Nondestructive inspection method for enhancing defect detection efficiency through deep learning

Abstract

The invention relates to the technical field of non-destructive testing, and discloses a non-destructive testing method for enhancing defect detection efficiency through deep learning, which comprises the following steps: step 1, acquiring multi-modal data of a target material, including data acquired through X-ray imaging, ultrasonic detection and infrared thermal imaging; 2, signal denoising processing is carried out on the obtained multi-modal data, signal denoising is carried out through wavelet transformation, and the wavelet transformation comprises the steps that multi-scale decomposition is carried out on signals, and high-frequency noise is removed through a soft threshold method; and step 3, carrying out image enhancement processing on the de-noised multi-modal data. The internal structural features of the material are captured from multiple physical dimensions by fusing multi-modal data of X-ray imaging, ultrasonic detection and infrared thermal imaging. The multi-modal fusion can make up for the deficiency of a single detection technology, so that the defect identification in complex structures and materials is more comprehensive and accurate, thereby significantly improving the detection precision.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing technology, and specifically to a nondestructive testing method for enhancing defect detection efficiency through deep learning. Background Art

[0002] With the rapid development of modern industry, the complexity of materials and structures is increasing, and the requirements for their quality and safety are becoming higher and higher. As a method that can detect internal defects without destroying the material, non-destructive testing has been widely used in aerospace, automobile manufacturing, energy and other fields. The development of non-destructive testing technology has greatly promoted the improvement of industrial product quality and the enhancement of safety assurance, and has become an indispensable and important means in modern manufacturing.

[0003] Existing nondestructive testing technologies mainly rely on single-mode detection methods, such as X-ray imaging, ultrasonic detection or infrared thermal imaging. These traditional methods detect the internal structure of materials through simple signal processing and image enhancement technology, and rely on experienced operators to identify defects. However, due to the limitations of each of these technologies, single-mode detection is often difficult to fully and accurately identify tiny defects in complex structures. In addition, in the absence of automated processing, the detection results of traditional methods are easily affected by human factors, and there is a high rate of misjudgment and missed judgment.

[0004] A major problem with existing technologies is that single-mode nondestructive testing methods are difficult to fully and accurately identify tiny defects in the detection of complex materials and structures. This is because different detection modalities have their own limitations, and it is difficult to capture multiple physical properties of materials at the same time during the same detection process, resulting in insufficient detection accuracy and reliability. Therefore, how to improve detection accuracy through multi-modal data fusion and automation technology has become an urgent problem to be solved in the current non-destructive testing field. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a nondestructive flaw detection method that enhances defect detection efficiency through deep learning, solving the problem that the single-modality nondestructive flaw detection method in the prior art is difficult to comprehensively and accurately identify tiny defects in the detection of complex materials and structures.

[0006] To achieve the above objectives, the present invention is implemented by the following technical solution: a non-destructive testing method for enhancing defect detection efficiency through deep learning, comprising the following steps:

[0007] Step 1: Obtain multimodal data of the target material, including data obtained through X-ray imaging, ultrasonic detection and infrared thermal imaging;

[0008] Step 2: performing signal denoising on the acquired multimodal data, wherein the signal denoising is performed by wavelet transform, wherein the wavelet transform includes multi-scale decomposition of the signal and application of a soft threshold method to remove high-frequency noise;

[0009] Step 3: Perform image enhancement processing on the denoised multimodal data, including applying histogram equalization for global contrast enhancement and applying Laplace operator for local contrast enhancement;

[0010] Step 4: Input the enhanced multimodal data into the convolutional neural network for feature fusion, wherein the feature fusion includes using the convolutional layer for deep feature extraction, the pooling layer for size reduction and the fully connected layer for feature integration;

[0011] Step 5: Based on the fused feature map, the defects of the target material are identified and classified through the CNN model.

[0012] Preferably, the X-ray imaging data acquisition step includes:

[0013] Select the X-ray energy range of 80-150kV, and select the appropriate energy value according to the density and thickness of the target material to ensure the best imaging effect;

[0014] Set the exposure time to 10-100ms;

[0015] Acquire multiple X-ray images through multi-angle scanning to capture the three-dimensional structural characteristics of the material;

[0016] The resolution of the X-ray image is not less than 1024x1024 pixels, and the image format is DI COM.

[0017] Preferably, the data collection step of the ultrasonic detection includes:

[0018] Select the frequency range of the ultrasonic probe to be 1-10MHz or 10-50MHz. The specific frequency is selected according to the thickness and acoustic impedance matching characteristics of the target material.

[0019] Apply coupling agent evenly on the surface of the material;

[0020] The scanning motor controls the probe to move smoothly on the material surface with a scanning interval of 0.1mm;

[0021] The scanning speed and angle of the ultrasonic probe are monitored and adjusted in real time to obtain the best signal-to-noise ratio.

[0022] Preferably, the infrared thermal imaging data collection step includes:

[0023] Use an infrared thermal imager with a resolution of 640x480 pixels or higher;

[0024] Apply a short-term heat pulse stimulus to the surface of the material through laser or infrared light. The intensity and duration of the heat pulse are adjusted according to the thermal conductivity of the material.

