Method and System for Identifying Microscopic Defects of New Energy Materials Based on Computer Vision

Abstract

The present invention relates to the fields of computer vision and materials science and technology, and particularly to a method and system for identifying microscopic defects of new energy materials based on computer vision. The method includes: first, collecting data through multimodal optical interference and photoacoustic imaging technologies, and performing feature alignment and fusion; then, performing adaptive signal decoupling on the preprocessed data, and reconstructing the defect area through multi-scale iterative inversion and grid optimization; finally, dynamically adjusting the model parameters through transfer learning and multi-level feedback optimization to improve the recognition accuracy and adaptability of the system. The present invention can accurately identify microscopic defects in wind power materials and provide a scientific basis for equipment maintenance decisions.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer vision and materials science, and particularly to a method and system for identifying microscopic defects of new energy materials based on computer vision. Background Art

[0002] Wind power materials (such as epoxy resin and carbon fiber on wind turbine blades) are prone to microscopic defects such as cracks, sand holes, and delamination during actual use due to environmental factors (such as wind erosion, salt spray, lightning strikes, etc.). These defects will affect the structural integrity of wind turbine blades, resulting in a decline in equipment performance and even serious accidents. Therefore, effective defect detection and identification are of great significance for ensuring the safe operation of wind power equipment.

[0003] The prior art (Chinese invention patent, publication number: CN114565571A, title: A method and device for detecting defects of wind turbine blades based on computer vision) mainly uses deep learning models and non-destructive testing means (such as infrared thermal imaging, acoustic emission monitoring, etc.) for defect detection of wind turbine blades. These methods usually process image data by constructing multi-scale feature fusion and deep convolutional networks. However, these methods have certain limitations: the feature extraction module is difficult to fully capture multi-modal information; the detection accuracy of the detection model for small defects is relatively low; and the data source is single (such as only using optical images), which limits the accuracy and adaptability of the detection. Summary of the Invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a method and system for identifying microscopic defects of new energy materials based on computer vision. The present invention is based on multi-modal data fusion and adaptive learning methods, and through the joint processing and dynamic optimization of multi-modal features, realizes the accurate identification and prediction of microscopic defects of wind power materials. Its core principles include the acquisition, fusion, adaptive signal decoupling, three-dimensional defect reconstruction, and multi-level feedback optimization of multi-modal feature data, improving the robustness and accuracy of the detection.

[0005] A method for identifying microscopic defects of new energy materials based on computer vision includes the following steps:

[0006] Perform multi-modal optical interference and photoacoustic imaging scans on new energy materials, collect phase interference data, polarized light data, and photoacoustic data, generate multi-modal response data through multi-modal feature alignment and multi-modal feature fusion, and then perform denoising processing and preprocessing on the multi-modal response data to obtain preprocessed multi-modal data;

[0007] Perform adaptive signal decoupling of sparse representation on the preprocessed multi-modal data to generate abnormal signal data, and perform non-linear enhancement processing and feature normalization processing on the abnormal signal data to obtain clustered abnormal signal data;

[0008] Reconstruct the defect area data of the clustered abnormal signal data through multi-scale iterative inversion, combine with the material physical model for adaptive grid optimization and dynamic regulation, and generate the final three-dimensional reconstruction data;

[0009] Conduct spatio-temporal perturbation simulation on the final three-dimensional reconstruction data, extract dynamic features, and perform multi-level feedback optimization based on correlation inversion to obtain defect evolution prediction data;

[0010] Based on the defect evolution prediction data, combine optical features, acoustic features, and mechanical features, and through cross-domain adaptive learning of the multi-level feedback optimization algorithm and the neural network model, dynamically adjust the weights and algorithm parameters of the neural network model to achieve system update and iterative optimization.

[0011] Preferably, in the multi-modal optical interference and photoacoustic imaging joint scanning step, by adjusting the phase angle and polarization angle, phase interference data and polarized light data are collected at different angles to cover the surface characteristics and internal structure characteristics of the new energy material; the multi-modal feature alignment and multi-modal feature fusion include registering the phase interference data, polarized light data, and photoacoustic data to align the phase interference data, polarized light data, and photoacoustic data in the same coordinate system, and fusing the eigenvalue of the phase interference data, polarized light data, and photoacoustic data through a weighted average method to generate multi-modal response data.

[0012] Preferably, in the adaptive signal decoupling step of sparse representation, the preprocessed multi-modal data is decomposed into normal signal data and abnormal signal data, where the sparse representation method is used to solve the sparse coefficients of the preprocessed multi-modal data, and the abnormal signal data is extracted by minimizing the reconstruction error, and the abnormal signal data is used for non-linear enhancement processing.

[0013] Preferably, the multi-scale iterative inversion step includes:

[0014] According to the clustered abnormal signal data, construct a material physical model, and perform initial grid division based on the stress and strain characteristics of the material to generate an initial defect area model; the grid division adopts an adaptive algorithm, and the grid in the high stress concentration area is refined by adjusting the grid density;

[0015] In each iteration process, calculate the inversion error and optimize the error term in the inversion process by dynamically adjusting the regularization parameter in the inversion algorithm, and use the iterative method to gradually update the defect area model to obtain three-dimensional reconstruction data with higher accuracy.

[0016] Preferably, the adjustment of the regularization parameter is based on the following expression: where, represents the The regularization parameter in the -th iteration; The regularization parameter in the -th iteration; represents the change in error in the current iteration, calculated as the difference between the error in this iteration and the error in the previous iteration;

[0017] Preferably, in the spatio-temporal perturbation simulation step, multiple perturbation conditions are imposed on the final three-dimensional reconstruction data, including successively adjusting the temperature field and stress field according to a preset perturbation scheme, and solving the spatio-temporal evolution equation of the material by numerical methods to extract the dynamic characteristics under different perturbation conditions, and the dynamic characteristics are used to reflect the changes in the defect area.

[0018] Preferably, the associated inversion processes the multi-scale spatio-temporal feature data through joint optimization, where the weight coefficient in the weighted convolution algorithm is calculated according to the following expression: where represents the weight coefficient of the -th feature layer; represents the feature importance coefficient of the -th feature layer; represents the total number of all feature layers; represents the weight normalization factor of all feature layers.

[0019] Preferably, in the cross-domain adaptive learning step of the multi-level feedback optimization algorithm and the neural network model, by combining the feedback results of optical features, acoustic features, and mechanical features, transfer learning is used to adjust the weights and algorithm parameters of the neural network model, where the transfer learning process dynamically adjusts by comparing the similarities between optical features, acoustic features, and mechanical features and updating the weights of the neural network model to optimize the overall performance of the system.

[0020] Preferably, in the transfer learning process, by constructing a cross-domain feature fusion model for optical features, acoustic features, and mechanical features, eigenvalue normalization processing is performed on optical features, acoustic features, and mechanical features, and the similarities between optical features, acoustic features, and mechanical features are calculated, and the weight update strategy of the neural network model is adjusted according to the similarities.

