A method for defect detection in eddy current pulse thermography based on weighted principal component analysis
Patent Information
- Application Number
- CN202610775878.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-01
- Publication Date
- 2026-08-18
AI Technical Summary
[0008]本发明的目的在于克服现有主成分热成像方法对弱缺陷检测能力不足的缺陷，提出一种基于加权主成分分析的涡流脉冲热成像缺陷检测方法，将盲统计分析转化为以物理先验为导向的目标化分析，显著提升弱缺陷的检测信噪比
（1）本发明利用缺陷区域时间积分热能显著高于背景区域这一物理规律，构建热显著性空间权重图，将物理先验知识融入主成分分析框架，使统计分析聚焦于缺陷相关区域，克服了传统PCT盲统计的局限性，显著提升了弱缺陷的检测信噪比。
Smart Images
Figure CN122597366A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of eddy current pulse thermal imaging nondestructive testing technology, specifically an eddy current pulse thermal imaging defect detection method based on weighted principal component analysis. Background Technology
[0002] Non-destructive testing (NDT) technology is an important means of ensuring the safety of industrial products and in-service equipment. Eddy current pulse thermal imaging (ECPT) combines the principles of electromagnetic induction and infrared thermal imaging, enabling non-contact, large-area, and rapid detection of defects at different depths in conductive materials. It has been widely used in aerospace, energy, and rail transportation fields.
[0003] The working principle of ECPT is that an excitation coil generates short-duration high-frequency AC pulses, inducing eddy currents on the surface of a conductive sample. The Joule heating effect of the eddy currents causes a transient increase in the surface temperature of the sample, and the infrared thermal imager records the subsequent temperature decay process. When defects are present in the sample, the defects act as thermal barriers, hindering the vertical diffusion of heat. This results in the defect area exhibiting a higher temperature peak and a slower decay characteristic, thus forming a recognizable thermal anomaly in the thermal image.
[0004] However, the quality of raw ECPT thermal data is severely affected by two types of physical phenomena: lateral thermal diffusion, where heat diffuses laterally from the defect location, blurring defect boundaries and reducing image contrast; and non-uniform heating, where the inherent non-uniform induction characteristics of the excitation coil superimpose a strong thermal gradient background onto the image, masking subtle defect thermal features. These two problems are particularly prominent when detecting small defects such as flat-bottomed holes or fine natural cracks with diameters of 1-3 mm.
[0005] To overcome these limitations, researchers have developed various signal post-processing algorithms. Pulse Phase Thermal Imaging (PPT) transforms thermal attenuation to the frequency domain, obtaining a phase map insensitive to non-uniform heating. Thermal Imaging Signal Reconstruction (TSR) extracts time derivative features by fitting the temperature attenuation curve with a logarithmic polynomial. Principal Component Analysis (PCT) performs principal component analysis (PCA) on thermal image sequences, utilizing data dimensionality reduction to extract features; this is a widely used method. Independent Component Analysis (ICA), Robust Principal Component Analysis (RPCA), and Sparse Principal Component Analysis (SPCA) have also been studied in ECPT post-processing.
[0006] However, statistical methods, such as PCT, are essentially blind statistical tools. They perform statistical analysis on all pixels indiscriminately and lack prior guidance on defect location. Therefore, they are unable to separate weak defect thermal features from the dominant background noise, resulting in a low signal-to-noise ratio (SNR) for weak defects.