[0025] Record the temperature change curve after the heat pulse is applied. The thermal image is collected at a frame rate of 30 Hz or higher and saved in video format to capture the dynamic thermal response.

[0026] Analyze areas of temperature anomalies in infrared thermal images to identify possible surface or near-surface defects.

[0027] Preferably, the signal denoising processing step includes:

[0028] The Daubechies wavelet basis is selected as the basis function of wavelet transform, and three-level wavelet decomposition is performed to extract the low-frequency and high-frequency components in the signal;

[0029] The soft threshold method is used to denoise the high-frequency components. To remove noise, where σ is the noise standard deviation and n is the signal length;

[0030] The denoised low-frequency and high-frequency components are inversely transformed to reconstruct the denoised signal.

[0031] Preferably, the image enhancement processing step includes:

[0032] Apply histogram equalization to X-ray images and infrared thermal images to even out the grayscale distribution of the image and enhance the overall contrast;

[0033] Calculate the cumulative distribution function of the image and extend it to the full grayscale range by linear interpolation;

[0034] The Laplacian operator is further applied to perform local contrast enhancement. The edge information in the image is detected by the Laplacian filter, and the enhancement coefficient λ is adjusted.

[0035] Preferably, the feature extraction step of the convolutional neural network includes:

[0036] Use multiple convolutional layers to extract features of the input multimodal image layer by layer, the convolution kernel size of the convolutional layer is 3x3 or 5x5, and the activation function is the ReLU function;

[0037] A pooling layer is applied after each convolutional layer, and the pooling layer uses a 2x2 maximum pooling operation;

[0038] The feature maps after multi-layer convolution and pooling are integrated into a feature vector through a fully connected layer.

[0039] Preferably, the input features in the feature fusion step include:

[0040] The edge features extracted from the X-ray image are extracted by Sobel operator to reflect the density change of the internal structure of the material;

[0041] The frequency characteristics extracted from the ultrasonic signal are extracted through Fourier transform and reflect the acoustic characteristics inside the material;

[0042] The temperature abnormality area extracted from the infrared thermal image is identified by the threshold-based heat zone detection algorithm, reflecting the abnormal thermal distribution on the material surface.

[0043] Preferably, the convolutional neural network model used in the defect identification step includes:

[0044] Add a batch normalization layer on top of the convolutional and pooling layers;

[0045] Using a fully connected layer with a cross entropy loss function to classify multiple types of defects, the classification results include the type, location and confidence of the defect;

[0046] During model training, data augmentation techniques are applied, including random cropping, rotation, and scaling.

[0047] Preferably, the defect identification result is used to generate a defect assessment report containing the following contents:

[0048] Classification of defect types, including surface cracks, internal pores, and interlayer separation;

[0049] The location and size of the defect, including the three-dimensional coordinates, depth and size of the defect in the material;

[0050] Defect severity assessment, combined with defect density and distribution to generate statistical reports;

[0051] Recommendations for repairing different types of defects, including which repair techniques to use and the repair steps required.

[0052] The present invention provides a non-destructive testing method that enhances defect detection efficiency through deep learning. It has the following beneficial effects:

[0053] 1. The present invention captures the internal structural characteristics of materials from multiple physical dimensions by fusing multimodal data from X-ray imaging, ultrasonic detection and infrared thermal imaging. This multimodal fusion can make up for the shortcomings of a single detection technology, making the identification of defects in complex structures and materials more comprehensive and accurate, thereby significantly improving the detection accuracy.

[0054] 2. The present invention uses convolutional neural network (CNN) to identify defects on the feature map after multi-modal fusion, relying on the powerful classification ability of the deep learning model, and can realize automatic identification and classification of various defect types. This not only improves the accuracy of identification, but also reduces the impact of human operation on the detection results, and significantly improves the detection efficiency.

[0055] 3. The present invention uses advanced wavelet transform technology (such as Daubechies wavelet basis) to denoise the collected signal. The present invention can effectively remove noise while retaining key structural information in the signal. This provides high-quality data input for subsequent image enhancement and defect identification, ensuring the reliability and stability of the overall detection process.

[0056] 4. The present invention introduces techniques such as histogram equalization and Laplace operator enhancement in the image processing step, which significantly improves the global and local contrast of the image, making the key features in the material more clearly visible. Especially in the detection of complex materials and structures, these enhancement techniques can effectively highlight defect features and facilitate subsequent automated analysis and identification.

[0057] 5. The defect assessment report generated by the present invention not only includes the type, location and severity of the defect, but also provides targeted repair suggestions. These reports are standardized and structured and can be directly applied to industrial quality control and maintenance decisions, helping enterprises to timely discover and repair potential safety hazards and improve overall product quality and safety.