[0021] A system for implementing the above-mentioned computer vision-based new energy material micro-defect identification method, the system includes:

[0022] A multi-modal data acquisition module for performing multi-modal optical interference and photoacoustic imaging scans on new energy materials to collect phase interference data, polarized light data, and photoacoustic data;

[0023] A data preprocessing module, connected to the multi-modal data acquisition module, is used to perform feature alignment and multi-modal feature fusion on multi-modal data, generate multi-modal response data, and perform denoising and preprocessing on the multi-modal response data to obtain preprocessed multi-modal data;

[0024] A signal decoupling and anomaly detection module, connected to the data preprocessing module, is used to perform adaptive signal decoupling of sparse representation on the preprocessed multi-modal data, generate anomaly signal data, and perform non-linear enhancement and feature normalization processing on the anomaly signal data to obtain clustered anomaly signal data;

[0025] A multi-scale inversion and three-dimensional reconstruction module, connected to the signal decoupling and anomaly detection module, is used to perform multi-scale iterative inversion on the clustered anomaly signal data, combine with the material physical model for adaptive grid optimization and dynamic regulation, and generate the final three-dimensional reconstruction data;

[0026] A spatio-temporal evolution modeling module, connected to the multi-scale inversion and three-dimensional reconstruction module, is used to perform spatio-temporal perturbation simulation on the final three-dimensional reconstruction data, extract dynamic features, and perform multi-level feedback optimization based on associated inversion to obtain defect evolution prediction data;

[0027] A feedback optimization and adaptive learning module, connected to the spatio-temporal evolution modeling module, is used to perform cross-domain adaptive learning of a multi-level feedback optimization algorithm and a neural network model based on the defect evolution prediction data, combined with optical features, acoustic features, and mechanical features, and dynamically adjust the weights and algorithm parameters of the neural network model to achieve the update and iterative optimization of the system.

[0028] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0029] The present invention collects phase, polarization, and acoustic data through multi-modal optical interference and photoacoustic imaging, and combines multi-modal feature alignment and fusion technology to achieve higher-precision defect detection.

[0030] The present invention effectively separates anomaly signals by decoupling preprocessing data through sparse representation, enhancing the detection sensitivity to defects.

[0031] The present invention adopts a grid optimization technology based on the material physical model to improve the three-dimensional reconstruction accuracy of the defect area.

[0032] The present invention uses transfer learning to dynamically adjust model parameters, combined with the feedback of multi-modal data, to optimize the recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic flowchart of the method of the present invention;

[0034] Figure 2Schematic diagram of the multi-modal data fusion process in the present invention;

[0035] Figure 3 Schematic diagram of the multi-scale iterative inversion and adaptive optimization process in the present invention;

[0036] Figure 4 Flow chart of the dynamic feedback optimization and transfer learning in the present invention;

[0037] Figure 5 System structure block diagram of the present invention. Detailed implementation manners

[0038] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth in order to provide a comprehensive understanding of the embodiments of the present disclosure. However, obviously, one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0039] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising" and the like used herein indicate the presence of the described features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0040] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0041] As Figure 1 shown, a method for identifying microscopic defects of new energy materials based on computer vision includes the following steps:

[0042] Perform multi-modal optical interference and photoacoustic imaging scans on new energy materials, collect phase interference data, polarized light data and photoacoustic data, generate multi-modal response data through multi-modal feature alignment and multi-modal feature fusion, and then perform denoising processing and preprocessing on the multi-modal response data to obtain preprocessed multi-modal data;

[0043] Preferably, in the joint scanning step of multimodal optical interference and photoacoustic imaging, by adjusting the phase angle and polarization angle, phase interference data and polarized light data are collected at different angles to cover the surface characteristics and internal structure characteristics of the new energy material; the multimodal feature alignment and multimodal feature fusion include registering the phase interference data, polarized light data and photoacoustic data to align the phase interference data, polarized light data and photoacoustic data in the same coordinate system, and fusing the eigenvalue of the phase interference data, polarized light data and photoacoustic data by a weighted average method to generate multimodal response data.

[0044] As Figure 2 shown, in the present invention, first, a joint scan of multimodal optical interference and photoacoustic imaging is performed on new energy materials (mainly wind turbine blade materials, such as epoxy resin and carbon fiber) to achieve efficient identification and analysis of material microdefects.

[0045] This method utilizes the advantages of the combination of multimodal optical interference and photoacoustic imaging technologies to collect detailed information on the surface and inside of wind power materials. The optical interference technology can provide high-resolution microscopic features of the material surface by analyzing the phase difference between the incident light and the reflected light; while photoacoustic imaging excites ultrasonic waves on the material surface by laser irradiation, and obtains the internal structure features of the material according to the time and intensity changes of the ultrasonic waves propagating inside the material. Therefore, these two imaging methods complement each other and can simultaneously detect surface and internal defects of the material.

[0046] In specific operations, the system will scan the wind power material at multiple angles, and by adjusting the phase angle and polarization angle of the optical interference, phase interference data and polarized light data are collected at different angles. This multi-angle scanning method can cover the surface microscopic features of the material (such as surface scratches and cracks) and the internal structure features of the material (such as internal voids and delaminations). In this way, the comprehensiveness and accuracy of the detection can be significantly improved. In addition, photoacoustic imaging supplements the internal information that cannot be detected by optical interference, making the scanning results more complete.

[0047] The data collected include phase interference data, polarized light data and photoacoustic data. To ensure the accuracy of multimodal data, these data need to be processed by multimodal feature alignment and multimodal feature fusion. Specifically, multimodal feature alignment refers to mapping the data of each modality into the same coordinate system so that different modality data have a consistent reference standard in space. This step is achieved through registration technology, aligning the phase interference data, polarized light data and photoacoustic data into the same coordinate system to ensure the correlation between the data in the subsequent processing.

[0048] The aligned multimodal data needs to be fused. The purpose of multimodal feature fusion is to integrate the useful information from data of different modalities into a whole to generate more accurate multimodal response data. In the present invention, the eigenvalues of phase interference data, polarized light data, and photoacoustic data are weighted and fused by the method of weighted average. The weight settings for weighted fusion are adjusted according to data quality and modal importance. For example, in the detection of wind turbine blade materials, photoacoustic data may be more important for describing internal defects of the material. Therefore, the weight of photoacoustic data can be appropriately increased during weighting to improve the sensitivity of detecting internal defects of the material. The generated multimodal response data after fusion contains a comprehensive description of the surface and internal characteristics of the material, providing more comprehensive data support for subsequent defect analysis.

[0049] To further improve the quality of the multimodal response data, the fused multimodal response data also needs to be denoised and preprocessed. Denoising is mainly achieved through filters to eliminate the possible noise interference during the acquisition process, such as the influence of external environmental vibrations or measurement device errors on the data; while preprocessing includes operations such as normalization and scale adjustment to make the data more suitable for subsequent signal processing and defect analysis steps in terms of numerical range and distribution.

