Cable lead sealing defect identification and uncertainty evaluation method and system

Abstract

The invention provides a cable lead sealing defect identification and uncertainty evaluation method. The method comprises the following steps: acquiring full-matrix capture ultrasonic data of cable lead sealing; processing the full-matrix captured data by a full-focusing method, and reconstructing a full-focusing image of the cable lead sealing; inputting the full-focus image into a probabilistic image segmentation neural network model, and sampling the same input image for multiple times to obtain a plurality of segmentation samples; calculating a pixel-level uncertainty quantitative index, and generating an uncertainty map; and outputting a final defect identification and uncertainty evaluation result of cable lead sealing according to a statistical result of the segmented sample and the uncertainty quantitative index. The invention further discloses a corresponding system. By implementing the method, the accuracy of structural integrity evaluation and defect evaluation can be improved, and the reliability and interpretability of a detection result are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing, and in particular to a method and system for identifying defects and assessing uncertainties in cable lead sealing. Background Technology

[0002] Currently, phased array ultrasonic imaging technology based on the full-matrix acquisition and full-focusing method has become an important technique in the field of nondestructive testing due to its synthetic focusing capability and advantages in reconstructing complex defect patterns. This technology generates high-resolution defect images by acquiring all transmit-receive combined signals and performing time-delay superposition processing on each pixel in the imaging plane. However, in practical applications, even with the full-focusing method, accurate identification and segmentation of defects in the reconstructed images still face many challenges. These challenges mainly stem from the physical characteristics of the defects themselves (such as shadow areas and multiple reflection effects) and scattering noise interference caused by the microstructure of the material.

[0003] In existing technologies, deep learning models, especially convolutional neural networks represented by U-Net, have been widely used in automated defect identification and segmentation tasks of ultrasound images. However, these methods still have significant limitations: First, most models adopt a deterministic modeling approach, outputting only a single segmentation result for a given input image, which cannot effectively reflect the inherent uncertainties in the segmentation process, such as blurred boundary regions caused by imaging physical limitations or defect sizes smaller than wavelength; second, the "spurious certainty" output by deterministic models cannot quantify the model's confidence in the prediction results, making it difficult for operators to determine which regions have reliable segmentation results and which require further verification.

[0004] In nondestructive testing practice, the assessment of inspection reliability typically relies on probability detection curves. However, this method cannot explain why specific defects are difficult to identify and lacks the ability to perform fine-grained analysis of uncertainties in a single inspection. Therefore, existing technologies for defect segmentation of full-focus images generally suffer from insufficient interpretability and the inability to effectively identify and quantify sources of uncertainty, which directly affects the accuracy and reliability of structural integrity assessment. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for identifying defects and assessing uncertainties in cable lead sealing, which can improve the accuracy of structural integrity assessment and defect assessment, and significantly improve the reliability and interpretability of the test results.

[0006] To address the aforementioned technical problems, as one aspect of the present invention, a method for identifying cable lead sealing defects and assessing uncertainties is provided, comprising the following steps: Step S1: Acquire full-matrix capture ultrasound data of the cable lead seal using a phased array two-dimensional array probe; Step S2: Process the full matrix capture data using the full focusing method to reconstruct the full focusing image of the cable lead seal; Step S3: Input the fully focused image into a pre-trained probabilistic image segmentation neural network model, and sample the same input image multiple times to obtain multiple segmentation samples; Step S4: Based on the multiple segmentation samples, calculate the pixel-level uncertainty quantification index and generate an uncertainty map, in which the uncertainty region of the defect segmentation result in the full-focus image is identified; Step S5: Based on the statistical results of the segmented samples and the uncertainty quantification index, output the final defect identification and uncertainty assessment results for the cable lead sealing.

[0007] Preferably, in step S3: The probabilistic image segmentation neural network model is a probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generating the multiple segmentation samples by sampling from the latent space; or The probabilistic image segmentation neural network model is an ensemble U-Net model, which generates the multiple segmentation samples by integrating the prediction results of multiple independently trained deterministic U-Net models; or The probabilistic image segmentation neural network model is a U-Net model with Monte Carlo dropout, which enables a random dropout layer during the model inference phase and generates the multiple segmentation samples through multiple forward propagations.

[0008] The complex shape defect is defined using a randomly generated Bézier curve.

