Material defect detection method and system based on multi-dimensional feature fusion

Abstract

This invention relates to the field of materials evaluation technology, specifically proposing a material defect detection method and system based on multi-dimensional feature fusion. The method involves synchronously acquiring thermal response data through pulse and phase-locked loop dual-modal thermal excitation, extracting four-dimensional features in the time and frequency domains, and constructing a unified pixel-level feature vector through standardized preprocessing. Then, a response map is generated through weighted fusion, and candidate abnormal regions are screened. Finally, defect determination is completed by combining dual-branch classification and decision-level fusion. This approach addresses both shallow and deep defect detection in the tested flat plate material component, effectively distinguishing between various defects such as interlayer debonding, microcracks, and porosity, solving the problems of poor accuracy and easy confusion in single thermal excitation detection. Furthermore, the standardized preprocessing process improves feature stability, enabling accurate determination of defect location and type, thus enhancing the accuracy and comprehensiveness of defect detection in the tested flat plate material component.

Description

Technical Field

[0001] This invention relates to the field of materials evaluation technology, and in particular to a method and system for detecting material defects based on multi-dimensional feature fusion. Background Technology

[0002] MI-CMC (Melt Infiltration Ceramic Matrix Composite) flat panel structures are widely used in aerospace thermal protection systems. Due to the unique nature of their woven structure and manufacturing process, various types of micro-defects are easily generated within the material, mainly including interlaminar debonding, microcracks, and pores (voids). The presence of these defects can significantly reduce the structural strength and service reliability of components; therefore, non-destructive testing is a crucial means to ensure the quality of material components.

[0003] Infrared thermal imaging is widely used due to its advantages such as non-contact operation, high detection efficiency, and good visualization. However, traditional infrared detection often uses a single thermal excitation mode, which can only extract single-dimensional thermal response features and has obvious technical shortcomings: simple pulse thermal imaging is only sensitive to surface and shallow defects and has difficulty distinguishing different types of deep defects; although simple lock-in thermal imaging has the ability to detect deep defects, its single-dimensional frequency domain features can easily confuse the thermal response signals of different defects.

[0004] In summary, infrared detection using a single thermal excitation mode suffers from poor detection accuracy. Summary of the Invention

[0005] This invention aims to at least address the technical problem of poor detection accuracy in infrared detection of flat materials using a single thermal excitation mode in related technologies. Therefore, the purpose of this invention is to propose a material defect detection method, system, and storage medium based on multi-dimensional feature fusion.

[0006] In a first aspect, the present invention proposes a material defect detection method based on multi-dimensional feature fusion, comprising the following steps: applying pulsed thermal excitation and periodic phase-locked thermal excitation to the flat plate material component to be tested, simultaneously acquiring instantaneous infrared thermogram sequences under pulsed thermal excitation and phase-locked thermal response sequences under periodic phase-locked thermal excitation; extracting instantaneous temperature difference features and temperature change rate features from the instantaneous infrared thermogram sequences, and extracting phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequences; sequentially performing noise suppression, background correction, spatial registration, and normalization processing on the instantaneous temperature difference features, the temperature change rate features, the phase-locked amplitude features, and the phase-locked phase features to construct a dual-modal basic feature vector under the same pixel coordinates; and performing weighted fusion calculation on the dual-modal basic feature vector. The system integrates response values ​​and constructs a fusion defect response map based on the fusion response values ​​of different pixels. Image segmentation and connected component analysis are performed on the fusion defect response map to extract candidate abnormal regions. These candidate abnormal regions are then expanded to obtain a region-level expanded feature vector. This region-level expanded feature vector is split into an instantaneous feature subset and a phase-locked feature subset. The instantaneous and phase-locked feature subsets are input into a preset criterion library for independent dual-branch classification. The candidate defect categories and classification confidence scores corresponding to the instantaneous and phase-locked feature branches are output. A decision-level fusion strategy is used to fuse the two classification results, outputting the defect type, defect location, and corresponding confidence score of the test area. The preset criterion library is constructed based on the region-level expanded feature vector of a standard sample through clustering and supervised classification.

[0007] According to embodiments of the present invention, the material defect detection method based on multi-dimensional feature fusion synchronously acquires thermal response data through pulse and phase-locked loop dual-mode thermal excitation, extracts four-dimensional features in the time and frequency domains, and constructs a unified pixel-level feature vector through standardized preprocessing. Then, a response map is generated through weighted fusion, and candidate abnormal regions are screened. Finally, defect determination is completed by combining dual-branch classification and decision-level fusion. This method not only takes into account the detection of shallow and deep defects in the flat plate material component under test, effectively distinguishes various defects such as interlayer debonding, microcracks, and pores, and solves the problems of poor accuracy and easy confusion of defects by single thermal excitation detection, but also improves feature stability through standardized preprocessing process, and achieves accurate determination of defect location and type, which helps to improve the accuracy and comprehensiveness of defect detection in the flat plate material component under test.

[0008] In addition, the material defect detection method based on multi-dimensional feature fusion according to embodiments of the present invention may also have the following additional technical features:

[0009] Further, the step of extracting phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequence includes: obtaining the time-temperature curve of each pixel in the phase-locked thermal response sequence; performing Fourier transform or cross-correlation operation on the time-temperature curve, and extracting the phase-locked amplitude features and phase-locked phase features from the processing result.

[0010] Furthermore, the step of performing image segmentation and connected component analysis on the fusion defect response map to extract candidate abnormal regions includes: performing image segmentation on the fusion defect response map based on an adaptive threshold to extract abnormal response regions; performing connected component analysis on the abnormal response regions to remove isolated noise points and obtain multiple connected regions; and combining minimum area constraints, aspect ratio constraints, and edge continuity constraints to select effective candidate abnormal regions from the connected regions.

[0011] Further, the step of extending the candidate anomaly region to obtain a region-level extended feature vector includes: performing statistical analysis on the bimodal basic feature vectors of all pixels within the candidate anomaly region, and combining them with the morphological parameters of the candidate anomaly region to form a region-level extended feature vector; wherein, the region-level extended feature vector includes at least instantaneous feature statistical parameters, phase-locked feature statistical parameters, and morphological parameters, wherein the instantaneous feature statistical parameters include at least one of the instantaneous temperature difference mean, instantaneous temperature difference variance, temperature change rate mean, and temperature change rate variance, wherein the phase-locked feature statistical parameters include at least one of the phase-locked amplitude mean, phase-locked amplitude variance, phase-locked phase mean, and phase-locked phase variance, and wherein the morphological parameters include at least one of the region area, aspect ratio, edge sharpness, linear continuity, connected area, and phase gradient change degree.