[0058] 6. The automated testing process of the present invention greatly reduces the reliance on experienced operators and reduces the impact of human factors on the test results. This not only makes the test results more objective and consistent, but also reduces the professional requirements for operators and improves the operability and popularity of the test process. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Embodiment 1:

[0062] Please refer to the attached Figure 1, an embodiment of the present invention provides a nondestructive flaw detection method for enhancing defect detection efficiency through deep learning, comprising the following steps:

[0063] 1. Multimodal data collection

[0064] Step 1.1: X-ray imaging data acquisition

[0065] In this embodiment, industrial-grade X-ray equipment is used for data acquisition, and appropriate parameters are selected to adapt to the characteristics of the material. Specifically, the energy range of the X-ray should be set between 80-150kV, which can be done through the software that comes with the equipment or manual adjustment. The exposure time is set between 10-100ms to avoid overexposure or underexposure of the image. Through multi-angle scanning technology, multiple X-ray images are obtained from different angles to capture the three-dimensional structural information inside the material. The resolution of these images is not less than 1024x1024 pixels and is saved in DI COM format to ensure the readability of the image and the compatibility of subsequent processing.

[0066] Step 1.2: Ultrasonic detection data acquisition

[0067] In this embodiment, an ultrasonic probe with a frequency range of 1-10MHz is selected for the detection of thick materials, while a probe with a frequency range of 10-50MHz is selected for the detection of thin materials. Before scanning, a coupling agent (such as glycerin or water) should be evenly applied to the surface of the material to eliminate the loss of sound wave energy caused by the air interface and ensure the effective transmission of ultrasonic waves. By using a scanning motor, the probe is controlled to move smoothly on the surface of the material, and the scanning interval should be set to 0.1mm to ensure high-resolution data acquisition. The scanning speed and angle of the probe can be optimized through real-time monitoring and adjustment to obtain the best signal-to-noise ratio and signal stability. The calculation of the detection depth can be performed according to the following formula:

[0068]

[0069] Where c is the propagation speed of ultrasound in the material, t is the reflection time, and c can be determined by the following formula:

[0070]

[0071] Where ρ is the material density and k is the bulk compression modulus of the material.

[0072] Step 1.3: Infrared thermal imaging data acquisition

[0073] In this embodiment, an infrared thermal imager with a resolution of 640x480 pixels or higher is used to ensure that the thermal distribution on the surface of the material can be accurately captured. First, a short-term heat pulse stimulation is applied to the surface of the material using a laser or an infrared lamp. The intensity and duration of the heat pulse should be adjusted according to the thermal conductivity of the material to avoid overheating or insufficient thermal response to the surface of the material. After the heat pulse is applied, the temperature change curve of the material is recorded at a frame rate of 30Hz or higher using a thermal imager. The data should be saved in a video format (such as .avi or .mp4) to capture the complete dynamic thermal response process. By analyzing the temperature abnormality areas in the infrared thermal image, defects on the surface or near the surface of the material can be identified. The heat diffusion process can be described by the following heat conduction equation:

[0074]

[0075] Where T is the temperature, is the thermal diffusion coefficient, k is the thermal conductivity, ρ is the density, c p is the specific heat capacity.

[0076] 2. Signal processing and image enhancement

[0077] Step 2.1: Signal denoising

[0078] In this embodiment, the collected multimodal data is subjected to signal denoising, and the specific operation includes selecting Daubechies wavelet basis function and performing three-level wavelet decomposition. Through wavelet transform, the signal is subjected to multi-scale analysis to extract low-frequency and high-frequency components. For the high-frequency component, the soft threshold method is used for denoising, and the threshold λ is determined according to the following formula:

[0079]

[0080] Among them, σ is the standard deviation of noise, n is the signal length, and the threshold can be set by step-by-step experimentation to select a suitable threshold to retain key structural information while removing high-frequency noise. After that, the denoised signal is reconstructed through inverse wavelet transform to generate clear and stable image or signal data.

[0081] Step 2.2: Image enhancement processing

[0082] In this embodiment, the denoised X-ray image and infrared thermal image are subjected to image enhancement processing, which specifically includes two steps: histogram equalization and Laplace enhancement. First, the grayscale distribution of the image is adjusted to make it more uniform through the histogram equalization technology. The cumulative distribution function of the image is calculated and extended to the full grayscale range through linear interpolation to enhance the global contrast of the image. Next, the Laplace operator is applied to enhance the local contrast of the image, and the key structural information in the image is highlighted through edge detection and enhancement algorithms. The Laplace enhancement formula is as follows:

[0083]

[0084] Where I(x, y) is the original image, I′(x, y) is the enhanced image, and λ is the enhancement coefficient. According to the image quality, adjust the value of λ to achieve the best sharpening effect while avoiding artifacts.

[0085] 3. Multimodal Data Fusion

[0086] Step 3.1: Feature extraction

[0087] In this embodiment, features are extracted from the enhanced multimodal images, specifically including three aspects: extracting edge features from X-ray images, extracting frequency features from ultrasonic signals, and extracting temperature abnormality areas from infrared thermal images. For X-ray images, the Sobel operator is used for edge detection to obtain an edge feature map reflecting the internal structure of the material. For ultrasonic signals, Fourier transform is used to extract frequency features and analyze the acoustic properties inside the material. For infrared thermal images, a threshold-based algorithm is used to identify temperature abnormality areas and mark possible defect locations. The results after feature extraction are saved in the form of feature vectors or feature maps to provide input for subsequent feature fusion and deep learning models.