[0009] Preferably, the uncertainty quantification index is a measure H based on information entropy, and its calculation formula is as follows:

[0010] in, This represents the probability of predicting a value of class c given the training data and the current pixel, where C is the set of all classes.

[0011] Preferably, the method further includes a model training step, wherein the dataset used to train the probabilistic image segmentation neural network model is synthetic data generated by a numerical simulation method, and the dataset includes at least one of circular hole defects in lead seal, complex shape defects in lead seal, and circular defects in the microstructure scattering noise of lead seal.

[0012] Preferably, the full-focusing method processing of the full matrix capture data includes: Perform a Hilbert transform on each A-scan signal in the full matrix data to form an analytic signal; Based on the analytical signal, the signal amplitude is coherently superimposed at each pixel in the imaging area; The method further includes a defect size performance analysis step: Calculate the F1 score of the segmentation evaluation index corresponding to defects of different sizes; Analyze the curve of F1 score as a function of defect size; The evaluation model's ability to identify defects smaller than half the wavelength of ultrasound is based on the aforementioned variation curve, as well as its uncertainty. The method further includes: performing correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

[0013] As another aspect of the present invention, a cable lead sealing defect identification and uncertainty assessment system is also provided, comprising: The data acquisition module is used to acquire full-matrix capture ultrasound data of the cable lead seal using a phased array two-dimensional array probe. The image reconstruction module is used to process the full matrix capture data using the full focusing method to reconstruct a full focusing image of the cable lead seal; The probabilistic segmentation module is used to input the fully focused image into a probabilistic image segmentation neural network model, and to sample the same input image multiple times to obtain multiple segmentation samples; An uncertainty quantification module is used to calculate pixel-level uncertainty quantification indices based on the multiple segmentation samples and generate an uncertainty map; the uncertainty map identifies uncertain regions of defect segmentation results in the fully focused image. The result output module is used to output the final defect identification and uncertainty assessment results of the cable lead sealing based on the statistical results of the segmented samples and the uncertainty quantification index.

[0014] Preferably, the probabilistic image segmentation neural network model used by the probabilistic segmentation module is any one of the following: The probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generates the multiple segmentation samples by sampling from the latent space; or An ensemble U-Net model generates the multiple segmentation samples by integrating the predictions from multiple independently trained deterministic U-Net models; or The U-Net model with Monte Carlo dropout enables a random dropout layer during the model inference phase, generating the multiple segmentation samples through multiple forward propagations.

[0015] Preferably, the uncertainty quantification module calculates the information entropy-based metric H using the following formula:

[0016] in, This represents the probability of predicting a value of class c given the training data and the current pixel, where C is the set of all classes.

[0017] Preferably, the system further includes: The model training module is used to train the probabilistic image segmentation neural network model using synthetic data, the synthetic data including at least one of circular hole defects in lead seal, complex shape defects, and circular defects in the microstructure scattering noise of lead seal.

[0018] Preferably, the system further includes: The defect size performance analysis module is used to calculate the F1 score, a segmentation evaluation index, for defects of different sizes; analyze the change curve of the F1 score with defect size; and evaluate the model's ability to identify defects smaller than half the wavelength of ultrasound based on the change curve.

[0019] The correlation analysis module is used to perform correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

[0020] Implementing this embodiment has the following beneficial effects: This invention provides a method and system for identifying defects in cable lead sealing and assessing uncertainty. It generates an uncertainty map by outputting multiple segmented samples through a probabilistic neural network and combining information entropy calculation, thereby improving the interpretability of the detection results. In embodiments of the present invention, probabilistic segmentation and multiple sampling are used to improve the accuracy and reliability of identifying defects smaller than λ / 2.

[0021] In embodiments of the present invention, defect segmentation and reliability assessment can be integrated: the output includes not only the defect location and shape, but also the reliability of the segmentation results, supporting more comprehensive non-destructive testing decisions.

[0022] In embodiments of the present invention, the generalization ability and robustness of the model can be improved: by probabilistic modeling and training with synthetic data, the model's adaptability to complex defects and noise can be enhanced.