[0012] Furthermore, a preset criterion library is constructed based on the regional extended feature vectors of the standard sample through clustering and supervised classification, including: extracting multiple sets of the regional extended feature vectors as training samples based on the known defect regions and normal regions of the standard sample; performing unsupervised clustering analysis on the training samples based on K-means clustering to obtain natural cluster centers in the feature space; semantically labeling the clustering results based on known labeled samples, mapping each natural cluster center to the target defect type and normal region respectively, forming an initial defect feature template; and establishing a category discrimination boundary on the initial defect feature template based on a support vector machine to form the preset criterion library.

[0013] Furthermore, the decision-level fusion strategy can be any one of a voting mechanism, a weighted decision-making mechanism, or a Bayesian decision-making mechanism.

[0014] Furthermore, the material defect detection method based on multi-dimensional feature fusion also includes: laser scanning excitation supplementary verification, including: scanning and heating the detection surface of the flat plate material component under test based on a laser source, and confirming the direction, size and distribution characteristics of the defect based on the thermal wave diffusion characteristics during the heating process, so as to verify the defect type and the defect location.

[0015] Furthermore, the material of the flat plate component to be tested includes one or more of ceramic matrix composites, resin matrix composites, metal materials, and adhesive composite structures; wherein, based on the thermophysical properties of different materials, the thermal excitation parameters of the corresponding materials, the feature weight coefficients in the weighted fusion, and the preset criterion library are adapted to achieve defect detection of the corresponding materials.

[0016] Secondly, based on the same inventive concept, this invention also proposes a material defect detection system based on multi-dimensional feature fusion, comprising: an acquisition module, used to apply pulsed thermal excitation and periodic phase-locked thermal excitation to the flat plate material component under test, and simultaneously acquire instantaneous infrared thermogram sequences under pulsed thermal excitation and phase-locked thermal response sequences under periodic phase-locked thermal excitation; a construction module, used to extract instantaneous temperature difference features and temperature change rate features from the instantaneous infrared thermogram sequences, and extract phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequences, and sequentially perform noise suppression, background correction, spatial registration, and normalization processing on the instantaneous temperature difference features, the temperature change rate features, the phase-locked amplitude features, and the phase-locked phase features to construct a dual-modal basic feature vector under the same pixel coordinates; and an extraction module, used to process the dual-modal basic feature vector... The system performs weighted fusion to calculate the fusion response value, constructs a fusion defect response map based on the fusion response values ​​of different pixels, performs image segmentation and connected component analysis on the fusion defect response map, and extracts candidate abnormal regions. A detection module is used to extend the candidate abnormal regions to obtain a region-level extended feature vector, splits the region-level extended feature vector into an instantaneous feature subset and a phase-locked feature subset, inputs the instantaneous feature subset and the phase-locked feature subset into a preset criterion library for dual-branch independent classification, outputs the candidate defect category and classification confidence corresponding to the instantaneous feature branch and the phase-locked feature branch respectively, and uses a decision-level fusion strategy to fuse the two classification results, outputting the defect type, defect location, and corresponding confidence of the test area. The preset criterion library is constructed based on the region-level extended feature vector of the standard sample through clustering and supervised classification.

[0017] The material defect detection system based on multi-dimensional feature fusion provided in the second aspect has the same technical features as the material defect detection method based on multi-dimensional feature fusion provided in the first aspect. Therefore, it also has similar technical effects as the first aspect, which will not be described in detail here.

[0018] Thirdly, based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a material defect detection program based on multi-dimensional feature fusion, wherein the material defect detection program based on multi-dimensional feature fusion, when executed by a processor, implements the material defect detection method based on multi-dimensional feature fusion as described in the above embodiments of the present invention.

[0019] The computer-readable storage medium provided in the third aspect has the same technical features as the material defect detection method based on multi-dimensional feature fusion provided in the first aspect, and therefore also has similar technical effects as the first aspect, which will not be described in detail here.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a material defect detection method based on multi-dimensional feature fusion according to an embodiment of the present invention; Figure 2 This is a general flowchart of a material defect detection method based on multi-dimensional feature fusion according to a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a multimodal infrared detection platform according to a specific embodiment of the present invention; Figure 4 This is an imaging principle diagram of a multimodal infrared detection platform according to a specific embodiment of the present invention; Figure 5 This is a structural block diagram of a material defect detection system based on multi-dimensional feature fusion according to an embodiment of the present invention.

[0022] Figure label: 1-Signal generator; 2-Power amplifier; 3-Computer; 4-Infrared thermal imager; 5-First halogen lamp; 6-Two-dimensional motion platform; 7-Ceramic matrix composite plate specimen; 8-Second halogen lamp; 100-Material defect detection system based on multi-dimensional feature fusion; 110-Acquisition module; 120-Construction module; 130-Extraction module; 140-Detection module. Detailed Implementation

[0023] The embodiments of the present invention are described in detail below, and the embodiments described with reference to the accompanying drawings are exemplary.

[0024] Currently, related technologies typically employ infrared detection techniques with a single excitation mode: simple pulsed thermal imaging (instantaneous thermal imaging) is sensitive to surface and shallow defects, but it is difficult to accurately distinguish different types of deep defects; simple lock-in thermal imaging, although it can eliminate the influence of uneven surface emissivity through phase information and detect at a deeper depth, is prone to confusing hot spots generated by pores with small-area debonding signals in the absence of instantaneous feature assistance; in addition, MI-CMC materials have significant anisotropy and large differences in in-plane and interlayer thermal conductivity, and a single criterion often leads to the missed detection of "linear" defects such as cracks.

[0025] To address the above problems, this invention provides a material defect detection method, system, and storage medium based on multi-dimensional feature fusion. By fusing time-domain and frequency-domain multi-dimensional features under pulse and phase-locked loop thermal excitation, and combining pixel-level feature processing with decision-level fusion judgment, it achieves accurate identification, location, and type determination of internal defects in various material components, while improving the versatility and robustness of the detection method. See below for reference. Figures 1-5 A method, system, and storage medium for detecting material defects based on multidimensional feature fusion according to embodiments of the present invention are described.

[0026] Figure 1 This is a flowchart of a material defect detection method based on multi-dimensional feature fusion according to an embodiment of the present invention. Figure 1 As shown, a material defect detection method based on multi-dimensional feature fusion according to an embodiment of the present invention includes the following steps: Step S1: Apply pulsed thermal excitation and periodic phase-locked thermal excitation to the flat plate material component to be tested, and simultaneously acquire the instantaneous infrared thermogram sequence under pulsed thermal excitation and the phase-locked thermal response sequence under periodic phase-locked thermal excitation.