[0050] This step realizes the comprehensive detection of wind turbine blade materials and the comprehensive processing of multimodal data. It can not only identify the tiny cracks and defects on the material surface but also detect the internal structural defects (such as voids, delaminations, or debonds) of the material. The multi-angle scanning method and multimodal feature fusion significantly improve the detection resolution and accuracy, enabling the present invention to effectively handle complex defect situations. In addition, through the alignment and fusion of multimodal data, a comprehensive analysis of the material microstructure can be achieved. Especially for the composite materials of wind turbine blades, the defect distribution and characteristics between different material layers can be accurately distinguished.

[0051] For example, assume that there is an incompletely bonded area between the epoxy resin layer and the carbon fiber layer of a wind turbine blade. In the phase interference data, it may be manifested as a tiny amplitude change on the surface, while photoacoustic imaging may detect an abnormal local acoustic wave propagation speed. Through the multimodal feature fusion processing in the present invention, these two abnormal features can be considered and combined simultaneously to accurately locate the position and range of the defect and evaluate its impact. Finally, through the denoising and preprocessing of the data, the reliability and accuracy of the defect detection results are further ensured.

[0052] Perform adaptive signal decoupling of sparse representation on the preprocessed multimodal data to generate abnormal signal data, and perform nonlinear enhancement processing and feature normalization processing on the abnormal signal data to obtain clustered abnormal signal data;

[0053] Preferably, in the adaptive signal decoupling step of sparse representation, the preprocessed multimodal data is decomposed into normal signal data and abnormal signal data. The sparse coefficients of the preprocessed multimodal data are solved using the sparse representation method, and the extraction of abnormal signal data is achieved by minimizing the reconstruction error. The abnormal signal data is used for non-linear enhancement processing.

[0054] In the present invention, the adaptive signal decoupling of the preprocessed multimodal data by sparse representation is used to extract and identify abnormal signals in the material, so as to detect and analyze possible microscopic defects in wind turbine blade materials (such as epoxy resin and carbon fiber layers). This step uses sparse representation technology to process signal data, thereby highlighting abnormal features in the massive data and improving the accuracy and reliability of defect identification.

[0055] The preprocessed multimodal data is adaptively signal-decoupled by the sparse representation method. Sparse representation is a signal processing technology, and its core idea is to represent a signal as a linear combination of a few elements in a dictionary, so that most irrelevant or redundant information is ignored, and only important feature information is retained. In the present invention, the goal of sparse representation is to decompose the preprocessed multimodal data into normal signal data and abnormal signal data. The normal signal data reflects the background information of the material and the characteristics of the defect-free area, while the abnormal signal data represents the potential defect area in the material.

[0056] Specifically, the adaptive signal decoupling step of sparse representation first sparsifies the preprocessed multimodal data using the sparse coding method. The process of sparsification can be described as solving the sparse coefficients to minimize the reconstruction error of the signal. The mathematical expression of this step is usually: where, represents the preset dictionary matrix, represents the sparse coefficient vector, is the preprocessed multimodal data, is a balance parameter used to control the balance between sparsity and reconstruction error. By solving the above optimization problem, the sparse coefficient vector can be obtained, and then the abnormal signal data and the normal signal data are decomposed. The abnormal signal data usually shows a significant enhancement of the sparse coefficients at certain positions, indicating that these signal features may correspond to potential defects in the material.

[0057] To further highlight the characteristics of the abnormal signal data, the present invention introduces non-linear enhancement processing. The purpose of non-linear enhancement is to amplify the weak features in the abnormal signal data, making them more easily recognizable and separable in subsequent analysis. This step usually adopts methods such as logarithmic transformation or exponential enhancement to stretch the amplitude of the signal to improve the contrast. For example, assuming that the response of the detected abnormal signal data is weak in certain frequency bands, logarithmic enhancement can be applied to the signals in these frequency bands, so that smaller signal changes are more prominent in the enhanced signal. This is particularly important for detecting micro-cracks or delaminations in wind turbine blades, as these defects may appear as signals with very low amplitudes in the original data.

[0058] After completing the non-linear enhancement processing, feature normalization is performed on the abnormal signal data. The purpose of feature normalization is to map data of different modalities and scales to a unified numerical range to eliminate the bias caused by different signal amplitudes or units, ensuring that subsequent clustering analysis can be based on the same standard. Normalization can be achieved through linear normalization (such as min-max normalization) or standardization (such as zero-mean normalization). The normalized abnormal signal data will be used as the clustered abnormal signal data for further defect area clustering and analysis.

[0059] Through the adaptive signal decoupling of sparse representation, the normal features and abnormal features of the wind turbine blade material can be effectively separated, significantly improving the detection ability of micro-defects. After the abnormal signal data undergoes non-linear enhancement and feature normalization processing, the recognizability of micro-defects can be significantly enhanced, making the present invention show higher sensitivity in detecting micro-cracks, delaminations or other hidden defects. For example, if there is an unbonded area between the epoxy resin layer and the carbon fiber layer of a wind turbine blade, sparse representation can separate the abnormal signal data in this area, non-linear enhancement processing will amplify its signal amplitude, and normalization processing can ensure that this signal has significant distinguishable features in subsequent analysis.

[0060] This method has important application value in the daily detection and maintenance of wind turbine blades. By detecting and identifying micro-defects in a timely manner, the expansion of defects can be prevented and the service life of the blades can be extended. In addition, the adaptability of the sparse representation technology also enables the present invention to be flexibly adjusted according to different material characteristics and detection environments, improving the versatility and adaptability of detection.

[0061] Reconstruct the defect area of the clustered abnormal signal data through multi-scale iterative inversion, combine the material physical model for adaptive grid optimization and dynamic regulation to generate the final three-dimensional reconstruction data;

[0062] Preferably, as Figure 3 shown, the multi-scale iterative inversion step includes:

[0063] Based on the clustered abnormal signal data, a physical model of the material is constructed, and an initial mesh division is performed based on the stress and strain characteristics of the material to generate an initial defect region model; the mesh division uses an adaptive algorithm, and the mesh in the high stress concentration region is made more refined by adjusting the mesh density;

[0064] In each iteration process, the inversion error is calculated and the error term in the inversion process is optimized by dynamically adjusting the regularization parameter in the inversion algorithm, and the defect region model is gradually updated iteratively to obtain three-dimensional reconstruction data with higher accuracy.

[0065] In the present invention, a three-dimensional reconstruction of the defect region is performed on the clustered abnormal signal data through a multi-scale iterative inversion method, and adaptive mesh optimization and dynamic regulation are combined with the physical model of the material to generate three-dimensional reconstruction data with high precision. This process uses the clustered abnormal signal data to reflect the defect region inside the material, continuously improves the accuracy of the inversion through iterative optimization, and combines the stress and strain characteristics of the material for mesh division and optimization, finally forming an accurate reconstruction of the micro-defects of the wind turbine blade.

[0066] Multi-scale iterative inversion is a method of gradually updating and optimizing the target model through multiple iterations, which is particularly suitable for the problem of material defect reconstruction. In the present invention, the defect region is reconstructed on the clustered abnormal signal data through multi-scale iterative inversion, and a corresponding physical model is established in combination with the physical characteristics of the material to guide and optimize the inversion process.