[0023] This invention achieves precise defect segmentation and reliability assessment through a complete process, providing more comprehensive decision support for nondestructive testing. The final output includes not only the location and shape information of the defect, but also a reliable quantification of the segmentation results, significantly improving the reliability and interpretability of the test results. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention. Figure 1 This is a schematic diagram of the main flow of an embodiment of a cable lead sealing defect identification and uncertainty assessment method provided by the present invention; Figure 2 yes Figure 1 A schematic diagram of the full-focus image reconstruction process involved in the process; Figure 3 yes Figure 1 A schematic diagram of the probabilistic segmentation and reasoning process involved; Figure 4 This is a schematic diagram of an embodiment of a cable lead sealing defect identification and uncertainty assessment system provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] like Figure 1 The diagram shown illustrates the main flowchart of an embodiment of a cable lead sealing defect identification and uncertainty assessment method provided by the present invention; in conjunction with... Figure 2 and Figure 3 As shown, in this embodiment, the method aims to segment defects in ultrasonic images of lead-sealed high-voltage cables using a probabilistic deep learning model based on full-matrix capture and full-focusing methods, and to quantify the uncertainty of the segmentation results. The entire process can be divided into the following main stages: obtaining three-dimensional full-matrix capture data through a two-dimensional phased array probe, TFM image reconstruction, probabilistic neural network segmentation and inference, and uncertainty quantification and result analysis.

[0027] More specifically, in this embodiment, the method includes at least: Step S1: Acquire full matrix capture ultrasound data of the cable lead seal using a phased array two-dimensional area array probe. Specifically, acquire the three-dimensional full matrix capture raw data and binary image mask.

[0028] A two-dimensional phased array probe is used to inspect the lead seal of high-voltage cables. Because a two-dimensional ultrasonic phased array probe is used, a complete three-dimensional full-matrix raw dataset can be directly acquired through hardware, with one dimension being... A three-dimensional matrix (where It is the number of array elements. (This refers to the number of time sampling points for each A-scan signal). In this invention, the original data matrix is ​​referred to as the information matrix. Its elements Indicates the first Each element is launched, the first The time-domain signal received by each array element.

[0029] A binary image mask is generated, with each pixel corresponding one-to-one with the pixels of the subsequent full-focus image. Regions containing defects are labeled 1 (foreground), and non-defect regions are labeled 0 (background). This labeling is generated based on the precise geometric location of the defects in the numerical model.

[0030] Step S2: Process the full-matrix capture data using the full-focusing method to reconstruct a full-focus image of the cable lead seal; specifically, this step involves reconstructing the full-matrix capture image. The purpose of this step is to convert the original time-domain signal data from the full-matrix capture into an energy-focused grayscale image to facilitate subsequent visual analysis and neural network processing.

[0031] The input signal in this stage is the original dataset of the three-dimensional full matrix capture obtained in step S1, namely the information matrix A.

[0032] First, the information matrix Each A scan signal in Perform Hilbert transform To form an analytical signal.

[0033] The core objective of this transformation is to accurately extract the envelope of the original real signal by constructing an analytic signal. This envelope smoothly outlines the time-varying profile of the signal energy, effectively stripping away the rapid oscillations caused by the center frequency of the ultrasound wave, thus laying the foundation for subsequent imaging. Simultaneously, it preserves all the frequency information of the signal, ensuring the fidelity of subsequent processing. (Analytic signal) The calculation formula is:

[0034] in It is the imaginary unit.

[0035] Then, after completing the signal preprocessing and obtaining the analyzed signal... Then, the process enters the core compositing and focusing stage, which is the superposition of pixel amplitudes.

[0036] This step aims to calculate each specific pixel in the imaging plane. The final intensity or brightness value To achieve this, the algorithm first needs to process each pair of transmitting array elements. and receiving array elements Calculate the ultrasonic waves from the transmitting array element propagate to target pixel After being reflected at that point, it reaches the receiving array element. Total path time required The calculation of this time depends strictly on the longitudinal wave velocity of the lead seal of the high-voltage cable. .

[0037] Subsequently, the algorithm performs coherent superposition: for the current pixel... From all Analysis of the signal of a transmit-receive pair In the middle, extract the corresponding calculation time respectively. The signal complex values ​​at each time point. These complex values ​​from all channels are vector-summed, and the magnitude of this summation is defined as the intensity value of that pixel.