[0027] In a specific embodiment, the plate material component under test is, for example, an MI-CMC plate structure. For example, a dedicated multimodal infrared detection platform can be used to provide thermal excitation and acquire temperature field sequences. The dedicated multimodal infrared detection platform includes, for example, an infrared thermal imager, a halogen lamp array excitation source, a multidimensional motion platform, and a signal control unit. The multidimensional motion platform (e.g., a two-dimensional motion platform) is used to fix the plate material component under test. The signal control unit is used to output corresponding control signals to control the halogen lamp array excitation source to output short-time high-energy pulses or perform periodic intensity modulation (e.g., sine or square wave modulation). The infrared thermal imager is used to acquire instantaneous infrared thermal image sequences or multiple-cycle thermal response sequences during the cooling phase of the short-time high-energy pulse.

[0028] Specifically, addressing the issue that related technologies using a single thermal excitation mode can only extract single-dimensional features of flat materials, failing to meet the needs of detecting both shallow and deep defects, this invention employs dual-modal thermal excitation to simultaneously acquire two types of thermal response data. This provides a data foundation for multi-dimensional feature extraction. Pulse thermal excitation can capture the instantaneous thermal anomaly features of shallow defects, while phase-locked loop thermal excitation can capture the frequency domain thermal response features of deep defects, achieving full coverage detection of both shallow and deep defects. This solves the problems of difficulty in qualitatively identifying deep defects and inaccurate detection of shallow defects with single thermal excitation. Simultaneous acquisition ensures the spatiotemporal consistency of the two types of data, laying the foundation for subsequent pixel-level fusion of multi-dimensional features and avoiding feature registration errors caused by asynchronous acquisition.

[0029] Step S2: Extract instantaneous temperature difference features and temperature change rate features from the instantaneous infrared thermal image sequence, and extract phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequence. Perform noise suppression, background correction, spatial registration and normalization processing on the instantaneous temperature difference features, temperature change rate features, phase-locked amplitude features and phase-locked phase features in sequence to construct a dual-modal basic feature vector under the same pixel coordinates.

[0030] In specific embodiments, noise suppression refers to filtering the infrared thermal image to remove random noise from the environment and equipment, thereby improving the feature signal-to-noise ratio; background correction refers to deducting temperature shifts caused by non-defects on the component surface, eliminating uneven surface emissivity and environmental temperature interference, and highlighting the true thermal response of defects; spatial registration refers to aligning images under both pulse and phase-locked loop excitation to the same pixel coordinates, ensuring a one-to-one correspondence between different modal features and avoiding misjudgments due to misalignment; normalization refers to mapping features of different magnitudes to a unified interval, eliminating dimensional differences, and making subsequent weighted fusion and classification more stable and accurate.

[0031] Specifically, addressing the limitation of related technologies to single-dimensional feature extraction, this invention extracts four core features—instantaneous temperature difference, temperature change rate, phase-locked loop amplitude, and phase-locked loop phase—to achieve comprehensive extraction of multi-dimensional features in both the time and frequency domains. This multi-dimensional feature complementarity comprehensively characterizes the thermal response differences of different types of defects: instantaneous features reflect the instantaneous thermal response intensity and evolution of the defect; phase-locked loop features reflect the steady-state thermal response characteristics of the defect; and multi-dimensional features can accurately distinguish easily confused defect types (such as pores and interlayer debonding). Simultaneously, preprocessing the features, such as noise suppression and background correction, effectively eliminates detection errors caused by environmental interference and uneven emissivity of the material surface, improving the effectiveness of the features. Finally, a dual-modal basic feature vector is constructed based on the same pixel coordinates, achieving pixel-level fusion of multi-dimensional features. This ensures a one-to-one correspondence between time-domain and frequency-domain features at the same detection location, avoiding misjudgments of defects caused by feature misalignment.

[0032] Step S3: Calculate the fusion response value by weighted fusion of the dual-modal basic feature vectors, construct a fusion defect response map based on the fusion response values ​​of different pixels, perform image segmentation and connected component analysis on the fusion defect response map, and extract candidate abnormal regions.

[0033] In a specific embodiment, when calculating the fused response value by weighting and fusing the dual-modal basic feature vectors, the sensitivity of different defects to features in each dimension can be highlighted by the feature weight coefficients. Specifically, for example, crack defects are more sensitive to phase-locked loop features, so the weight of phase features can be increased; pore defects are more sensitive to instantaneous temperature difference features, so the weight of temperature difference features can be increased.

[0034] Specifically, for single-dimensional feature defect determination in related technologies, this embodiment of the invention calculates a fusion response value by weighting and fusing the dual-modal basic feature vectors, which can achieve targeted identification of different defects and give full play to the fusion advantages of multi-dimensional features. At the same time, by constructing a fusion defect response map based on the fusion response value, the abstract information of four-dimensional features is transformed into a visual defect response map, which helps to intuitively reflect the location and outline of the defect. Finally, candidate abnormal regions are extracted through image segmentation and connected component analysis to extract the defect region, which facilitates the accurate determination of the defect type based on the features of the defect region.

[0035] Step S4: Expand the candidate abnormal region to obtain a region-level expanded feature vector. Split the region-level expanded feature vector into an instantaneous feature subset and a phase-locked feature subset. Input the instantaneous feature subset and the phase-locked feature subset into a preset criterion library for dual-branch independent classification. Output the candidate defect category and classification confidence corresponding to the instantaneous feature branch and the phase-locked feature branch, respectively. Use a decision-level fusion strategy to fuse the two classification results and output the defect type, defect location and corresponding confidence of the test area. The preset criterion library is constructed based on the region-level expanded feature vector of the standard sample through clustering and supervised classification.

[0036] In a specific embodiment, when the regional extended feature vector is split into an instantaneous feature subset and a phase-locked feature subset, the instantaneous feature subset has high classification accuracy for shallow and surface defects, while the phase-locked feature subset has high classification accuracy for deep and internal defects. The preset criterion library is constructed based on the multi-dimensional features of standard samples and has iterative properties. The criterion library can be continuously optimized by adding new standard sample samples to adapt to the detection needs of more materials and defect types.

[0037] Specifically, addressing the issue of related technologies relying solely on single-pixel features for judgment, this invention extends pixel-level multi-dimensional features to region-level extended feature vectors. This comprehensively reflects the overall characteristics and morphological patterns of the defect region, avoiding misjudgments caused by feature noise from a single pixel. Furthermore, addressing the problem that related technologies often rely on empirical rules for defect judgment, resulting in poor objectivity and portability, this invention constructs a pre-defined criterion library based on the region-level extended feature vectors of standard samples through clustering and supervised classification. This transforms empirical criteria into data-driven objective criteria, improving judgment accuracy. Moreover, by splitting the region-level extended feature vectors into instantaneous and phase-locked feature subsets for independent dual-branch classification, this invention fully leverages the classification advantages of different dimensional features for different defects. By outputting classification results and confidence levels through dual branches, the contribution of different dimensional features to defect judgment can be quantified, providing quantitative evaluation indicators for detection results and enhancing the accuracy and reliability of the detection results.