[0067] Specifically, first, an initial physical model of the material is constructed based on the clustered abnormal signal data. This model uses the stress-strain characteristics of the wind turbine blade material for initial mesh division, that is, according to the stress concentration degree at different positions in the material, the density of the mesh is automatically adjusted to make the mesh in the high stress concentration region more refined. This can ensure that the inversion algorithm has higher resolution in the key area, thereby improving the recognition accuracy of defects.

[0068] Next, multi-scale iterative inversion is performed. In each iteration process, the inversion algorithm calculates the inversion error according to the clustered abnormal signal data and the current defect region model, and optimizes the error term by dynamically adjusting the regularization parameter. The regularization parameter is used to balance the relationship between data fitting degree and model complexity, ensuring that the inversion result not only conforms to the known data, but also has good smoothness and robustness. During the inversion process, dynamically regulating the regularization parameter can further optimize the accuracy and stability of the inversion result, making the defect region model gradually approach the true defect morphology after each iteration.

[0069] In specific implementation, the following process can be adopted:

[0070] 1. Initial mesh generation: Based on the material properties of wind turbine blades, such as the composite structure of epoxy resin and carbon fiber, the initial mesh generation differentiates the stress-strain behavior of these different material layers. For example, in the carbon fiber layer on the blade surface, higher mesh density may be required in high-stress areas to capture micro-cracks or delaminations, while in the epoxy resin layer, relatively coarser meshes can be used.

[0071] 2. Multi-scale inversion iteration: The inversion algorithm is carried out at different scales, starting from a coarser mesh and iteratively inverting, gradually refining the mesh. During each iteration, the current inversion error is calculated, and the regularization parameter is dynamically adjusted according to the error change.

[0072] 3. Adaptive mesh optimization: During the inversion process, the mesh density is dynamically adjusted as the inversion progresses. For example, for surface cracks in wind turbine blades, as the inversion accuracy improves, the local mesh can be further refined to capture more details, while for larger-scale delamination defects, coarser meshes can be maintained to save computational resources.

[0073] Through multi-scale iterative inversion and adaptive mesh optimization, the three-dimensional morphology of the defect area can be reconstructed with high precision in a relatively short computational time. This method has a significant effect on the identification of micro-defects in wind turbine blade materials, especially in accurately locating the specific position and scope of defects in the multi-layer structure of composite materials. For example, if there are small delamination or debonding areas in the carbon fiber layer under the epoxy resin layer, this method can accurately reconstruct the defect area and provide clear three-dimensional reconstruction data. This is of great significance for evaluating the structural integrity of wind turbine blades and performing preventive maintenance.

[0074] Preferably, the adjustment of the regularization parameter is based on the following expression: where represents the regularization parameter in the th iteration; represents the regularization parameter in the th iteration; represents the error change in the current iteration, calculated as the difference between the error in this iteration and the error in the previous iteration; represents the reference error value of the initial iteration, which is used to dynamically control the regularization strength in the inversion process.

[0075] The core of the regularization parameter adjustment lies in dynamically controlling the regularization strength of the inversion through the error change. Regularization is used in inversion problems to prevent overfitting and improve the stability of the results by imposing a smoothing constraint on the inversion solution.

[0076] During each iteration, the system calculates the inversion error, which is the difference between the model prediction value and the actual observation value. If the error of the current iteration changes significantly compared to the previous iteration (i.e., is large), it indicates that the fitting effect of the model on the data has changed significantly. At this time, the regularization strength is increased (by increasing ) to prevent the risk of overfitting; on the contrary, if the error change is small, the regularization strength can be appropriately weakened to enhance the model's fitting ability for details. In this way, the dynamic adjustment of the regularization parameter achieves the balance between fitting accuracy and model smoothness during the inversion process.

[0077] In the present invention, the dynamic regulation of the regularization parameter is mainly used in the process of multi-scale iterative inversion to achieve the micro-defect reconstruction of wind turbine blade materials (such as epoxy resin and carbon fiber). The specific operation process is as follows:

[0078] 1. Initial mesh generation and inversion initialization: Before the start of multi-scale iterative inversion, the initial mesh is first generated based on the material physical model. For wind turbine blade materials, the initial mesh generation is carried out according to the stress-strain characteristics of different material layers (such as carbon fiber layer and epoxy resin layer), ensuring a higher mesh density in high stress concentration areas to improve the inversion accuracy. The initial regularization parameter is set to a moderate value to balance the quality of the initial inversion solution.

[0079] 2. Error calculation and parameter update: In each inversion iteration, the system calculates the inversion error of the current iteration and compares it with the error of the previous iteration to obtain . Based on and the initial error value , the value of the regularization parameter is adjusted according to the above formula. This update process ensures that the inversion algorithm can dynamically adapt to different error change situations, thereby optimizing the accuracy and stability of the inversion solution. For example, if the inversion error of the carbon fiber layer suddenly increases in a certain iteration, it indicates that the current model cannot well fit the defects in this area. At this time, the system will increase the regularization strength to suppress the violent fluctuation of the inversion solution, thereby ensuring the smoothness of the reconstruction.

[0080] 3. Gradually refined multi-scale inversion: As the iteration progresses, the inversion algorithm gradually refines the mesh on the basis of a coarser mesh and performs more refined defect reconstruction for local areas. This multi-scale strategy helps to improve the computational efficiency of the algorithm and ensure the global and local reconstruction accuracy. By dynamically adjusting the regularization parameter, the inversion process can maintain good fitting quality and solution stability at different scales.

[0081] Termination condition and generation of 3D reconstruction results: When the iterative process meets the preset termination condition (such as the error change is less than a certain threshold or the number of iterations reaches the upper limit), the inversion process stops and the final 3D reconstruction data is generated. The reconstructed data at this time can accurately reflect the defect areas inside the wind turbine blade and their distribution characteristics.

[0082] By dynamically adjusting the regularization parameter, the inversion algorithm can adaptively control the inversion intensity at different iterative stages, balancing the fitting accuracy and model smoothness. This method makes the present invention show higher flexibility and adaptability in detecting micro-defects of wind turbine blades. For example, in the defect identification at the junction of the carbon fiber layer and the epoxy resin layer, since there are usually complex stress concentration phenomena in such interface regions, the regularization intensity needs to be frequently adjusted during the inversion process to capture the detailed changes. The dynamically adjusted regularization parameter enables the system to better adapt to these changes, and the generated 3D reconstruction data is more accurate and delicate.

[0083] Perform spatio-temporal perturbation simulation on the final 3D reconstruction data, extract dynamic features, and perform multi-level feedback optimization based on correlation inversion to obtain defect evolution prediction data;

[0084] In the present invention, spatio-temporal perturbation simulation is performed on the final 3D reconstruction data to extract dynamic features, and multi-level feedback optimization is performed based on correlation inversion technology to predict the evolution trend of defects and generate defect evolution prediction data. This process uses the 3D reconstruction data to simulate the response characteristics of the material under different external conditions, so as to deeply analyze and predict the changes of micro-defects of the material under different time and space conditions, providing a reliable basis for the preventive maintenance of defects.