[0038] The essence of this operation is the coherent convergence of energy. By performing complex vector superposition on the analytic signal (rather than the original real signal), the destructive interference problem caused by phase difference is effectively avoided, ensuring that the reflected signal from the real defect region can be superimposed, thus appearing as a bright pixel in the image.

[0039] Finally, by traversing all pixels within the imaging plane and repeating this process, a complete, clear, and well-focused full-focus image of the lead seal of the high-voltage cable can be generated.

[0040] After superimposing the pixel magnitudes of the entire matrix data, the resulting image is a two-dimensional matrix composed of complex values, whose range is typically large and uncertain. To transform this intermediate result into a standard input format suitable for subsequent neural network processing, image normalization is necessary. The input signal for this step is the output of the previous stage. As an input signal.

[0041] The specific mathematical processing for image normalization is as follows: First, traverse the entire image matrix and find the global maximum value among all pixel values. and global minimum Then, for each pixel value in the image... Perform a linear transformation to map it to the closed interval [0, 1]. The formula for this transformation is:

[0042] In this formula, the minimum value is subtracted from each pixel value. This achieves data shifting, adjusting the lower limit of the numerical range to zero; subsequently, it is divided by the numerical range. This achieves data scaling, compressing the entire range to between 0 and 1. After this operation, the highest value representing the strongest reflection in the original signal is mapped to 1, while the lowest value representing the background noise level is mapped to 0, and all intermediate values ​​are distributed proportionally in between.

[0043] The output signal of this step is a normalized TFM grayscale image. This image is a [image with dimensions of height]. Multiply by width The two-dimensional matrix, where the grayscale value of each pixel no longer represents the original physical voltage or absolute energy, but rather the relative intensity of the ultrasonic reflected signal at that spatial location. This standardized format is crucial for deep learning models, eliminating the overall gain differences that may exist in the original data due to different detection conditions each time, ensuring the stability and consistency of the neural network input, thereby greatly promoting the convergence speed and final performance of model training.

[0044] Finally, this normalized TFM grayscale image is matched with the segmentation ground truth labels prepared in the data generation stage to form a complete training sample pair. This sample serves as the output signal for this step.

[0045] Step S3: Input the fully focused image into a pre-trained probabilistic image segmentation neural network model, and sample the same input image multiple times to obtain multiple segmentation samples; In step S3: The probabilistic image segmentation neural network model is a probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generating the multiple segmentation samples by sampling from the latent space; or The probabilistic image segmentation neural network model is an ensemble U-Net model, which generates the multiple segmentation samples by integrating the prediction results of multiple independently trained deterministic U-Net models; or The probabilistic image segmentation neural network model is a U-Net model with Monte Carlo dropout, which enables a random dropout layer during the model inference phase and generates the multiple segmentation samples through multiple forward propagations.

[0046] In a specific example, the above output signal is input into the Probabilistic U-Net model for training and inference.

[0047] The input signal is: the normalized TFM grayscale image X from the previous step, and the segmentation true label Y prepared in the data generation stage, forming a complete training sample pair (X,Y).

[0048] 1) Then the U-Net backbone network processes the input normalized TFM grayscale image X to generate preliminary image features.

[0049] 2) The prior encoder learns the prior distribution parameters of the latent space z using only X as input:

[0050] 3) The posterior encoder receives both X and the true label Y simultaneously, and calculates the posterior distribution parameters of the latent variables:

[0051] 4) Training is achieved by minimizing a joint loss function:

[0052] The first term, reconstruction loss, ensures that z sampled from the posterior distribution can accurately reconstruct the true label Y through the decoder (U-Net); the second term, KL divergence, forces the posterior distribution to move closer to the prior distribution, so that after training, the prior encoder can provide a meaningful latent distribution for a new input image.

[0053] After the four steps described above, the model completes the training phase and proceeds to the next inference phase. It is understood that in the model training step, the dataset used to train the probabilistic image segmentation neural network model is synthetic data generated through numerical simulation methods. This dataset contains at least one of the following: circular hole defects in lead seals, complex-shaped defects in lead seals, and circular defects in the microstructure scattering noise of lead seals. In a specific example, complex-shaped defects are defined using randomly generated Bézier curves.

[0054] For a new normalized TFM grayscale image X, the posterior encoder is no longer used (because the posterior encoder needs to use the encoded label Y).