[0038] Therefore, the material defect detection method based on multi-dimensional feature fusion according to embodiments of the present invention synchronously acquires thermal response data through pulse and phase-locked loop dual-mode thermal excitation, extracts four-dimensional features in the time and frequency domains, constructs a unified pixel-level feature vector through standardized preprocessing, generates a response map through weighted fusion, filters candidate abnormal regions, and finally completes defect determination by combining dual-branch classification and decision-level fusion. This method not only takes into account the detection of shallow and deep defects in the flat plate material component under test, effectively distinguishes various defects such as interlayer debonding, microcracks, and pores, and solves the problems of poor accuracy and easy confusion of defects by single thermal excitation detection, but also improves feature stability through standardized preprocessing process, and achieves accurate determination of defect location and type, which helps to improve the accuracy and comprehensiveness of defect detection in the flat plate material component under test.

[0039] In one embodiment of the present invention, step S2 extracts phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequence, including: obtaining the time-temperature curve of each pixel in the phase-locked thermal response sequence; performing Fourier transform or cross-correlation operation on the time-temperature curve, and extracting phase-locked amplitude features and phase-locked phase features from the processing result.

[0040] In a specific embodiment, cross-correlation demodulation is a classic method for extracting amplitude and phase features in phase-locked thermal imaging. It is equivalent to Fourier transform. The core is to perform cross-correlation calculation between the time-temperature response signal of the pixel under test and the reference signal of thermal excitation to demodulate the amplitude (phase-locked amplitude) and phase lag (phase-locked phase) of the response signal, which is adapted to the fast demodulation requirements of engineering.

[0041] Specifically, the embodiments of the present invention perform Fourier transform or cross-correlation operations on the time-temperature curve, which not only ensures the accuracy of frequency domain feature extraction in multidimensional features, but also expands the scope of application: Fourier transform is suitable for high-precision, slow-speed detection scenarios, while cross-correlation operations are suitable for engineering, fast detection scenarios. The appropriate method can be flexibly selected according to actual detection needs. At the same time, both methods can effectively eliminate the interference of thermal excitation signals, improve the extraction accuracy of phase-locked features, and ensure the effectiveness of multidimensional features.

[0042] In one embodiment of the present invention, step S3 performs image segmentation and connected component analysis on the fusion defect response map to extract candidate abnormal regions, including: performing image segmentation on the fusion defect response map based on an adaptive threshold to extract abnormal response regions; performing connected component analysis on the abnormal response regions to remove isolated noise points and obtain multiple connected regions; and combining minimum area constraints, aspect ratio constraints, and edge continuity constraints to select effective candidate abnormal regions from the connected regions.

[0043] In specific embodiments, adaptive threshold segmentation automatically calculates and determines the segmentation threshold based on the grayscale distribution of the fused defect response map itself, rather than using a fixed threshold. This automatically adapts to differences in thermal response under different materials, defects, and environments, accurately distinguishing defect areas from normal areas and reducing missegmentation and missed segmentation. Connected component analysis divides the abnormal regions obtained after image segmentation into independent, continuous connected regions based on pixel adjacency. This aggregates discrete abnormal points into complete defect regions while identifying and removing isolated noise points, providing basic units for subsequent screening. Minimum area constraint sets a lower limit for the area, retaining only areas larger than this limit. The product of connected regions can eliminate small invalid regions such as noise and small interference points, reducing the amount of subsequent calculations and avoiding the identification of non-defect interference as defects. The aspect ratio constraint limits the range of the length-to-width ratio of the connected region based on the geometric characteristics of the defect. It can distinguish different types of defects (such as cracks being long and thin, and holes being nearly circular), filter interference regions with obviously abnormal shapes, and improve the accuracy of defect identification. The edge continuity constraint determines whether the contour of the connected region is continuous and smooth, and eliminates interference regions with broken or discontinuous edges. It can preserve the continuous contour of the real defect, eliminate noisy regions with messy and irregular edges, and improve the reliability of candidate abnormal regions.

[0044] Specifically, the embodiments of the present invention extract candidate abnormal regions through adaptive threshold segmentation, connected component analysis, and multi-constraint screening. The adaptive threshold can adapt to the multi-dimensional feature response patterns of different materials and different detection environments, avoiding missed or false detections caused by fixed thresholds. Multi-constraint screening of connected regions can effectively remove invalid connected regions corresponding to isolated noise points, ensuring the effectiveness of candidate abnormal regions, significantly reducing the computational load of subsequent classification, and improving detection efficiency.

[0045] In one embodiment of the present invention, step S4 expands the candidate anomaly region to obtain a region-level expanded feature vector, including: performing statistical analysis on the bimodal basic feature vectors of all pixels within the candidate anomaly region, and combining them with the morphological parameters of the candidate anomaly region to form a region-level expanded feature vector; wherein, the region-level expanded feature vector includes at least instantaneous feature statistical parameters, phase-locked feature statistical parameters, and morphological parameters, the instantaneous feature statistical parameters include at least one of the instantaneous temperature difference mean, instantaneous temperature difference variance, temperature change rate mean, and temperature change rate variance, the phase-locked feature statistical parameters include at least one of the phase-locked amplitude mean, phase-locked amplitude variance, phase-locked phase mean, and phase-locked phase variance, and the morphological parameters include at least one of the region area, aspect ratio, edge sharpness, linear continuity, connected area, and phase gradient change degree.

[0046] Specifically, this embodiment of the invention configures the regional extended feature vector to integrate feature statistical parameters and morphological parameters. The statistical parameters can reflect the overall multidimensional thermal response characteristics of the defect region, while the morphological parameters can reflect the geometric characteristics of the defect. This achieves a dual representation of multidimensional thermal response characteristics and geometric characteristics, providing a more comprehensive feature basis for accurate defect classification. At the same time, the diversity of feature parameters can adapt to the differences in defect characteristics of different materials, providing a foundation for multi-material detection and helping to solve the problem that single-dimensional features cannot be adapted to multi-material detection.

[0047] In one embodiment of the present invention, a preset criterion library is constructed based on the regional extended feature vectors of a standard sample through clustering and supervised classification. This includes: extracting multiple sets of regional extended feature vectors as training samples based on the known defect regions and normal regions of the standard sample; performing unsupervised clustering analysis on the training samples based on K-means clustering to obtain natural cluster centers in the feature space; semantically labeling the clustering results based on known labeled samples, mapping each natural cluster center to the target defect type and normal region respectively to form an initial defect feature template; and establishing a category discrimination boundary on the initial defect feature template based on a support vector machine to form a preset criterion library.