[0085] Spatio-temporal perturbation simulation is a technique for simulating material response by applying different time and space perturbation conditions. This method is used in the present invention to analyze the dynamic behavior of wind turbine blade materials (such as epoxy resin and carbon fiber layers) under various external conditions, so as to capture the evolution characteristics of their defect areas. Specifically, the spatio-temporal perturbation conditions can include changes in factors such as temperature, stress loading, and vibration, which will affect the micro-defects inside the material and cause their positions, sizes, and shapes to change over time.

[0086] In the specific implementation process, spatio-temporal perturbation conditions are first applied to the 3D reconstruction data. Taking the wind turbine blade as an example, when performing spatio-temporal perturbation simulation, different temperature changes (such as day-night temperature difference) and stress loading (such as the wind load on the blade) can be applied to the reconstructed defect area. These perturbation conditions are applied to the 3D model through numerical simulation methods to generate corresponding material response data. The response data usually includes changes in the stress field, strain field, and other physical quantities inside the material, which to a certain extent reflect the dynamic evolution trend of the internal defects of the material.

[0087] Next, extract the dynamic features under spatio-temporal perturbations. The extraction of dynamic features is achieved by analyzing the change patterns of material response data under spatio-temporal perturbation conditions. Specific implementation methods can use feature decomposition techniques, such as principal component analysis (PCA) or Fourier analysis in the spatio-temporal domain, to extract the feature vectors representing the dynamic behavior of the material. For example, for the fine cracks in the carbon fiber layer, by analyzing the crack propagation trend under temperature change and stress loading conditions, the change features of the crack propagation rate, direction, and crack shape can be extracted, and these dynamic features can be used for subsequent defect evolution prediction.

[0088] After extracting the dynamic features, the present invention performs multi-level feedback optimization based on correlation inversion. Correlation inversion is a method of reverse reasoning and analysis of complex data by combining multiple physical and mathematical models. The multi-level feedback optimization here includes multiple feedback loops, which are optimized for different dynamic features of the material respectively to improve the accuracy of defect evolution prediction. During the feedback optimization process, the system will repeatedly adjust parameters (such as the elastic modulus and thermal conductivity of the material) to make the simulation results coincide with the actually observed dynamic features, thereby correcting the errors of the prediction model and finally generating more accurate defect evolution prediction data.

[0089] Preferably, in the spatio-temporal perturbation simulation step, various perturbation conditions are applied to the final three-dimensional reconstruction data, including successively adjusting the temperature field and stress field according to a preset perturbation scheme, and solving the spatio-temporal evolution equation of the material by numerical methods to extract the dynamic features under different perturbation conditions, and the dynamic features are used to reflect the changes in the defect area.

[0090] In the present invention, various perturbation conditions are applied to the final three-dimensional reconstruction data through spatio-temporal perturbation simulation to extract dynamic features and reflect the changes in the defect area. This process aims to simulate the response behavior of the material under different external environmental conditions, so as to predict and analyze the evolution trend of material defects. For the complex structure of wind turbine blade materials (such as epoxy resin and carbon fiber layer), by adjusting perturbation conditions such as the temperature field and stress field, the deterioration process and defect expansion of the material under actual working conditions can be more comprehensively understood.

[0091] The core of spatio-temporal perturbation simulation is to simulate the response characteristics of the material by applying different external conditions (perturbation conditions), and solve the spatio-temporal evolution equation of the material by numerical methods, and then extract the dynamic features. The perturbation conditions can include temperature change, mechanical stress, vibration, and humidity change, etc. These factors will cause changes in the internal stress field, strain field, and temperature field of the material, thereby affecting the microscopic defect behavior of the material. In the present invention, the implementation steps of spatio-temporal perturbation simulation are as follows:

[0092] 1. Apply multiple perturbation conditions: Apply different perturbation conditions to the three-dimensional reconstruction data, such as changes in temperature fields and stress fields. For the materials of wind turbine blades, different perturbation situations such as day-night temperature changes, wind load changes, and humidity changes in the marine environment can be considered separately. Specifically, temperature field perturbations can be simulated by setting different temperature distribution models, such as linear temperature rise models or periodic temperature fluctuation models, to simulate the thermal expansion effects of the blades under high and low temperature conditions; while stress field perturbations can simulate the stress conditions of the blades under different wind conditions and study the influence of stress concentration on defect propagation.

[0093] 2. Numerical solution of the spatio-temporal evolution equations: After applying the perturbation conditions, it is necessary to establish the spatio-temporal evolution equations of the material to describe the response behavior of the material under the perturbation conditions. These equations usually include heat conduction equations, stress-strain equations, and momentum balance equations, etc. By using the finite element method (FEM) or other numerical solution techniques to solve these equations, the time and space response data of the material under different perturbation conditions can be obtained. Taking epoxy resin and carbon fiber layers as an example, the heat conduction equation can be used to describe the change of the temperature field and simulate the thermal expansion and contraction of the material during temperature fluctuations; the stress-strain equation can be used to describe the change of the stress concentration area when the blade is subjected to wind load and help analyze the crack propagation trend.

[0094] 3. Extract dynamic features: After obtaining the response data of the material, analyze these data to extract dynamic features. Dynamic features include the change rate, direction, and range of the defect area, etc., which are used to reflect the evolution trend of the defect under different perturbation conditions. The method of extracting dynamic features can adopt time-frequency domain analysis techniques, such as short-time Fourier transform (STFT) or wavelet transform, to analyze the change patterns of the defect area on different time scales. For example, during the crack propagation process in the carbon fiber layer, if it is found through spatio-temporal perturbation simulation that the crack propagation rate increases significantly under high stress conditions, the risk of structural strength decline in this area can be predicted.

[0095] Through spatio-temporal perturbation simulation, the present invention can effectively evaluate the defect evolution behavior of wind turbine blade materials under various environmental conditions, thereby providing important technical support for the preventive maintenance of materials and structural health monitoring. For epoxy resin and carbon fiber composite layers, spatio-temporal perturbation simulation can reveal the comprehensive influence of temperature changes and wind loads on material deterioration. For example, in the simulation, it is found that with the periodic fluctuation of temperature, the stress concentration area in the carbon fiber layer gradually expands and peels off from the interface of the epoxy resin layer. Then, by extracting the dynamic feature data, the exacerbation of interface peeling can be further predicted and the time of blade failure can be estimated.

[0096] Preferably, the associated inversion processes the multi-scale spatio-temporal feature data in a jointly optimized manner, where the weight coefficients in the weighted convolution algorithm Calculated according to the following expression: Wherein, represents the weight coefficient of the th feature layer; represents the feature importance coefficient of the th feature layer; represents the total number of all feature layers; represents the weight normalization factor of all feature layers.

[0097] In the present invention, the multi-scale spatio-temporal feature data is processed through associative inversion to optimize the accuracy and robustness of defect detection. Associative inversion is a reverse reasoning technique that combines multiple physical and mathematical models and is used to inversely deduce the specific characteristics of defects based on the response data of materials. In this process, a weighted convolution algorithm is used to jointly optimize the multi-scale features, and the weight coefficients of the weighted convolution are calculated based on the importance of the features to ensure a reasonable distribution of the contributions of different feature layers to the final result.