[0088] Step 3.2: Feature Fusion

[0089] In this embodiment, the extracted multimodal features are input into a convolutional neural network (CNN) for fusion. The architecture design of CNN includes multiple layers of convolutional layers, pooling layers, and fully connected layers. First, the convolutional layer extracts deep features through convolution kernels of different sizes, and each convolutional layer is followed by a pooling layer to reduce the size and amount of calculation of the feature map. The pooling layer uses the maximum pooling operation to further reduce the computational complexity and retain the key information of the feature map. Finally, the fully connected layer integrates the multimodal features into a single feature vector for classification or regression tasks. The activation function in the convolutional layer uses the ReLU function, and the formula is as follows:

[0090] f(x)=max(0,x)

[0091] This activation function can effectively enhance the nonlinear expression ability of the feature map, thereby improving the recognition accuracy and robustness of the model.

[0092] 4. Defect identification and assessment

[0093] Step 4.1: Defect Identification

[0094] In this embodiment, a trained convolutional neural network model is used to identify defects on the fused feature map. The design of the CNN model includes multiple convolutional layers, pooling layers, and fully connected layers, plus a batch normalization layer to reduce the offset of internal covariates and accelerate the training process. During the model training process, the cross entropy loss function is used to ensure the classification accuracy of multiple types of defects. The generalization ability of the model is further improved through data enhancement techniques such as random cropping, rotation, and scaling. The output of the model includes the type, location, and confidence of the defect. With this information, different types of defects in the material can be quickly and accurately identified and classification results can be generated.

[0095] Step 4.2: Defect assessment and report generation

[0096] In this embodiment, based on the defect recognition results output by the CNN model, a defect assessment report containing the following contents is generated: defect type, location, size, depth and severity. The specific operation includes integrating the three-dimensional coordinates and size information of the defect into the report, and generating the density and distribution map of the defect through statistical analysis. The report should also include repair suggestions for different types of defects, providing detailed repair techniques and steps. The generation of the assessment report is carried out through a standardized template to ensure the structure and readability of the report, which is convenient for engineers to refer to and make decisions in practical applications.

[0097] Embodiment 2:

[0098] The difference between the second embodiment and the first embodiment is that the frequency of the ultrasonic probe selected in the ultrasonic detection data collection step is adjusted. Specifically, the frequency range of the ultrasonic probe in the second embodiment is set to a fixed 10 MHz, while in the first embodiment, the frequency of the ultrasonic probe is dynamically selected according to the thickness and acoustic impedance characteristics of the material, and specifically includes the following steps:

[0099] 1. Multimodal data collection

[0100] Step 1.1: X-ray imaging data acquisition

[0101] In this embodiment, the same X-ray equipment and parameter settings as those in the first embodiment are used for data acquisition, including an X-ray energy range of 80-150 kV and an exposure time of 10-100 ms. Multiple X-ray images are acquired by multi-angle scanning and saved in DI COM format. This part of the operation is completely consistent with the first embodiment, ensuring the consistency and comparability of the X-ray imaging data.

[0102] Step 1.2: Ultrasonic detection data acquisition

[0103] In this embodiment, an ultrasonic probe with a fixed frequency of 10 MHz is selected for data acquisition. Different from the first embodiment, the frequency of the probe is not dynamically adjusted according to the thickness and acoustic impedance characteristics of the material, but a 10 MHz probe is uniformly used to detect all materials. During the scanning process, a scanning interval of 0.1 mm is still used, and a coupling agent is evenly applied on the surface of the material to ensure the effective transmission of the ultrasonic wave. The scanning speed and angle of the probe are adjusted through real-time monitoring to obtain the best possible signal-to-noise ratio. Due to the fixed frequency of the probe, the detection effect of some materials may not be as good as that of the first embodiment, which will be explained in detail in the subsequent result analysis.

[0104] Step 1.3: Infrared thermal imaging data acquisition

[0105] In this embodiment, the infrared thermal imaging data acquisition steps are consistent with those in the first embodiment. An infrared thermal imager with a resolution of 640x480 pixels is used to apply a short-term heat pulse stimulus and record the thermal diffusion process of the material at a frame rate of 30Hz. The data is saved in a video format, and surface or near-surface defects are identified by analyzing the temperature abnormality area in the thermal image. The thermal diffusion process is still described by the heat conduction equation.

[0106] 2. Signal processing and image enhancement

[0107] Step 2.1: Signal denoising

[0108] In this embodiment, the signal denoising process is exactly the same as that in the first embodiment. The multimodal data is subjected to three-level wavelet decomposition using Daubechies wavelet basis functions, and high-frequency noise is removed by a soft threshold method. The denoised data is reconstructed by an inverse wavelet transform to generate a clear and stable signal or image.