[0048] In a specific embodiment, K-means clustering is an unsupervised machine learning algorithm that does not rely on sample labels. It automatically divides a large number of high-dimensional feature samples into a preset number of clusters by calculating the distance similarity between feature vectors, minimizing the feature differences within the same cluster and maximizing the feature differences between different clusters. Finally, the central features of each cluster are extracted as the cluster center. Its function is to: expand the feature vectors at the regional level for standard samples, explore the natural distribution patterns in the multi-dimensional feature space, automatically classify defects and normal areas with similar features, quickly distinguish feature clusters corresponding to different defect types, form an initial feature classification template, provide a standardized feature benchmark for subsequent supervised classification, solve the problem that it is impossible to directly construct discrimination rules for unlabeled samples, and at the same time, it fits the real distribution patterns of actual defect features, avoiding the subjective bias of manual experience classification.

[0049] In a specific embodiment, SVM (Support Vector Machine) is a supervised machine learning algorithm that constructs an optimal classification boundary in a high-dimensional feature space based on sample data with known labels, accurately separating feature samples of different categories. Even when faced with small samples and high-dimensional feature data, it can establish stable classification rules. Its function is to optimize and construct a clear defect category discrimination boundary based on the initial feature template obtained by K-means clustering, combined with known defect label samples, forming a standardized and reusable preset criterion library. For high-dimensional regional extended features, it can achieve accurate classification and determination between defect areas and normal areas, as well as between different defect types, making up for the shortcomings of single clustering algorithms in accurately labeling categories, improving the accuracy and robustness of defect classification, and adapting to high-dimensional feature classification scenarios of various material defects.

[0050] Specifically, this embodiment of the invention employs K-means clustering combined with support vector machine (SVM) to construct a pre-defined criterion library based on multi-dimensional features, transforming empirical criteria into data-driven objective criteria: K-means clustering performs unsupervised analysis on unlabeled samples, uncovering natural clustering patterns in the multi-dimensional feature space, closely matching the true feature distribution of different defects; semantic annotation is performed using known labeled samples, achieving accurate mapping between cluster centers and defect types; SVM can establish clear category discrimination boundaries in the high-dimensional feature space, effectively solving the defect classification problem under small samples and high-dimensional features, and improving the classification accuracy of the criterion library.

[0051] In one embodiment of the present invention, the decision-level fusion strategy is any one of a voting mechanism, a weighted decision mechanism, or a Bayesian decision mechanism.

[0052] In a specific embodiment, the voting mechanism is a basic majority decision-making mechanism. The instantaneous feature branch and the phase-locked feature branch are treated as two independent voting entities. The candidate defect categories output by the two branches are compared, and the category with the majority agreement is used as the final judgment result. If the categories of the two branches are consistent, the defect type is directly determined. If there is a category divergence, a preset priority or secondary verification can be used for judgment. The voting mechanism has simple logic and high computational efficiency, requiring no additional weight parameters. It is suitable for rapid detection scenarios in engineering fields and can effectively eliminate misjudgments caused by feature noise and interference from a single branch. Through dual-branch cross-validation, the stability of the defect classification results is initially guaranteed, reducing classification errors caused by random interference.

[0053] In a specific embodiment, the weighted decision mechanism assigns corresponding decision weights to the instantaneous feature branch and the phase-locked loop feature branch based on the classification reliability of different feature branches. It then performs a weighted sum of the confidence scores for each category output by both branches, selecting the defect category with the highest weighted score as the final result. The weights can be pre-calibrated or adaptively adjusted based on the material's thermophysical properties, defect type, and detection scenario, with a total weight sum of 1. The weighted decision mechanism specifically highlights the classification results of highly reliable branches, adapting to the varying sensitivity of different materials and defects to the two types of features. For example, it increases the weight of the phase-locked loop feature branch when detecting deep defects and increases the weight of the instantaneous feature branch when detecting shallow defects. Compared to the voting mechanism, it is more flexible, has higher classification accuracy, and can maximize the advantages of multi-dimensional feature fusion to achieve precise classification.

[0054] In a specific embodiment, the Bayesian decision mechanism is an optimal decision-making mechanism based on probability statistics and Bayes' theorem. It combines prior probabilities (the probability of each type of defect appearing in the component under test) with the conditional probabilities of bi-branch classification to calculate the posterior probability of each type of defect. The defect category with the highest posterior probability is selected as the final judgment result, achieving optimal decision-making based on probability statistics. The Bayesian decision mechanism is a high-precision decision fusion method that fully considers the prior patterns of defect distribution and the confidence probability of bi-branch classification. It can effectively handle complex scenarios with significant discrepancies in bi-branch classification results and ambiguous defect features, reducing the missed detection rate of low-probability defects. It is suitable for scenarios with extremely high requirements for classification accuracy, such as high-precision detection and defect determination of complex components, improving the scientific validity and reliability of classification results.

[0055] Specifically, the embodiments of the present invention employ voting, weighted, or Bayesian decision-level fusion strategies to fuse classification results. This can combine the advantages of dual-branch judgment and make up for the shortcomings of single-dimensional feature classification. For example, a region that is misjudged as a hole in an instantaneous branch can be determined as a crack through the phase features of the phase-locked branch. The fusion achieves accurate determination of the defect type.

[0056] In one embodiment of the present invention, the material defect detection method based on multi-dimensional feature fusion further includes: laser scanning excitation supplementary verification, including: scanning and heating the detection surface of the flat plate material component under test based on a laser source, and confirming the direction, size and distribution characteristics of the defect based on the thermal wave diffusion characteristics during the heating process, so as to verify the defect type and defect location.

[0057] Specifically, the embodiments of the present invention supplement the verification by adding laser scanning excitation, which can perform secondary verification on the core detection results obtained based on multi-dimensional feature fusion, further improving the accuracy of defect judgment; and the thermal wave diffusion characteristics of laser line scanning heating can accurately reflect the direction, size and distribution characteristics of linear defects such as cracks, making up for the problem that the multi-dimensional features of infrared thermal imaging are not clear in characterizing the geometric features of cracks; at the same time, the high directionality of laser excitation can accurately focus and detect tiny defects, and combined with multi-dimensional feature fusion, it can effectively reduce the false negative rate of tiny defects, thereby improving the robustness of the detection method.

[0058] In one embodiment of the present invention, the material of the flat plate material component to be tested includes one or more of ceramic matrix composites, resin matrix composites, metal materials, and adhesive composite structures; wherein, according to the thermophysical properties of different materials, the thermal excitation parameters of the corresponding materials, the feature weight coefficients in the weighted fusion, and the preset criterion library are adapted to achieve defect detection of the corresponding materials.

[0059] In specific embodiments, for example, for resin-based composite materials with low thermal conductivity, the instantaneous characteristic weight can be increased and the high-frequency parameters of the phase-locked excitation can be adjusted; for metal materials with high thermal conductivity, the phase-locked characteristic weight can be increased and high-energy short-time pulse excitation can be used.