[0098] The core of associative inversion is to process the multi-scale spatio-temporal feature data to accurately reflect the microscopic defect distribution and evolution trend of the wind turbine blade material. The multi-scale spatio-temporal feature data includes feature information extracted from different scales and different time resolutions, such as temperature changes, stress concentration conditions, and crack propagation rates. In order to comprehensively analyze these features, a weighted convolution algorithm is adopted.

[0099] In actual operation, the determination of the feature importance coefficient can be statistically analyzed according to the performance of the feature in multiple experiments. For example, in the composite material structure of a wind turbine blade, if it is found that the temperature change data at a certain scale has a greater impact on the accuracy of crack propagation prediction, then the feature importance coefficient of this feature layer will be relatively high, thereby increasing the weight in the weighted convolution.

[0100] First, different feature layers are extracted from the multi-scale spatio-temporal data, which may include different time resolutions, spatial resolutions, and changes in physical fields (such as temperature fields and stress fields). For the detection of wind turbine blade materials, feature data can be extracted from multiple perspectives such as stress concentration under high temperature conditions, material shrinkage behavior under low temperature, and dynamic strain changes under wind loads. The extracted feature layer data calculates the feature importance coefficient according to its importance, and then calculates the weight coefficient of each feature layer according to the above expression.

[0101] After obtaining the weight coefficient After that, weighted convolution processing is performed on all feature layer data, that is, the data of each feature layer is weighted and summed according to the corresponding weights. In this way, the influence of different feature layers on the final reconstruction result will be adjusted according to the magnitude of their weights. For example, in wind turbine blade materials, if the strain change of the carbon fiber layer under high stress is more important for predicting crack propagation, then the weight of the corresponding feature layer in the weighted convolution will be relatively high, so that the final correlation inversion result will be more dependent on this feature layer.

[0102] The feature data after weighted convolution is used for correlation inversion analysis, and multi-level joint optimization is carried out on this basis. The optimization process includes repeatedly adjusting the model parameters to make the inversion result more consistent with the experimental data. This multi-level feedback optimization can continuously improve the prediction accuracy by analyzing the dynamic characteristics of defect evolution, and finally generate detailed three-dimensional reconstruction and evolution trend data of the defects.

[0103] Through the joint optimization of correlation inversion and weighted convolution, the present invention can more accurately reconstruct the defect distribution in wind turbine blade materials and predict its evolution trend. The weighted convolution algorithm can make full use of the advantages of multi-scale feature data, and the complementary effect between different feature layers is effectively exerted. For example, in the interface region between epoxy resin and carbon fiber, if the stress concentration under different temperature conditions has a significant impact on crack propagation, the weight calculation method of weighted convolution can automatically assign higher weights to these important feature layers, so that the contributions of these features are preferentially considered in the correlation inversion process. This method can better capture the various dynamic behaviors of the material in a complex environment, significantly improving the accuracy and robustness of defect identification.

[0104] Based on the defect evolution prediction data, combined with optical features, acoustic features and mechanical features, through the cross-domain adaptive learning of the multi-level feedback optimization algorithm and the neural network model, the weights and algorithm parameters of the neural network model are dynamically adjusted to achieve system update and iterative optimization.

[0105] Preferably, as Figure 4 shown, in the cross-domain adaptive learning step of the multi-level feedback optimization algorithm and the neural network model, by combining the feedback results of optical features, acoustic features and mechanical features, the weights and algorithm parameters of the neural network model are adjusted using transfer learning, where the transfer learning process realizes dynamic adjustment by comparing the similarities between optical features, acoustic features and mechanical features and updating the weights of the neural network model to optimize the overall performance of the system.

[0106] The multi-level feedback optimization algorithm combines feedback information from multiple modalities to continuously iteratively adjust the neural network model, enabling it to have higher adaptability to data of different feature modalities. In the present invention, the optical features, acoustic features, and mechanical features are respectively obtained from visual inspection, ultrasonic inspection, and stress-strain analysis of the material surface. These feature information can respectively reflect the surface topography, internal structure state, and mechanical response behavior of the material. Based on the integration of multi-modal data, through comparing the similarity between different modal features, the transfer learning process is used to update the weights of the neural network model, enabling the model to better capture the correlation between these features.

[0107] During the transfer learning process, the system first calculates the similarity between the optical features, acoustic features, and mechanical features. The calculation of similarity can adopt methods such as Euclidean distance or cosine similarity between feature vectors to measure the degree of similarity between different modalities. For example, when abnormal reflected light intensity is found in a certain area during optical inspection and significant changes in ultrasonic reflection of acoustic features are accompanied, the feature similarity between these two modalities will be relatively high, indicating that they may reflect the existence of the same defect. Based on the result of this similarity calculation, transfer learning will adjust the weights of the neural network model, enabling the model to more accurately integrate information from different modalities.

[0108] Through the multi-level feedback optimization algorithm, the feature feedback results of different modalities are used as inputs to continuously iteratively adjust the weights and algorithm parameters of the neural network. Specifically, when the changes in optical features are frequent and consistent with the strain changes in mechanical features, the weights of these two modal features in the model can be increased, thereby enhancing the sensitivity of the model to these features. This weight update strategy can be achieved through the error backpropagation technique. After each update, the system will recalculate the error of the model to determine whether the performance of the adjusted model on the new dataset has improved.

[0109] Cross-domain adaptive learning enables the neural network model to perform knowledge transfer between different modal domains (such as optics, acoustics, and mechanics). The core of this process is to use a pre-trained neural network model (a model trained on a certain modal feature) to fine-tune the data of other modalities, enabling the model to adapt to the feature distribution of the new modality. For example, assume that a detection model has been trained on optical feature data. This model can be fine-tuned using acoustic feature data, and through an appropriate transfer learning process, its adaptability to ultrasonic inspection data can be enhanced, ultimately achieving accurate identification of potential defects in the carbon fiber layer.

[0110] Through multi-level feedback optimization and transfer learning, the present invention can significantly improve the accuracy of defect detection and identification of wind turbine blade materials. The similarity calculation between multi-modal features enables the model to effectively utilize data from different sensors and improve the multi-dimensional information fusion ability for composite materials. For example, in the carbon fiber layer below the epoxy resin layer, if abnormal optical and acoustic features are detected and their change trends are similar, the system can strengthen the reconstruction and identification of this area through transfer learning, making the final defect prediction result more accurate. In addition, through cross-domain adaptive learning, the system can achieve flexible knowledge transfer and parameter adjustment between different modal data, further improving the ability to handle complex defect situations.

[0111] Preferably, during the transfer learning process, by constructing a cross-domain feature fusion model for optical features, acoustic features, and mechanical features, eigenvalue normalization processing is performed on the optical features, acoustic features, and mechanical features, and the similarity between the optical features, acoustic features, and mechanical features is calculated. According to the similarity, the weight update strategy of the neural network model is adjusted.