[0109] Step 2.2: Image enhancement

[0110] In this embodiment, the image enhancement processing steps are consistent with those in Embodiment 1. The image processing techniques including histogram equalization and Laplace enhancement are intended to improve the global and local contrast of the image. The edge details of the image are enhanced by calculating the cumulative distribution function of the image and applying the Laplace operator.

[0111] 3. Multimodal Data Fusion

[0112] Step 3.1: Feature extraction

[0113] In this embodiment, the feature extraction step is consistent with that in the first embodiment. Edge features are extracted from the X-ray image, frequency features are extracted from the ultrasonic signal, and temperature abnormality areas are extracted from the infrared thermal image. These features are generated by Sobel operator, Fourier transform and threshold detection algorithm, and saved in the form of feature vectors or feature maps.

[0114] Step 3.2: Feature Fusion

[0115] In this embodiment, the feature fusion step is consistent with that in the first embodiment. A convolutional neural network (CNN) is used to fuse multimodal features. The architecture of CNN includes multiple convolutional layers, pooling layers, and fully connected layers. Through these layers, the model performs deep learning on the input feature map and finally integrates the multimodal features into a single vector for subsequent classification or regression tasks.

[0116] 4. Defect identification and assessment

[0117] Step 4.1: Defect Identification

[0118] In this embodiment, the defect identification step is consistent with that in the first embodiment. A convolutional neural network model is used to identify defects on the fused feature map. The design and training process of the CNN model is the same as that in the first embodiment, including the application of data enhancement technology and cross entropy loss function. The model output includes the type, location and confidence of the defect.

[0119] Step 4.2: Defect assessment and report generation

[0120] In this embodiment, the defect assessment and report generation steps are consistent with those in Embodiment 1. A detailed assessment report is generated based on the defect identification results, and the report content includes the defect type, location, size, depth and severity, and provides repair suggestions. The report is generated using a standardized template to ensure structure and readability.

[0121] By comparing the detection effects of Example 1 and Example 2, it can be seen that in Example 1, since the probe frequency is dynamically selected according to the thickness and acoustic impedance characteristics of the material in the ultrasonic detection data acquisition step, the best signal-to-noise ratio and signal stability can be obtained when detecting different types of materials, thereby performing better in the accuracy of defect identification and classification. In Example 2, a fixed-frequency ultrasonic probe is used to detect all materials. Although the operation process is simplified, the detection effect is not as good as that of Example 1 for certain specific materials (such as materials with thicker thickness or higher density), which is reflected in a slight decrease in recognition accuracy.

[0122] Embodiment three:

[0123] The difference between the third embodiment and the first embodiment is that the wavelet basis function used in the signal denoising process is adjusted. Specifically, the Haar wavelet basis is selected for signal denoising in the third embodiment, while the Daubechies wavelet basis is used in the first embodiment, which specifically includes the following steps:

[0124] 1. Multimodal data collection

[0125] Step 1.1: X-ray imaging data acquisition

[0126] In this embodiment, the same X-ray equipment and parameter settings as in the first embodiment are used for data acquisition. Specific parameters include an X-ray energy range of 80-150 kV and an exposure time of 10-100 ms. Multiple high-resolution X-ray images are acquired by multi-angle scanning technology, and the images are saved in DI COM format. These operations ensure the comprehensive capture of the internal structural information of the material, which is consistent with the first embodiment.

[0127] Step 1.2: Ultrasonic detection data acquisition

[0128] In this embodiment, the ultrasonic detection data acquisition steps are consistent with those in the first embodiment. According to the thickness and acoustic impedance characteristics of the material, the frequency of the ultrasonic probe is dynamically selected (1-10MHz for thick materials and 10-50MHz for thin materials). Coupling agent is evenly applied to the surface of the material to reduce the energy loss of the sound wave, and high-resolution data is obtained through a scanning interval of 0.1mm. The scanning speed and angle of the probe are monitored and adjusted in real time to ensure the best signal-to-noise ratio and signal quality. The calculation formula for the detection depth is the same as that in the first embodiment.

[0129] Step 1.3: Infrared thermal imaging data acquisition

[0130] In this embodiment, the infrared thermal imaging data acquisition steps are consistent with those in the first embodiment. An infrared thermal imager with a resolution of 640x480 pixels is used to apply a short-term high-intensity heat pulse stimulation, and the temperature change curve of the material is recorded at a frame rate of 30Hz. The data is saved in a video format, and the defects on the surface or near the surface of the material are identified by analyzing the temperature abnormality area in the infrared thermal image. The heat diffusion process is still described by the heat conduction equation.

[0131] 2. Signal processing and image enhancement

[0132] Step 2.1: Signal denoising

[0133] In this embodiment, the signal denoising process is different from that in the first embodiment, in that the Haar wavelet basis function is used for denoising instead of the Daubechies wavelet basis. The specific operations include:

[0134] Wavelet transform: The collected multimodal data are decomposed into three levels of wavelets using the Haar wavelet basis. The Haar wavelet basis is the simplest wavelet basis with good time resolution but low frequency resolution.