[0060] Specifically, the embodiments of the present invention specify that the flat plate material components to be tested include ceramic-based, resin-based, metal materials and adhesive composite structures. Based on the thermophysical properties of different materials, thermal excitation parameters, multi-dimensional feature weight coefficients and preset criterion libraries are specifically adapted to achieve the universality of one method for multi-material detection. Through customized parameters and a criterion library based on multi-dimensional features, the detection accuracy of different materials is guaranteed, solving the problem of requiring multiple methods for multi-material detection, thereby reducing detection costs.

[0061] The following describes the material defect detection method based on multi-dimensional feature fusion of the above embodiments of the present invention in further detail with reference to a specific embodiment. In this specific embodiment, a material defect detection method based on multi-dimensional feature fusion is provided, wherein the flat plate material component to be tested is an MI-CMC flat plate component.

[0062] Figure 2 This is a general flowchart of a material defect detection method based on multi-dimensional feature fusion according to a specific embodiment of the present invention. Figure 2As shown in this specific embodiment, the material defect detection method based on multi-dimensional feature fusion includes the following steps: Step S10: Thermal excitation and data acquisition.

[0063] Step S20: Multidimensional feature extraction and construction of bimodal basic feature vectors.

[0064] Step S30: Fusion response and candidate region extraction.

[0065] Step S40: Feature expansion and region-level expanded feature vector construction.

[0066] Step S50: Construct the preset criterion library.

[0067] Step S60: Dual-branch classification and decision-level fusion determination.

[0068] Step S70: Laser scanning stimulation supplementary verification.

[0069] Figure 3 This is a schematic diagram of the structure of a multimodal infrared detection platform according to a specific embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the imaging principle of a multimodal infrared detection platform according to a specific embodiment of the present invention. Figure 3 and Figure 4 As shown, in this specific embodiment, step S10, thermal excitation and data acquisition, includes: Step S11: Construct a multimodal infrared detection platform. The multimodal infrared detection platform includes a computer 3, an infrared thermal imager 4 with a thermal sensitivity of <50mK and a resolution of 384×288, a 1000W halogen lamp excitation source, a two-dimensional motion platform 6, and a signal control unit. The computer 3 includes a data processing unit, the signal control unit includes a signal generator 1 and a power amplifier 2, and the halogen lamp excitation source includes a first halogen lamp 5 and a second halogen lamp 8.

[0070] Step S12: Fix the ceramic matrix composite plate specimen 7 on the two-dimensional motion platform 6, arrange the first halogen lamp 5 and the second halogen lamp 8 on both sides of the normal direction of the ceramic matrix composite plate specimen 7, and place the infrared thermal imager 4 directly in front of the detection surface of the ceramic matrix composite plate specimen 7.

[0071] Step S13: The halogen lamp excitation source is controlled by the signal control unit to apply pulsed thermal excitation and periodic phase-locked thermal excitation respectively. The pulsed thermal excitation is a short-duration high-energy pulse, and the phase-locked thermal excitation adopts 0.1Hz square wave modulation (heating for 5s, cooling for 5s). The instantaneous infrared thermal image sequence under pulsed thermal excitation and the phase-locked thermal response sequence under phase-locked thermal excitation are acquired simultaneously and stored as .csv format files to provide basic data for multidimensional feature extraction.

[0072] In this specific embodiment, step S20, multidimensional feature extraction and dual-modal basic feature vector construction, includes: Step S21: Use Python to write a data processing algorithm to extract the instantaneous temperature difference feature of each pixel from the instantaneous infrared thermal image sequence. and temperature change rate characteristics .

[0073] Step S22: Obtain the time-temperature curve of each pixel from the phase-locked thermal response sequence, perform a discrete Fourier transform on the curve, and extract the phase-locked amplitude features. Phase-locked loop characteristics This completes the extraction of four-dimensional core features in the time and frequency domains.

[0074] Step S23: Perform Gaussian filtering noise suppression, background temperature correction, pixel coordinate space registration, and normalization on the four-dimensional features in sequence to construct a dual-modal basic feature vector under the same pixel coordinates. This enables pixel-level fusion of multi-dimensional features.

[0075] ,in, Characterized by instantaneous temperature difference. Characteristic of temperature change rate For phase-locked loop amplitude characteristics, This represents the phase-locked loop characteristic.

[0076] In this specific embodiment, step S30, fusion response and candidate region extraction, includes: Step S31: Perform weighted fusion of the bimodal basic feature vectors and calculate the fusion response value. : ,in, Let be the feature weight coefficients for the corresponding features, and satisfy: .

[0077] Step S32: Construct a fusion defect response map based on the fusion response values ​​of all pixels, and perform image segmentation on the response map based on the adaptive Otsu threshold to extract abnormal response regions. Specifically, different defects have different distribution patterns in the feature space: interlayer debonding usually manifests as a large instantaneous temperature difference, a low thermal decay rate, a high phase-locked amplitude, and obvious phase lag; cracks usually manifest as an insignificant instantaneous temperature difference but a significant abrupt change in the phase-locked phase gradient, combined with a high aspect ratio and linear connectivity; pores usually manifest as localized abnormal instantaneous temperature differences and relatively weak phase-locked phase lag, forming spots.

[0078] Step S33: Perform connected component analysis on the abnormal response region, remove isolated noise points with less than 50 pixels, and obtain multiple connected regions.

[0079] Step S34: Combining minimum area constraint (area > 0.1 mm²), aspect ratio constraint (crack aspect ratio > 5, hole aspect ratio < 3), and edge continuity constraint, select valid candidate abnormal regions from the connected regions.

[0080] In this specific embodiment, step S40, feature expansion and region-level expanded feature vector construction, includes: Step S41: Perform statistical analysis on the bimodal basic feature vectors of all pixels in each candidate anomaly region, and calculate multidimensional feature statistical parameters such as the mean / variance of instantaneous temperature difference, the mean / variance of temperature change rate, the mean / variance of phase-locked amplitude, and the mean / variance of phase-locked phase.

[0081] Step S42: Extract morphological parameters such as region area, aspect ratio, edge sharpness, and phase gradient change of candidate anomaly regions.

[0082] Step S43: Concatenate the statistical parameters and morphological parameters to form a regional extended feature vector, and split it into an instantaneous feature subset (including instantaneous temperature difference and temperature change rate related parameters) and a phase-locked feature subset (including phase-locked amplitude and phase related parameters).

[0083] In this specific embodiment, step S50, constructing the preset criterion library, includes: Step S51: Prepare standard specimens of ceramic matrix composites containing known defects such as interlaminar debonding, cracks, and pores, and extract the regional extended feature vectors of the known defect areas and normal areas of the standard specimens as training samples.

