function is the weighted sum of the four sub-loss functions, wherein the four sub-loss functions are: coating thickness loss function, element concentration loss function, substrate scattering loss function and spectral peak drift loss function; the coating thickness loss function constrains the coating thickness according to the Beer-Lambert law, the element concentration loss function constrains the element concentration according to the fluorescence yield model, the substrate scattering loss function constrains the substrate scattering according to the scattering-absorption mechanism, and the spectral peak drift loss function constrains the spectral peak drift according to the electron migration theory.

7. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 6, characterized in that: The four sub-loss functions are specifically expressed as: Coating thickness loss function: ； Among them, the represents the coating thickness loss function, represents the measured XRF signal intensity of the i-th pixel, represents the intensity of the incident X-ray, represents the linear absorption coefficient of the element, represents the coating thickness of the i-th layer, represents the constant of the coating, An index indicating the number of measurement points; Element concentration loss function: ； Among them, the represents the element concentration loss function, represents the measured fluorescence signal intensity of the ith element, represents the theoretically calculated fluorescence signal intensity of the ith element; Substrate scattering loss function: ； Among them, the represents the substrate scattering loss function, represents the scatter correction coefficient, represents the incident spectral intensity, represents the linear absorption coefficient of the substrate, Indicates the thickness of the substrate; Spectral peak drift loss function: ； Among them, the represents the spectral peak drift loss function, represents the measured spectral peak shift, Represents the theoretical spectral peak shift calculated according to the model.

8. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 1, characterized in that: The step S3 specifically includes the following sub-steps: S301. Preliminary estimation of physical constraint parameters based on specific features of input spectral data combined with the Physics-Informed Neural Network model; S302. Constraining the initially estimated physical constraint condition parameters according to the physical constraint loss function; S303. Taking the physical constraint condition parameters constrained by the physical constraint loss function as the output of the Physics-Informed Neural Network model.

9. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 8, characterized in that: The specific process of step S301 is: estimating the element concentration in the physical constraint condition parameters through the spectral intensity characteristics of the spectral data, estimating the spectral peak drift in the physical constraint condition parameters through the wavelength characteristics of the spectral data, estimating the substrate scattering correction in the physical constraint condition parameters through the absorption peak position characteristics of the spectral data, and estimating the coating thickness in the physical constraint condition parameters through the spectral line width characteristics of the spectral data.

10. The method for intelligent nondestructive testing of coatings based on XRF spectrum and neural network as claimed in claim 1, characterized in that: In step S4, the element interaction graph is globally corrected by the graph neural network, and the physical constraint condition parameters after secondary correction are output. The specific process steps include: S401. Through the message passing mechanism in GNN, the node feature receives information from neighbor nodes and updates the node feature according to the edge weight; S402. Node features are updated and iterated by aggregating global information; S403. Through iteration, the estimation results of the physical constraint condition parameters are corrected at the global level; S404. Output the physical constraint condition parameters after the last iteration update.

Images