[0007] Currently, there is a lack of ECPT post-processing methods that can effectively integrate physical prior information into the principal component analysis framework while taking into account computational efficiency, in order to meet the needs of high-speed industrial detection applications. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of existing principal component thermal imaging methods in detecting weak defects. It proposes an eddy current pulse thermal imaging defect detection method based on weighted principal component analysis, which transforms blind statistical analysis into a target-oriented analysis guided by physical priors, and significantly improves the signal-to-noise ratio of weak defect detection.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a defect detection method based on weighted principal component analysis using eddy current pulse thermal imaging, comprising the following steps: The thermal saliency weight map generation step involves calculating the time integral thermal energy of each pixel from the original three-dimensional thermal sequence data acquired by the eddy current pulse thermal imaging system to generate an initial spatial weight map. The nonlinear weight enhancement step involves normalizing the initial spatial weight map and then applying a power-law gamma transform to generate an enhanced spatial weight map, thereby amplifying the weight of the defect region and suppressing the weight of the background region. The spatially weighted principal component thermal imaging step integrates the enhanced spatial weight map into the principal component thermal imaging algorithm, constructs a weighted covariance matrix, performs eigenvalue decomposition, projects the original data onto the weighted principal components, and generates an enhanced SW-PCT defect feature image.
[0010] Preferably, the step of generating the thermal significance weight map specifically involves: assuming the original thermal sequence data is... Where H×W is the image spatial resolution and N is the number of frames; for each pixel position (x,y), its time integral thermal energy is calculated as the initial weight.
[0011] in The total number of frames; this operation generates a two-dimensional saliency map. The defective area corresponds to a larger weight value, while the background area corresponds to a smaller weight value.
[0012] Preferably, the nonlinear weight enhancement step specifically comprises: firstly, processing the initial weight graph... Perform min-max normalization to obtain the normalized weight graph. Map its range to the interval [0,1]; then apply a power-law gamma transform.
[0013] The exponent γ>1 is the nonlinear enhancement coefficient; the larger the value of γ, the stronger the contrast between the high-weight defect area and the low-weight background area, and the precise positioning of the defect area is achieved by sparse weight distribution.
[0014] Preferably, the spatially weighted principal component thermal imaging step specifically comprises: The original thermal sequence Remodeled into a two-dimensional data matrix ,in This represents the total number of pixels. Enhance weight graph Reshape into a weight vector and normalization ; Calculate the weighted mean curve ; Centering the data matrix ; Calculate the weighted covariance matrix ; right Perform eigenvalue decomposition and extract the eigenvectors corresponding to the K largest eigenvalues. ; Project the centered data onto the feature vector: The score matrix S is then reshaped into a three-dimensional feature image. .
[0015] Preferably, the selection principle of the nonlinear enhancement coefficient γ is as follows: for samples with weak initial thermal contrast and weak defect signals, a larger γ value of 3.0-4.0 is used to amplify the weak thermal features through aggressive contrast enhancement; for samples with strong initial thermal contrast and clear defect signals, a smaller γ value of 1.5-2.5 is used to avoid over-enhancement leading to amplification of background texture noise; for samples with moderate thermal contrast, γ is taken as approximately 3.0 as a robust starting point.
[0016] Preferably, the method is applicable to the following defect types: flat-bottomed hole defects on the surface and subsurface of steel plates, groove defects on the surface of steel plates, natural crack defects on the surface of steel plates, and V-shaped welding defects on the subsurface of nickel alloys; the method covers the defect size range of flat-bottomed holes and small natural cracks with diameters of 1-3 mm.
[0017] Preferably, the eddy current pulse thermal imaging system includes a high-frequency pulse excitation power supply, an excitation coil, an infrared thermal imager, and a data acquisition and processing unit; the excitation pulse duration is 0.2-0.8 seconds, the excitation frequency range is 150-300kHz, and the infrared thermal imager sampling frequency is 50-80Hz.
[0018] Preferably, the weighted covariance matrix The method is constructed using a vectorized implementation, with a computational complexity of approximately O(N²M). Compared to the standard PCT, it only adds linear overhead of the order of O(MN), ensuring high computational efficiency and making it suitable for high-speed industrial detection applications.