[0084] Step S52: Use K-means clustering to perform unsupervised clustering on the training samples to obtain natural cluster centers for 3 types of defective regions and 1 type of normal region.

[0085] Step S53: Combine known label samples to semantically annotate the cluster centers and map them as interlayer debonding, cracks, holes, and normal areas to form an initial defect feature template.

[0086] Step S54: Use Support Vector Machine (SVM) to establish category discrimination boundaries for the initial defect feature template and construct a pre-defined criterion library for ceramic matrix composites based on multi-dimensional features.

[0087] In this specific embodiment, step S60, dual-branch classification and decision-level fusion determination, includes: Step S61: Input the instantaneous feature subset and the phase-locked feature subset into the preset criterion library for dual-branch independent classification, and output the candidate defect categories and classification confidence of the instantaneous feature branch and the phase-locked feature branch.

[0088] Step S62: Use a weighted decision-making mechanism to perform fusion and calculate the final defect category. : ， in, Indicates the final defect category. Indicates the instantaneous feature branch for the first Classification confidence of defects Represents the phase-locked characteristic branch pair of the first Classification confidence of defects and These are the decision weights for the corresponding branches.

[0089] Step S63: Output the specific location, type (interlayer debonding / crack / void) and corresponding confidence level of the defect.

[0090] In this specific embodiment, step S70, laser scanning excitation supplementary verification, includes: fixing the ceramic matrix composite plate specimen 7 on the two-dimensional motion platform 6, using a line laser source to scan and heat the candidate defect area line by line, acquiring thermal wave diffusion images during the scanning process, analyzing the thermal wave diffusion characteristics, verifying the direction, size, and distribution characteristics of the crack defect, and performing secondary confirmation on the detection results obtained based on multi-dimensional feature fusion to correct deviations.

[0091] As can be seen in this specific embodiment, the material defect detection method based on multi-dimensional feature fusion achieves multi-dimensional feature extraction through dual-modal thermal excitation. Combined with pixel-level feature fusion, region-level feature expansion, and decision-level fusion judgment, it constructs a complete material defect detection system based on multi-dimensional feature fusion. Compared with related technologies, it has the following significant advantages: 1. High accuracy of defect detection: By fusing multi-dimensional thermal response features in the time and frequency domains, it achieves full coverage detection of shallow and deep defects, solving the problems of missed crack detection and confusion between pores and debonding due to single-dimensional features; through bi-branch classification and decision-level fusion judgment, it fully leverages the classification advantages of different dimensional features, significantly improving the accuracy of defect classification and reducing the missed detection rate and false detection rate.

[0092] 2. High versatility: It can be adapted to the detection of various materials such as ceramic matrix, resin matrix, metal materials and adhesive composite structures. By adapting thermal excitation parameters, multi-dimensional feature weights and criterion libraries according to the thermophysical properties of the materials, it can ensure the detection accuracy of different materials and solve the problem of poor multi-material compatibility of existing technologies.

[0093] 3. Automated and efficient detection process: The modular detection system and digital program enable automated execution of the detection process, avoiding human error; pixel-level multi-dimensional feature processing and multi-constraint candidate region extraction significantly reduce subsequent computation and improve detection efficiency, making it suitable for industrial batch detection.

[0094] 4. Objective and iterative criteria: K-means clustering and SVM are used to build a data-driven pre-set criterion library based on multi-dimensional features, which transforms empirical rules into objective feature discrimination boundaries. At the same time, the criterion library can be continuously optimized by adding new standard sample samples to adapt to the detection needs of more materials and defect types.

[0095] 5. High reliability of detection results: Through laser scanning excitation for supplementary verification, the core detection results obtained based on multi-dimensional feature fusion can be confirmed a second time, accurately characterizing the direction, size and distribution characteristics of defects; at the same time, the confidence index of defects is output, providing a quantitative evaluation of the detection results and improving the robustness and reliability of the detection results.

[0096] A further embodiment of the present invention discloses a material defect detection system based on multi-dimensional feature fusion. Figure 5 This is a structural block diagram of a material defect detection system based on multi-dimensional feature fusion according to an embodiment of the present invention. Figure 5 As shown, in one embodiment of the present invention, the material defect detection system 100 based on multi-dimensional feature fusion includes: an acquisition module 110, a construction module 120, an extraction module 130, and a detection module 140.

[0097] Specifically, the acquisition module 110 is used to apply pulsed thermal excitation and periodic phase-locked thermal excitation to the flat plate material component under test, and simultaneously acquire the instantaneous infrared thermogram sequence under pulsed thermal excitation and the phase-locked thermal response sequence under periodic phase-locked thermal excitation.

[0098] The construction module 120 is used to extract instantaneous temperature difference features and temperature change rate features from instantaneous infrared thermal image sequences, and to extract phase-locked amplitude features and phase-locked phase features from phase-locked thermal response sequences. The instantaneous temperature difference features, temperature change rate features, phase-locked amplitude features, and phase-locked phase features are sequentially subjected to noise suppression, background correction, spatial registration, and normalization processing to construct a dual-modal basic feature vector under the same pixel coordinates.

[0099] The extraction module 130 is used to perform weighted fusion calculation of the fusion response value on the dual-modal basic feature vector, construct a fusion defect response map based on the fusion response value of different pixels, perform image segmentation and connected component analysis on the fusion defect response map, and extract candidate abnormal regions.

[0100] The detection module 140 is used to expand the candidate abnormal region to obtain a region-level expanded feature vector. The region-level expanded feature vector is split into an instantaneous feature subset and a phase-locked feature subset. The instantaneous feature subset and the phase-locked feature subset are input into a preset criterion library for dual-branch independent classification. The candidate defect category and classification confidence corresponding to the instantaneous feature branch and the phase-locked feature branch are output respectively. A decision-level fusion strategy is used to fuse the two classification results and output the defect type, defect location and corresponding confidence of the test area. The preset criterion library is constructed based on the region-level expanded feature vector of the standard sample through clustering and supervised classification.

[0101] It should be noted that the specific implementation of the material defect detection system 100 based on multi-dimensional feature fusion in the embodiments of the present invention is similar to the specific implementation of the material defect detection method based on multi-dimensional feature fusion described in the above embodiments of the present invention, and therefore has similar technical effects. For details, please refer to the description of the material defect detection method based on multi-dimensional feature fusion. To reduce redundancy, it will not be repeated here.

[0102] A further embodiment of the present invention discloses a computer-readable storage medium storing a material defect detection program based on multidimensional feature fusion. When the material defect detection program based on multidimensional feature fusion is executed by a processor, it implements the material defect detection method based on multidimensional feature fusion as described in any of the above embodiments of the present invention.