[0135] Denoising: The high-frequency components are denoised using the soft threshold method. The threshold setting still follows the following formula:

[0136]

[0137] Among them, σ is the standard deviation of the noise, and n is the length of the signal. After denoising, the low-frequency and high-frequency components are inversely transformed to reconstruct the denoised signal.

[0138] Step 2.2: Image enhancement

[0139] In this embodiment, the image enhancement processing steps are consistent with those in the first embodiment. It includes histogram equalization and Laplace enhancement technology, which improves the global and local contrast of the image by adjusting the grayscale distribution of the image and enhancing the edge details. The formula of Laplace enhancement is:

[0140]

[0141] Among them, I(x,y) is the original image, I′ is the enhanced image, and λ is the enhancement coefficient.

[0142] 3. Multimodal Data Fusion

[0143] Step 3.1: Feature extraction

[0144] In this embodiment, the feature extraction step is consistent with that in the first embodiment. Edge features are extracted from the X-ray image, frequency features are extracted from the ultrasonic signal, and temperature abnormality areas are extracted from the infrared thermal image. These features are generated by Sobel operator, Fourier transform and threshold detection algorithm, and saved in the form of feature vectors or feature maps to provide input for subsequent feature fusion and deep learning models.

[0145] Step 3.2: Feature Fusion

[0146] In this embodiment, the feature fusion step is consistent with that in the first embodiment. A convolutional neural network (CNN) is used to fuse multimodal features. The architecture of CNN includes multiple convolutional layers, pooling layers, and fully connected layers. Through these layers, the model performs deep learning on the input feature map and finally integrates the multimodal features into a single vector for subsequent classification or regression tasks.

[0147] 4. Defect identification and assessment

[0148] Step 4.1: Defect Identification

[0149] In this embodiment, the defect identification step is consistent with that in the first embodiment. A convolutional neural network model is used to identify defects on the fused feature map. The design and training process of the CNN model is the same as that in the first embodiment, including the application of data enhancement technology and cross entropy loss function. The model output includes the type, location and confidence of the defect.

[0150] Step 4.2: Defect assessment and report generation

[0151] In this embodiment, the defect assessment and report generation steps are consistent with those in Embodiment 1. A detailed assessment report is generated based on the defect identification results, and the report content includes the defect type, location, size, depth and severity, and provides repair suggestions. The report is generated using a standardized template to ensure structure and readability.

[0152] By comparing the detection effects of Example 1 and Example 3, it can be seen that the signal denoising process using the Haar wavelet basis in Example 3 has reduced the computational complexity, but due to the lack of frequency resolution of the Haar wavelet basis, the denoising effect is not as good as that of Example 1 using the Daubechies wavelet basis. Specifically, the reconstructed signal in Example 3 may not perform well when processing fine structures, especially in retaining high-frequency details, and the effect is weakened. This may affect the subsequent image enhancement and feature extraction steps, resulting in a slight decrease in the final defect recognition accuracy.

[0153] The Daubechies wavelet basis used in the first embodiment can better process complex signals and retain more useful information due to its good balance in time and frequency. Therefore, the first embodiment is superior to the third embodiment in terms of denoising effect, image quality and defect recognition accuracy.

[0154] Table 1 Comparison table of experimental data of embodiment

[0155]

[0156] Experimental data analysis

[0157] Signal-to-noise ratio (SNR): In Example 1, the Daubechies wavelet-based denoising process was used to obtain the highest SNR (25.4 dB), which shows that this solution can better retain the useful information in the signal and reduce noise. In Example 2, due to the fixed frequency of the ultrasonic probe, the signal acquisition quality on some materials decreased, resulting in a slightly lower SNR (23.8 dB). In Example 3, due to the lower frequency resolution of the Haar wavelet basis, the SNR (24.2 dB) was slightly inferior to that of Example 1.

[0158] Image Contrast Enhancement (ICE): Example 1 has the best image contrast enhancement effect (92%), which is higher than the other two examples. This shows that the use of high-quality denoising and appropriate thermal pulse stimulation can significantly improve the image contrast. Example 2 uses a fixed ultrasonic probe frequency, resulting in poor image enhancement for some materials (90%). Example 3 has a lower image enhancement quality (88%) because the denoising is not as good as Example 1.

[0159] Defect recognition accuracy (DRA): Example 1 has the highest defect recognition accuracy (96.5%), indicating that the best denoising, image enhancement and multimodal data fusion technology are combined. Example 2 has a lower recognition accuracy of some materials (94.8%) due to the fixed frequency of the ultrasonic probe. Example 3 has a lower defect recognition accuracy (93.2%) due to the poor signal denoising effect.

[0160] Total detection time (TDT): The detection time of Example 2 is the shortest (3.0 hours) because the fixed frequency ultrasonic probe reduces the parameter adjustment time. The detection time of Example 1 is slightly longer (3.2 hours), but since its detection accuracy and effect are optimal, the time increase is reasonable. The detection time of Example 3 is close to that of Example 1 (3.1 hours), but the effect is slightly inferior.