[0112] In the present invention, transfer learning constructs a cross-domain feature fusion model for optical features, acoustic features, and mechanical features to enhance the adaptability of the neural network model to multi-modal data. The cross-domain feature fusion model is used to normalize and calculate the similarity of features under different physical modalities, so as to adjust the weight update strategy of the neural network model according to the similarity during the transfer learning process, thereby optimizing the overall performance of the system. This process is of great significance for defect detection of wind power materials (such as epoxy resin and carbon fiber composite materials) because it can better utilize the complementarity of multi-modal data and enhance the accurate identification of defects.

[0113] The core of transfer learning lies in processing feature data from different modalities through a feature fusion model, enabling the neural network model to adapt to the complex distribution of multi-modal data. In the present invention, the optical features, acoustic features, and mechanical features respectively correspond to the visual features, ultrasonic detection features, and stress-strain behavior features of the material. To make full use of these feature information, it is necessary to first perform eigenvalue normalization processing on them to make the data scales of each modality consistent for effective fusion and similarity calculation.

[0114] The purpose of eigenvalue normalization is to eliminate the numerical range differences between different modal features, enabling them to be compared on the same scale. Specific methods can include standard normalization (subtracting the mean from the eigenvalue and then dividing by the standard deviation) or min-max normalization (mapping the eigenvalue to the range of 0 to 1). For example, for the surface reflected light intensity obtained from optical detection, there may be significant fluctuations in a large range, while the ultrasonic reflection intensity obtained from acoustic detection may be concentrated in a smaller numerical range. After normalization, these feature data can be compared and fused on the same scale.

[0115] The construction of a cross-domain feature fusion model is to integrate features of different modalities into a unified representation space. In the model, the normalized data of each modal feature serves as input, and further feature representations are generated for these data through feature extraction layers (such as convolutional layers or fully connected layers). The feature fusion layer can adopt weighted average, attention mechanism, or deep feature fusion techniques to synthesize the feature information of different modalities. For example, in the crack detection of wind turbine blades, the feature vectors of optical features (surface reflected light intensity), acoustic features (ultrasonic signals), and mechanical features (stress distribution) can be input into the fusion model, and the model will integrate this feature information to obtain a cross-domain comprehensive feature representation for subsequent analysis.

[0116] At the output layer of the feature fusion model, the similarity between optical features, acoustic features, and mechanical features is calculated to guide the process of transfer learning. The calculation of similarity can use Euclidean distance, cosine similarity, or other measurement methods to measure the similarity between different modal features. For features with higher similarity, the system will consider that these modalities may reflect the same physical phenomenon or defect characteristics, and thus appropriately increase the weights of relevant features during transfer learning. For example, if similar change patterns of optical and acoustic features are detected in the interface region between the carbon fiber layer and the epoxy resin layer of a wind turbine blade, the weights of these two modalities can be increased to make the model pay more attention to the multi-modal feature fusion results in this region.

[0117] Based on the results of similarity calculation, transfer learning optimizes the training process of the model by dynamically adjusting the weight update strategy of the neural network. Weight update can be achieved by using gradient descent method combined with error backpropagation. For each training iteration, the learning rate or weight adjustment amplitude is adjusted according to the similarity, so that features with high similarity contribute more to the training of the model, while features with low similarity have less influence.

[0118] By constructing a cross-domain feature fusion model and combining a similarity-adjusted weight update strategy in transfer learning, the present invention can significantly improve the accuracy and robustness of the microscopic defect detection of wind turbine blade materials. Normalization ensures that different modality data can be compared on the same scale, effectively eliminating the influence of numerical range differences. The construction of the feature fusion model enables the information of different physical modalities to be comprehensively analyzed in a unified feature space, thereby capturing more correlation information between multimodal data.

[0119] For example, in the carbon fiber layer of a wind turbine blade, the propagation of cracks may simultaneously cause changes in the surface optical reflection intensity and attenuation of ultrasonic signals. The feature fusion model of the present invention can capture these changes simultaneously and combine the similarity between the two to enhance the detection ability of the crack area. In this way, the system can not only identify existing defects but also predict the potential defect propagation trend, thus providing a more scientific basis for the maintenance of wind turbine blades.

[0120] As Figure 5 shown, a system for implementing the above-mentioned microscopic defect identification method of new energy materials based on computer vision, the system includes:

[0121] A multimodal data acquisition module, which is used to perform multimodal optical interference and photoacoustic imaging scans on new energy materials to collect phase interference data, polarized light data, and photoacoustic data; using multimodal optical interference and photoacoustic imaging technology, scan wind power materials (such as epoxy resin and carbon fiber composite materials) to collect phase interference data, polarized light data, and photoacoustic data. These data can reflect different physical properties of the materials, helping to obtain more comprehensive defect information.

[0122] A data preprocessing module, connected to the multimodal data acquisition module, which is used to perform feature alignment and multimodal feature fusion on the multimodal data to generate multimodal response data, and perform denoising and preprocessing on the multimodal response data to obtain preprocessed multimodal data; perform feature alignment and fusion on the collected multimodal data to generate multimodal response data, and through denoising and preprocessing, remove redundant information and retain effective features to obtain preprocessed multimodal data. This process enhances the usability of the data and lays a foundation for subsequent analysis.

[0123] The signal decoupling and anomaly detection module is connected to the data preprocessing module and is used to perform adaptive signal decoupling of the preprocessed multi-modal data for sparse representation, generate anomaly signal data, and perform non-linear enhancement and feature normalization on the anomaly signal data to obtain clustered anomaly signal data; use the sparse representation method to perform adaptive signal decoupling on the preprocessed multi-modal data to separate the anomaly signal data. Subsequently, perform non-linear enhancement and feature normalization on the anomaly signal data to obtain the clustered anomaly signal data, thereby more clearly identifying potential defects.

[0124] The multi-scale inversion and three-dimensional reconstruction module is connected to the signal decoupling and anomaly detection module and is used to perform multi-scale iterative inversion on the clustered anomaly signal data, combine with the material physical model for adaptive grid optimization and dynamic regulation, and generate the final three-dimensional reconstruction data; perform multi-scale iterative inversion on the clustered anomaly signal data, combine with the physical model of the material, and perform adaptive grid optimization and dynamic regulation to generate three-dimensional reconstruction data. This process improves the accuracy of defect localization through multi-scale analysis.

[0125] The spatio-temporal evolution modeling module is connected to the multi-scale inversion and three-dimensional reconstruction module and is used to perform spatio-temporal perturbation simulation on the final three-dimensional reconstruction data, extract dynamic features, and perform multi-level feedback optimization based on correlation inversion to obtain defect evolution prediction data; perform spatio-temporal perturbation simulation based on the three-dimensional reconstruction data to extract dynamic features. Combine with the correlation inversion technology to perform multi-level feedback optimization to generate the defect evolution prediction data, providing the possibility of the future development of material defects.