[0019] The present invention also provides an eddy current defect detection device based on thermal saliency weighted principal component thermal imaging, comprising: The data acquisition module is used to acquire raw thermal sequence images of defective samples through an eddy current pulse thermal imaging system; The thermal saliency calculation module is used to perform time-dimensional integration on the original thermal sequence to generate a two-dimensional spatial weight map that reflects the thermal accumulation of each pixel. The nonlinear enhancement module is used to perform normalization and power-law gamma transformation on the spatial weight map to generate a high-contrast enhanced weight map. The weighted PCT analysis module is used to integrate the enhanced weight map into the PCT algorithm, construct the weighted covariance matrix and perform eigenvalue decomposition, and output the SW-PCT defect feature image; The results output module is used to display and output the final defect detection feature image, and supports quantitative comparison and evaluation of the signal-to-noise ratio with other methods.
[0020] Preferably, the weighted PCT analysis module further includes a regularization submodule, which, when the weight distribution is extremely uneven, performs regularization on the weighted covariance matrix. A small number of regularization terms are added to the diagonal to mitigate potential singularity issues and ensure computational stability.
[0021] The beneficial effects of this invention are as follows: (1) This invention utilizes the physical law that the time integral thermal energy of the defect region is significantly higher than that of the background region to construct a thermal significance spatial weight map, integrates physical prior knowledge into the principal component analysis framework, and makes the statistical analysis focus on the defect-related region, overcoming the limitations of traditional PCT blind statistics and significantly improving the detection signal-to-noise ratio of weak defects.
[0022] (2) The present invention introduces power-law gamma transformation to nonlinearly enhance the initial weight map, which effectively widens the weight difference between the defect region and the background region, forming a sparse and highly discriminative weight distribution, and further enhances the ability to guide the thermal characteristics of defects.
[0023] (3) The present invention has been systematically verified on six different metal samples, including flat bottom holes, grooves, natural cracks and V-shaped welding defects. Compared with traditional representative methods such as PCT, TSR, ICA, RPCA and SPCA, it has a significant advantage in the signal-to-noise ratio index of weak defect enhancement. The SNR improvement of some samples can reach 100%.
[0024] (4) The present invention adopts a vectorized implementation method, and the computational complexity is comparable to that of the standard PCT. It only adds linear overhead of O(MN) level, does not require iterative optimization, has high computational efficiency, and meets the actual needs of industrial high-speed detection applications. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the overall workflow of the SW-PCT method of the present invention; Figure 2 This is a schematic diagram of the nonlinear enhancement process of the spatial weight map in this invention: the left side is the initial spatial weight map; the middle side is a schematic diagram of the gamma transform; and the right side is the enhanced spatial weight map. Figure 3 The graph shows the comparative experimental results of each method on six samples. Figure 4 This is a comparison of ablation experiments between the present invention and traditional PCT and linear SW-PCT; Figure 5 The graph shows the SNR variation of each sample under different γ values. Detailed Implementation
[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are preferred embodiments of the present invention, and those skilled in the art can make equivalent substitutions without departing from the principle of the present invention.
[0027] The core physical basis of this invention is that when defects exist in a conductive material, these defects act as thermal barriers, hindering the vertical diffusion path of heat within the material. This results in a longer heat accumulation time and a larger total heat accumulation on the surface of the defective region. Therefore, by integrating the thermal sequence data over time, the degree of heat accumulation at each pixel can be quantitatively characterized, thereby reliably identifying potential defective regions.
[0028] Based on the above physical mechanism, this invention defines the thermal significance metric as the time integral of the pixel temperature curve, that is, the cumulative sum of the temperature values of each pixel, which is used to generate an initial spatial weight map. Then, the discriminativeness of the weight distribution is enhanced by nonlinear transformation, and finally this spatial guidance information is embedded in the PCT statistical analysis framework.
[0029] like Figures 1 to 5 As shown, a defect detection method based on weighted principal component analysis using eddy current pulse thermal imaging includes the following steps: Step 1: Generating the heat significance weight map
[0030] Let the raw thermal sequence data acquired by the eddy current pulse thermal imaging system be... , where H and W are the height and width of the thermal image, respectively, and N is the number of acquisition frames.