[0161] Comparative Example 1:

[0162] Traditional nondestructive testing methods mainly rely on single-mode X-ray imaging technology and use basic signal processing and image enhancement methods to detect defects. The following are the specific steps of this traditional method:

[0163] 1. X-ray imaging data acquisition

[0164] Step 1.1: X-ray imaging

[0165] In the traditional method, standard industrial-grade X-ray equipment is used to image the internal structure of the material. The specific parameters are usually fixed at 80-120kV X-ray energy and the exposure time is set between 50-200ms. During the operation, the X-ray scan is only performed from a single angle and a two-dimensional image is obtained. The image is usually of low resolution, generally 512x512 pixels, and the data format is JPEG or BMP.

[0166] X-ray imaging reflects internal structure based on differences in material density, but due to the lack of multi-angle or multi-modal supplementation, images may have problems such as overlap and artifacts, making it difficult to accurately identify subtle or complex defects.

[0167] 2. Signal Processing

[0168] Step 2.1: Basic filtering

[0169] In traditional methods, the collected X-ray images are simply filtered. Median filtering or mean filtering is usually used to reduce noise, and the filter kernel size is generally 3x3 or 5x5. These basic filtering methods can only eliminate part of the noise, which can easily lead to image blurring and cannot effectively distinguish noise from key signals.

[0170] Median filtering reduces the impact of random noise by replacing pixel values ​​with the median value in the neighborhood, while mean filtering smoothes the image by calculating the average value of pixels in the neighborhood. Although these methods are simple, they have poor retention of high-frequency details and are prone to loss of important structural information.

[0171] 3. Image Enhancement

[0172] Step 3.1: Histogram Equalization

[0173] In the traditional method, the filtered image is subjected to histogram equalization. The grayscale distribution of the image is stretched to improve the global contrast of the image. This process helps to enhance the visibility of the image, but due to its global nature, it is easy to ignore the improvement of local contrast and has a greater impact on noise.

[0174] Histogram equalization adjusts the pixel value distribution to make the grayscale of the image more uniform, thereby improving the overall brightness and contrast of the image. However, due to the lack of local enhancement mechanism, the processed image may still be difficult to clearly show key defects in complex scenes.

[0175] 4. Defect Detection

[0176] Step 4.1: Manual inspection

[0177] In traditional methods, defect detection mainly relies on experienced operators to manually analyze and judge the enhanced X-ray images. The operator identifies the possible defect type and location by observing abnormal areas in the image (such as brightness changes, break lines, etc.). This method relies on the subjective experience of the operator, is easily affected by human factors, and lacks automation and efficiency.

[0178] Manual inspection is based on the human eye's ability to perceive images and identifies possible defects by comparing grayscale changes in different areas. However, due to image quality limitations and the lack of automated analysis tools, traditional methods have difficulty accurately identifying subtle or hidden defects in complex structures.

[0179] Traditional nondestructive testing methods rely on single-mode X-ray imaging and use relatively simple signal processing and image enhancement methods, which have significant limitations in the accuracy and efficiency of defect detection. In particular, the simplicity of signal processing and the lack of comprehensive analysis of multi-modal data make this method perform poorly when dealing with complex materials and structures, and the reliance on manual inspection is also prone to misjudgment and missed judgment.

[0180] Therefore, traditional methods are difficult to meet the needs of modern industry for high-precision non-destructive testing.

[0181] Embodiment 1 Embodiment 2 Signal-to-Noise Ratio (SNR) 25.4dB 20.1dB Image Contrast Enhancement (ICE) 92％ 75％ Defect Recognition Accuracy (DRA) 96.5％ 80.5％ Total Detection Time (TDT) 3.2 hours 2.8 hours

[0182] Experimental data analysis

[0183] Signal-to-noise ratio (SNR): The signal-to-noise ratio of the conventional comparative example is significantly lower than that of Example 1 (25.4dB), Example 2 (23.8dB) and Example 3 (24.2dB), and is only 20.1dB. This is because the conventional method does not use advanced signal denoising technology, but only relies on simple filtering, resulting in more noise and poor signal quality.

[0184] Image Contrast Enhancement (ICE): The image contrast enhancement effect (75%) of the conventional comparative example is significantly lower than that of the three embodiments of the present invention. This is because the conventional method only relies on basic histogram equalization and fails to effectively enhance the key details and contrast in the image.

[0185] Defect recognition accuracy (DRA): The defect recognition accuracy of the traditional comparison is only 80.5%, which is significantly lower than Example 1 (96.5%), Example 2 (94.8%) and Example 3 (93.2%). The traditional method lacks the support of multimodal data fusion and deep learning technology, resulting in insufficient defect recognition capabilities, especially poor results in identifying complex defects.

[0186] Total Detection Time (TDT): The detection time of the conventional comparative example is 2.8 hours, which is slightly shorter than that of the three embodiments of the present invention. This is because the technology used in the conventional method is relatively simple, and the detection process involves less complex algorithms and data processing. However, this time advantage is at the expense of detection accuracy and recognition ability.

[0187] By comparing the experimental data, it can be seen that the traditional nondestructive testing method lags behind the three embodiments of the present invention in all key indicators, especially in signal-to-noise ratio, image contrast enhancement and defect recognition accuracy. Although the traditional method has certain advantages in detection time, its low precision and low recognition ability make it difficult to meet the detection needs of complex structures and materials.