[0126] The feedback optimization and adaptive learning module is connected to the spatio-temporal evolution modeling module and is used to perform multi-level feedback optimization algorithms and cross-domain adaptive learning of neural network models based on the defect evolution prediction data, combined with optical features, acoustic features, and mechanical features, and dynamically adjust the weights and algorithm parameters of the neural network model to achieve the update and iterative optimization of the system. According to the defect evolution prediction data, combined with the feedback results of multi-modal features, use multi-level feedback optimization algorithms and cross-domain adaptive learning of neural network models to dynamically adjust the weights and algorithm parameters of the model to improve the iterative optimization effect of the system.

[0127] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0128] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for identifying microscopic defects of new energy materials based on computer vision, characterized in that: The following steps are involved: Perform multimodal optical interference and photoacoustic imaging scanning on new energy materials, collect phase interference data, polarization light data and photoacoustic data, and generate multimodal response data through multimodal feature alignment and multimodal feature fusion, and then denoise and preprocess the multimodal response data to obtain preprocessed multimodal data; Performing sparse representation of adaptive signal decoupling on preprocessed multimodal data to generate abnormal signal data, and performing nonlinear enhancement processing and feature normalization processing on the abnormal signal data to obtain clustered abnormal signal data; The defect area of ​​clustered abnormal signal data is reconstructed through multi-scale iterative inversion, and adaptive grid optimization and dynamic regulation are performed in combination with the material physical model to generate the final three-dimensional reconstruction data; Perform spatiotemporal disturbance simulation on the final 3D reconstructed data, extract dynamic features, and perform multi-level feedback optimization based on correlation inversion to obtain defect evolution prediction data; Based on the defect evolution prediction data, combined with optical, acoustic and mechanical characteristics, through the multi-level feedback optimization algorithm and cross-domain adaptive learning of the neural network model, the weights and algorithm parameters of the neural network model are dynamically adjusted to achieve system update and iterative optimization.

2. The computer vision-based new energy material micro-defect identification method according to claim 1, characterized in that: In the multimodal optical interference and photoacoustic imaging joint scanning step, by adjusting the phase angle and polarization angle, phase interference data and polarization light data are collected at different angles to cover the surface characteristics and internal structure characteristics of the new energy material; The multimodal feature alignment and multimodal feature fusion include aligning the phase interference data, polarization light data and photoacoustic data so that the phase interference data, polarization light data and photoacoustic data are aligned in the same coordinate system, and fusing the characteristic values ​​of the phase interference data, polarization light data and photoacoustic data by a weighted average method to generate multimodal response data.

3. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 1, characterized in that: In the adaptive signal decoupling step of the sparse representation, the preprocessed multimodal data is decomposed into normal signal data and abnormal signal data, wherein the sparse coefficients of the preprocessed multimodal data are solved using a sparse representation method, and the abnormal signal data is extracted by minimizing the reconstruction error, and the abnormal signal data is used for nonlinear enhancement processing.

4. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 1, characterized in that: The multi-scale iterative inversion step comprises: According to the clustered abnormal signal data, a material physical model is constructed, and initial meshing is performed based on the stress and strain characteristics of the material to generate an initial defect area model; the meshing adopts an adaptive algorithm to make the mesh in the high stress concentration area finer by adjusting the mesh density; In each iteration, the inversion error is calculated and the regularization parameters in the inversion algorithm are dynamically adjusted to optimize the error term in the inversion process. The defect area model is gradually updated in an iterative manner to obtain more accurate three-dimensional reconstruction data.

5. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 4, characterized in that: The regularization parameter The adjustment is based on the following expression: in, Indicates The regularization parameter in the iteration; Indicates The regularization parameter in the iteration; Represents the error change in the current iteration, calculated as the difference between the error of this iteration and the error of the previous iteration; Represents the baseline error value of the initial iteration, which is used to dynamically control the regularization strength during the inversion process.

6. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 4, characterized in that: In the spatiotemporal disturbance simulation step, a variety of disturbance conditions are applied to the final three-dimensional reconstructed data, including successively adjusting the temperature field and stress field according to a preset disturbance scheme, and solving the spatiotemporal evolution equation of the material by numerical methods to extract dynamic characteristics under different disturbance conditions, and the dynamic characteristics are used to reflect the changes in the defect area.

7. The computer vision-based new energy material micro-defect identification method according to claim 6, characterized in that: The correlation inversion processes multi-scale spatiotemporal feature data by means of joint optimization, wherein the weight coefficients in the weighted convolution algorithm Calculated according to the following expression: in, Indicates The weight coefficient of each feature layer; Indicates The feature importance coefficient of each feature layer; Represents the total number of all feature layers; Represents the weight normalization factor of all feature layers.

8. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 1, characterized in that: In the cross-domain adaptive learning step of the multi-level feedback optimization algorithm and the neural network model, the weights and algorithm parameters of the neural network model are adjusted by transfer learning by combining the feedback results of the optical features, acoustic features and mechanical features. The transfer learning process achieves dynamic adjustment by comparing the similarities between the optical features, acoustic features and mechanical features and updating the weights of the neural network model to optimize the overall performance of the system.

9. The method for identifying microscopic defects of new energy materials based on computer vision according to claim 8, characterized in that: In the transfer learning process, a cross-domain feature fusion model of optical features, acoustic features and mechanical features is constructed, the feature values ​​of the optical features, acoustic features and mechanical features are normalized, and the similarities between the optical features, acoustic features and mechanical features are calculated, and the weight update strategy of the neural network model is adjusted according to the similarities.

10. A system for implementing the computer vision-based new energy material micro-defect recognition method according to any one of claims 1 to 9, characterized in that: The system includes: Multimodal data acquisition module, used to perform multimodal optical interference and photoacoustic imaging scanning on new energy materials, and collect phase interference data, polarization data and photoacoustic data; A data preprocessing module, connected to the multimodal data acquisition module, is used to perform feature alignment and multimodal feature fusion on the multimodal data to generate multimodal response data, and to denoise and preprocess the multimodal response data to obtain preprocessed multimodal data; The signal decoupling and anomaly detection module is connected to the data preprocessing module and is used to perform adaptive signal decoupling of sparse representation on the preprocessed multimodal data, generate abnormal signal data, and perform nonlinear enhancement and feature normalization processing on the abnormal signal data to obtain clustered abnormal signal data; The multi-scale inversion and 3D reconstruction module is connected to the signal decoupling and anomaly detection module, and is used to perform multi-scale iterative inversion on the clustered anomaly signal data, and perform adaptive grid optimization and dynamic regulation in combination with the material physical model to generate the final 3D reconstruction data; The spatiotemporal evolution modeling module connects the multi-scale inversion and 3D reconstruction modules to simulate the spatiotemporal disturbance of the final 3D reconstruction data, extract dynamic features, and perform multi-level feedback optimization based on the correlation inversion to obtain defect evolution prediction data; The feedback optimization and adaptive learning module is connected to the spatiotemporal evolution modeling module. It is used to perform cross-domain adaptive learning of multi-level feedback optimization algorithms and neural network models based on defect evolution prediction data, combined with optical, acoustic and mechanical characteristics, and dynamically adjust the weights and algorithm parameters of the neural network model to achieve system update and iterative optimization.

Images