[0031] For each pixel position Calculate its time integral thermal energy as the initial weight:
[0032] in This represents the total number of frames. This operation sums the hot sequences over time to generate a two-dimensional significance map. In this image, pixels corresponding to defects such as flat-bottomed holes and cracks have higher weight values due to the heat accumulation effect, while the weight values of healthy background areas are relatively lower. Step 2: Nonlinear Weight Enhancement
[0033] The initial weight graph generated in step one The discrimination capability is limited, with the weights of a large number of pixels concentrated in a low value range, resulting in insufficient weight contrast between defective areas and background areas. Therefore, a nonlinear enhancement step is introduced.
[0034] First of all, Perform min-max normalization to map the weight values to the [0,1] interval, thus obtaining the normalized weights. :
[0035] Then, a power-law gamma transform is applied to the normalized weighted graph:
[0036] The exponent γ>1 represents the nonlinear enhancement coefficient. When γ>1, high-weight values close to 1 (corresponding to defect areas) are moderately amplified, while low-weight values close to 0 (corresponding to background areas) are significantly compressed, resulting in a sparse, high-contrast distribution in the weight map, with a few pixels holding significantly larger weights. This provides strong spatial guidance for subsequent PCT analysis. Step 3: Spatial Weighted Principal Component Thermal Imaging
[0037] The enhanced spatial weight map is then incorporated into the PCT algorithm, as follows: Data reshaping: transforming the original hot sequence Remodeled into a two-dimensional data matrix ,in For each row of the matrix, the total number of pixels is [number]. The time-temperature curve corresponding to a single pixel.
[0038] Weight vector construction: enhancing the weight graph Reshape into a weight vector and normalization makes This ensures numerical stability.
[0039] Weighted mean calculation: Calculate the weighted mean curve ,in Let X be the i-th row.
[0040] Data centerization: , where 1 is a column vector of all 1s.
[0041] Construction of the weighted covariance matrix: Calculation of the weighted covariance matrix: This matrix primarily reflects the thermal dynamic characteristics of important regions (i.e., potential defect regions) identified by thermal significance weights.
[0042] Eigenvalue decomposition: for Perform eigenvalue decomposition to obtain the eigenvector matrix V, and select the top K eigenvectors corresponding to the K largest eigenvalues. .
[0043] Feature image generation: Projecting centered data onto selected feature vectors: and the score matrix Reconstructed into a 3D feature image This is the final SW-PCT defect feature image.
[0044] Due to design biases, SW-PCT feature images will preferentially present the thermal dynamic patterns of high-weight regions, thereby achieving a significant enhancement of the thermal features of defects.
[0045] The computational efficiency analysis is as follows: The computational complexity of this invention is mainly determined by the steps involved in calculating the weighted covariance matrix, i.e. When implemented using vectorization, the complexity is approximately O(N²M), adding only a linear overhead of O(MN) compared to the standard PCT. The entire method is based on the closed-form solution of Singular Value Decomposition (SVD), requiring no iterative optimization, thus possessing inherent numerical stability and deterministic results, and exhibiting no convergence issues.
[0046] The following are guidelines for selecting the γ parameter: The nonlinear enhancement factor γ has a crucial impact on SW-PCT performance. Based on the intensity of the thermal characteristics of the sample defects, the following empirical selection criteria are provided: For samples with weak defect signals: It is recommended to use a larger γ value (3.0-4.0) to amplify weak thermal characteristics through aggressive contrast enhancement.
[0047] For samples with strong defect signals: It is recommended to use a smaller γ value (1.5-2.5) to avoid over-enhancing, which would cause the background texture or noise to be amplified into artifacts.
[0048] For medium signal samples: γ≈3.0 can be used as a robust starting reference value.
[0049] The γ parameter can be optimized using an automated selection algorithm based on thermal feature analysis to further improve the practicality of industrial automated inspection systems.