[0103] It should be noted that the specific implementation of the computer-readable storage medium in the embodiments of the present invention is similar to the specific implementation described in the material defect detection method based on multi-dimensional feature fusion in the above embodiments of the present invention, and therefore has similar technical effects. For details, please refer to the description in the material defect detection method based on multi-dimensional feature fusion section. To reduce redundancy, it will not be repeated here.

[0104] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0105] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A material defect detection method based on multi-dimensional feature fusion, characterized in that, Includes the following steps: Pulsed thermal excitation and periodic phase-locked thermal excitation were applied to the flat plate material component under test, and the instantaneous infrared thermogram sequence under pulsed thermal excitation and the phase-locked thermal response sequence under periodic phase-locked thermal excitation were acquired simultaneously. Instantaneous temperature difference features and temperature change rate features are extracted from the instantaneous infrared thermal image sequence, and phase-locked amplitude features and phase-locked phase features are extracted from the phase-locked thermal response sequence. The instantaneous temperature difference features, temperature change rate features, phase-locked amplitude features, and phase-locked phase features are sequentially subjected to noise suppression, background correction, spatial registration, and normalization to construct a dual-modal basic feature vector under the same pixel coordinates. The weighted fusion of the dual-modal basic feature vectors is used to calculate the fusion response value. A fusion defect response map is constructed based on the fusion response values ​​of different pixels. Image segmentation and connected component analysis are performed on the fusion defect response map to extract candidate abnormal regions. The candidate abnormal region is extended to obtain a region-level extended feature vector. The region-level extended feature vector is split into an instantaneous feature subset and a phase-locked feature subset. The instantaneous feature subset and the phase-locked feature subset are input into a preset criterion library for dual-branch independent classification. The candidate defect category and classification confidence corresponding to the instantaneous feature branch and the phase-locked feature branch are output respectively. A decision-level fusion strategy is used to fuse the two classification results and output the defect type, defect location and corresponding confidence of the test area. The preset criterion library is constructed based on the region-level extended feature vector of the standard sample through clustering and supervised classification.

2. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The extraction of phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequence includes: Obtain the time-temperature curve of each pixel in the phase-locked thermal response sequence; The time-temperature curve is subjected to Fourier transform or cross-correlation operation, and the phase-locked amplitude feature and the phase-locked phase feature are extracted from the processing result.

3. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The step of performing image segmentation and connected component analysis on the fusion defect response map to extract candidate abnormal regions includes: Image segmentation is performed on the fused defect response map based on an adaptive threshold to extract abnormal response regions; Connectivity analysis is performed on the abnormal response region to remove isolated noise points and obtain multiple connected regions; By combining minimum area constraints, aspect ratio constraints, and edge continuity constraints, valid candidate abnormal regions are selected from the connected regions.

4. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The step of expanding the candidate anomaly region to obtain a region-level expanded feature vector includes: Statistical analysis is performed on the bimodal basic feature vectors of all pixels within the candidate anomaly region, and combined with the morphological parameters of the candidate anomaly region, to form a region-level extended feature vector.

5. The material defect detection method based on multi-dimensional feature fusion according to claim 4, characterized in that, The region-level extended feature vector includes at least instantaneous feature statistical parameters, phase-locked feature statistical parameters, and morphological parameters. The instantaneous feature statistical parameters include at least one of the following: instantaneous temperature difference mean, instantaneous temperature difference variance, temperature change rate mean, and temperature change rate variance. The phase-locked feature statistical parameters include at least one of the following: phase-locked amplitude mean, phase-locked amplitude variance, phase-locked phase mean, and phase-locked phase variance. The morphological parameters include at least one of the following: region area, aspect ratio, edge sharpness, linear continuity, connected area, and phase gradient change degree.

6. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, A pre-defined criterion library is constructed based on the regional extended feature vectors of standard samples through clustering and supervised classification, including: Based on the known defect and normal regions of the standard specimen, multiple sets of regional extended feature vectors are extracted as training samples. Unsupervised clustering analysis is performed on the training samples based on K-means clustering to obtain natural cluster centers in the feature space. Based on known labeled samples, semantic annotation is performed on the clustering results, and each natural cluster center is mapped to the target defect type and normal region respectively to form an initial defect feature template. The initial defect feature template is used to establish a category discrimination boundary based on the support vector machine, thus forming the preset criterion library.

7. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The decision-level fusion strategy can be any one of a voting mechanism, a weighted decision mechanism, or a Bayesian decision mechanism.

8. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, Also includes: Laser scanning excitation supplementary verification, including: The test surface of the flat plate material component under test is scanned and heated by a laser source. Based on the thermal wave diffusion characteristics during the heating process, the direction, size and distribution characteristics of the defects are confirmed to verify the defect type and the defect location.

9. The material defect detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The material of the flat plate component to be tested includes one or more of the following: ceramic matrix composites, resin matrix composites, metallic materials, and adhesive composite structures; wherein... Based on the thermophysical properties of different materials, the corresponding thermal excitation parameters, feature weight coefficients in weighted fusion, and preset criterion libraries are adapted to achieve defect detection of the corresponding materials.

10. A material defect detection system based on multi-dimensional feature fusion, characterized in that, include: The acquisition module is used to apply pulsed thermal excitation and periodic phase-locked thermal excitation to the flat plate material component under test, and simultaneously acquire the instantaneous infrared thermogram sequence under pulsed thermal excitation and the phase-locked thermal response sequence under periodic phase-locked thermal excitation. The module is used to extract instantaneous temperature difference features and temperature change rate features from the instantaneous infrared thermal image sequence, and to extract phase-locked amplitude features and phase-locked phase features from the phase-locked thermal response sequence. The instantaneous temperature difference features, temperature change rate features, phase-locked amplitude features, and phase-locked phase features are sequentially subjected to noise suppression, background correction, spatial registration, and normalization processing to construct a dual-modal basic feature vector under the same pixel coordinates. The extraction module is used to perform weighted fusion calculation of the fusion response value on the dual-modal basic feature vector, construct a fusion defect response map based on the fusion response value of different pixels, perform image segmentation and connected component analysis on the fusion defect response map, and extract candidate abnormal regions. The detection module is used to expand the candidate abnormal region to obtain a region-level expanded feature vector, split the region-level expanded feature vector into an instantaneous feature subset and a phase-locked feature subset, input the instantaneous feature subset and the phase-locked feature subset into a preset criterion library for dual-branch independent classification, output the candidate defect category and classification confidence corresponding to the instantaneous feature branch and the phase-locked feature branch respectively, and use a decision-level fusion strategy to fuse the two classification results to output the defect type, defect location and corresponding confidence of the test area; wherein, the preset criterion library is constructed based on the region-level expanded feature vector of the standard sample through clustering and supervised classification.

Images