[0188] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A nondestructive testing method for enhancing defect detection efficiency through deep learning, characterized in that: The following steps are involved: Step 1: Obtain multimodal data of the target material, including data obtained through X-ray imaging, ultrasonic detection and infrared thermal imaging; Step 2: performing signal denoising on the acquired multimodal data, wherein the signal denoising is performed by wavelet transform, wherein the wavelet transform includes multi-scale decomposition of the signal and application of a soft threshold method to remove high-frequency noise; Step 3: Perform image enhancement processing on the denoised multimodal data, including applying histogram equalization for global contrast enhancement and applying Laplace operator for local contrast enhancement; Step 4: Input the enhanced multimodal data into the convolutional neural network for feature fusion, wherein the feature fusion includes using the convolutional layer for deep feature extraction, the pooling layer for size reduction and the fully connected layer for feature integration; Step 5: Based on the fused feature map, the defects of the target material are identified and classified through the CNN model.

2. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The data acquisition step of the X-ray imaging comprises: Select the X-ray energy range of 80-150kV, and select the appropriate energy value according to the density and thickness of the target material to ensure the best imaging effect; Set the exposure time to 10-100ms; Acquire multiple X-ray images through multi-angle scanning to capture the three-dimensional structural characteristics of the material; The resolution of the X-ray image is not less than 1024x1024 pixels, and the image format is DICOM.

3. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The data collection step of the ultrasonic detection comprises: Select the frequency range of the ultrasonic probe to be 1-10MHz or 10-50MHz. The specific frequency is selected according to the thickness and acoustic impedance matching characteristics of the target material. Apply coupling agent evenly on the surface of the material; The scanning motor controls the probe to move smoothly on the material surface with a scanning interval of 0.1mm; The scanning speed and angle of the ultrasonic probe are monitored and adjusted in real time to obtain the best signal-to-noise ratio.

4. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The data acquisition step of the infrared thermal imaging comprises: Use an infrared thermal imager with a resolution of 640x480 pixels or higher; Apply a short-term heat pulse stimulus to the surface of the material through laser or infrared light. The intensity and duration of the heat pulse are adjusted according to the thermal conductivity of the material. Record the temperature change curve after the heat pulse is applied. The thermal image is collected at a frame rate of 30 Hz or higher and saved in video format to capture the dynamic thermal response. Analyze areas of temperature anomalies in infrared thermal images to identify possible surface or near-surface defects.

5. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The signal denoising processing step comprises: The Daubechies wavelet basis is selected as the basis function of wavelet transform, and three-level wavelet decomposition is performed to extract the low-frequency and high-frequency components in the signal; The soft threshold method is used to denoise the high-frequency components. To remove noise, where σ is the noise standard deviation and n is the signal length; The denoised low-frequency and high-frequency components are inversely transformed to reconstruct the denoised signal.

6. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The image enhancement processing step comprises: Apply histogram equalization to X-ray images and infrared thermal images to even out the grayscale distribution of the image and enhance the overall contrast; Calculate the cumulative distribution function of the image and extend it to the full grayscale range by linear interpolation; The Laplacian operator is further applied to perform local contrast enhancement. The edge information in the image is detected by the Laplacian filter, and the enhancement coefficient λ is adjusted.

7. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The feature extraction step of the convolutional neural network includes: Use multiple convolutional layers to extract features of the input multimodal image layer by layer, the convolution kernel size of the convolutional layer is 3x3 or 5x5, and the activation function is the ReLU function; A pooling layer is applied after each convolutional layer, and the pooling layer uses a 2x2 maximum pooling operation; The feature maps after multi-layer convolution and pooling are integrated into a feature vector through a fully connected layer.

8. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The input features in the feature fusion step include: The edge features extracted from the X-ray image are extracted by the Sobel operator to reflect the density changes of the internal structure of the material; The frequency characteristics extracted from the ultrasonic signal are extracted through Fourier transform and reflect the acoustic characteristics inside the material; The temperature abnormality area extracted from the infrared thermal image is identified by the threshold-based heat zone detection algorithm, reflecting the abnormal thermal distribution on the material surface.

9. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The convolutional neural network model used in the defect identification step includes: Add a batch normalization layer on top of the convolutional and pooling layers; Using a fully connected layer with a cross entropy loss function to classify multiple types of defects, the classification results include the type, location and confidence of the defect; During model training, data augmentation techniques are applied, including random cropping, rotation, and scaling.

10. The nondestructive testing method for enhancing defect detection efficiency through deep learning according to claim 1, characterized in that: The results of the defect identification are used to generate a defect assessment report containing the following contents: Classification of defect types, including surface cracks, internal pores, and interlayer separation; The location and size of the defect, including the three-dimensional coordinates, depth and size of the defect in the material; Defect severity assessment, combined with defect density and distribution to generate statistical reports; Recommendations for repairing different types of defects, including which repair techniques to use and the repair steps required.

Images