[0050] This invention was systematically verified experimentally on six different metal samples, covering a variety of defect types: Sample 1: Flat-bottomed hole defect on smooth steel plate surface (3mm×7mm, diameter×depth); Sample 2: Flat-bottomed hole defect on the surface of painted steel plate (2mm×6mm); Sample 3: Flat-bottomed hole defect on smooth steel plate surface (1mm×5mm); Sample 4: Groove defects on the surface of a rusted steel plate (100mm × 1mm, length × width). Sample 5: Natural cracks on the surface of the steel plate (three small cracks); Sample 6: V-shaped welding defect on the subsurface of nickel alloy (depth 2mm).
[0051] The experimental system consists of a high-frequency pulse excitation power supply, an excitation coil, an infrared thermal imager (Optris-PI450 or Hikvision-MV-CI003-GL-T6), and a data acquisition and processing unit. The excitation frequency range is 100-300kHz, and the sampling frequency is 50-80Hz.
[0052] Using signal-to-noise ratio (SNR) as a quantitative evaluation index, the SW-PCT method (γ=3) of this invention was compared with five representative methods: traditional PCT, TSR, ICA, RPCA, and SPCA. Experimental results show that the SW-PCT method exhibits a significant improvement in SNR on all six samples, with the most prominent enhancement effect on weak signal defects. Ablation experiments further verified the independent contributions of the spatial weighting and nonlinear enhancement components: the spatial weighting component contributed approximately 60-75% of the SNR improvement, while the nonlinear enhancement component contributed approximately 25-40% of the SNR improvement; the two complement each other to achieve optimal performance.
[0053] Taking the detection of three flat-bottomed hole defects on the surface of a smooth steel plate (sample 1) as an example, the specific implementation process is as follows: A high-frequency excitation power supply with an excitation pulse duration of 0.5s and a frequency of 150-300kHz was used to drive the excitation coil to excite the sample. The infrared thermal imager acquired the thermal sequence at a sampling rate of 80Hz, with an image resolution of 382×288 pixels and 749 frames acquired. Calculate the time integral for each pixel in the obtained hot sequence I∈R^(382×288×749) to generate the initial weight map. ∈R^(382×288), the regions corresponding to the three flat-bottomed holes show obvious high weight values, especially the third defect located at the edge of the excitation coil, whose thermal characteristics are relatively weak; right After performing min-max normalization, a power-law transformation with γ=3 is applied to generate an enhanced weighted graph. The high-weight defect area is further highlighted, while the background area is effectively suppressed; Will Incorporating PCT Analysis: Constructing a Weighted Covariance Matrix ∈R^(749×749), perform eigenvalue decomposition, take the first 3 principal components, and generate SW-PCT feature image; The SW-PCT feature image clearly shows three flat-bottom hole defects, including a third weak defect located at the edge of the excitation coil. Its outline is clear, and the average SNR of the three defects is significantly improved compared with the traditional PCT method.
[0054] This invention encompasses any alternatives, modifications, or equivalent methods and solutions within the spirit and scope of this invention. The above descriptions are merely preferred embodiments of this invention. For those skilled in the art, various improvements and modifications can be made without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.
Claims
1. A defect detection method based on weighted principal component analysis using eddy current pulse thermal imaging, characterized in that, Includes the following steps: The steps for generating the thermal saliency weight map are as follows: For the original three-dimensional thermal sequence data acquired by the eddy current pulse thermal imaging system, the time integral thermal energy of each pixel is calculated to generate an initial spatial weight map. The nonlinear weight enhancement step involves normalizing the initial spatial weight map and then applying a power-law gamma transform to generate an enhanced spatial weight map, thereby amplifying the weight of the defect region and suppressing the weight of the background region. The spatially weighted principal component thermal imaging (SW-PCT) step integrates the enhanced spatial weight map into the PCT algorithm, constructs a weighted covariance matrix, performs eigenvalue decomposition, projects the original data onto the weighted principal components, and generates an enhanced SW-PCT defect feature image.
2. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 1, characterized in that, The specific steps for generating the thermal significance weight map are as follows: Let the original thermal sequence data be... Where H×W is the image spatial resolution and N is the number of frames; for each pixel position (x,y), its time integral thermal energy is calculated as the initial weight. , in The total number of frames; this operation generates a two-dimensional saliency map. The defective area corresponds to a larger weight value, while the background area corresponds to a smaller weight value.
3. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 1, characterized in that, The nonlinear weight enhancement step specifically involves: first, processing the initial weight graph... Perform min-max normalization to obtain the normalized weight graph. Map its range to the interval [0,1]; then apply a power-law gamma transform. The exponent γ>1 is the nonlinear enhancement coefficient; the larger the value of γ, the stronger the contrast between the high-weight defect area and the low-weight background area, and the precise positioning of the defect area is achieved by sparse weight distribution.
4. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 1, characterized in that, The spatially weighted principal component thermal imaging step is specifically as follows: The original thermal sequence Remodeled into a two-dimensional data matrix ,in This represents the total number of pixels. Enhance weight graph Reshape into a weight vector and normalization ; Calculate the weighted mean curve ; Centering the data matrix ; Calculate the weighted covariance matrix ; right Perform eigenvalue decomposition and extract the eigenvectors corresponding to the K largest eigenvalues. ; Project the centered data onto the feature vector: The score matrix S is then reshaped into a three-dimensional feature image. .
5. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 3, characterized in that, The selection principle of the nonlinear enhancement coefficient γ is as follows: for samples with weak initial thermal contrast and weak defect signals, a larger γ value of 3.0-4.0 is used to amplify the weak thermal features through aggressive contrast enhancement; for samples with strong initial thermal contrast and clear defect signals, a smaller γ value of 1.5-2.5 is used to avoid over-enhancement leading to amplification of background texture noise; for samples with moderate thermal contrast, γ is taken as approximately 3.0 as the robust starting point.
6. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 1, characterized in that, The method is applicable to the following defect types: flat-bottomed hole defects on the surface and subsurface of steel plates, groove defects on the surface of steel plates, natural crack defects on the surface of steel plates, and V-shaped welding defects on the subsurface of nickel alloys; the method covers the defect size range of flat-bottomed holes and small natural cracks with diameters of 1-3 mm.
7. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 1, characterized in that, The eddy current pulse thermal imaging system includes a high-frequency pulse excitation power supply, an excitation coil, an infrared thermal imager, and a data acquisition and processing unit; the excitation pulse duration is 0.2-0.8 seconds, the excitation frequency range is 150-300kHz, and the infrared thermal imager sampling frequency is 50-80Hz.
8. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 4, characterized in that, The weighted covariance matrix The method is constructed using a vectorized implementation, with a computational complexity of approximately O(N²M). Compared to the standard PCT, it only adds linear overhead of the order of O(MN), ensuring high computational efficiency and making it suitable for high-speed industrial detection applications.
9. A device for defect detection using eddy current pulse thermal imaging based on weighted principal component analysis, characterized in that, include: The data acquisition module is used to acquire raw thermal sequence images of defective samples through an eddy current pulse thermal imaging system; The thermal saliency calculation module is used to perform time-dimensional integration on the original thermal sequence to generate a two-dimensional spatial weight map that reflects the thermal accumulation of each pixel. The nonlinear enhancement module is used to perform normalization and power-law gamma transformation on the spatial weight map to generate a high-contrast enhanced weight map. The weighted PCT analysis module is used to integrate the enhanced weight map into the PCT algorithm, construct the weighted covariance matrix and perform eigenvalue decomposition, and output the SW-PCT defect feature image; The results output module is used to display and output the final defect detection feature image, and supports quantitative comparison and evaluation of the signal-to-noise ratio with other methods.
10. The eddy current pulse thermal imaging defect detection method based on weighted principal component analysis according to claim 9, characterized in that, The weighted PCT analysis module also includes a regularization submodule, which, when the weight distribution is extremely uneven, performs regularization on the weighted covariance matrix. A small number of regularization terms are added to the diagonal to mitigate potential singularity issues and ensure computational stability.