[0055] 1) Firstly, because the prior encoder can calculate the prior distribution parameters

[0056] 2) Subsequently, k independent samples are taken from this distribution to obtain k potential vectors z. k ~ Each z k The image features are fused with those in the U-net decoder to independently generate a segmentation probability map S. K Therefore, for the same input image X, the model outputs K segmentation samples. Each segmented sample S K They are all H×W matrices.

[0057] The final output signal of this process—K segmentation samples—is directly used as input for the next stage, "uncertainty quantification and result analysis." By performing statistical calculations on these samples, the model not only provides the location information of the defects, but more importantly, it quantifies the confidence level of the predictions, accurately revealing the segmentation ambiguity caused by physical phenomena, thereby achieving interpretable and evaluable intelligent segmentation of images obtained through the full-focus method.

[0058] Step S4: Based on the multiple segmentation samples, calculate the pixel-level uncertainty quantification index and generate an uncertainty map, in which the uncertainty region of the defect segmentation result in the full-focus image is identified; Specifically, this step is used to perform statistical analysis on the results of multiple segmentation samplings, generate the final segmentation results and the corresponding uncertainty map, thereby providing interpretable confidence information for defect assessment.

[0059] The input signal is: K segmented samples

[0060] K segmentation samples 𝑘 The probability values ​​at each pixel location (h, w) are arithmetically averaged to obtain a consensus probability map that integrates information from all samples. Its mathematical expression is:

[0061] in, Representing the probability value of the k-th sample at pixel (h, w), this average segmentation map is an H×W matrix where each pixel value represents the model's average confidence that the point is defective. It integrates the consensus of multiple predictions and is usually more accurate than any single sampling result.

[0062] Uncertainty quantification is then performed by calculating the information entropy H for each pixel. For a given pixel, its K predictions constitute an empirical probability distribution. The uncertainty metric H is based on information entropy and is calculated based on this distribution, as shown in the following formula:

[0063] in, This represents the probability of predicting a value as class c given the training data and the current pixel, where C = {defect, background} is the set of all classes, and p... c The uncertainty map is the average probability predicted for that category (for the defect class, it's the average probability map value p at that point; for the background class, it's 1−p). The uncertainty map is an H×W matrix, where each pixel value is its entropy value H. The entropy value reaches its maximum when the predicted probability p is close to 0.5, indicating that the model is most hesitant and uncertain about classifying the pixel; when p is close to 0 or 1, the entropy value approaches 0, indicating that the model is very certain that the pixel belongs to either the background or the defect.

[0064] The third step is result analysis and correlation. First, the average probability segmentation map is binarized by setting a threshold, which is 0.5 in this invention, to obtain the final binary segmentation result, which clearly marks the predicted defect area.

[0065] Subsequently, the segmentation results were compared and analyzed in conjunction with the uncertainty map and the original TFM image.

[0066] The output signals in this step are: The output signals in this stage include: Average probability segmentation plot: Represents the average confidence level of the model.

[0067] Final binary segmentation result: Clear defect localization obtained through thresholding.

[0068] Uncertainty Map: Visually displays the spatial distribution of the reliability of prediction results.

[0069] Step S5: Based on the statistical results of the segmented samples and the uncertainty quantification index, output the final defect identification and uncertainty assessment results for the cable lead sealing.

[0070] Furthermore, in this embodiment of the invention, the full-focusing method processing of the full matrix capture data includes: Perform a Hilbert transform on each A-scan signal in the full matrix data to form an analytic signal; Based on the analytical signal, the signal amplitude is coherently superimposed at each pixel in the imaging area; The method further includes a defect size performance analysis step: Calculate the F1 score of the segmentation evaluation index corresponding to defects of different sizes; Analyze the curve of F1 score as a function of defect size; The evaluation model's ability to identify defects smaller than half the wavelength of ultrasound is based on the aforementioned variation curve, as well as its uncertainty. The method further includes: performing correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

[0071] It is understood that this invention relates to a method for defect identification and uncertainty assessment in ultrasound total focusing images based on probabilistic deep learning, aiming to solve the problem of difficulty in quantifying the reliability of defect segmentation results in traditional nondestructive testing. This method implements a complete technical chain from data acquisition to uncertainty assessment through a systematic process.

[0072] The first stage involves data acquisition and processing. A phased array probe is used to scan the lead seal of the cable joint or cable terminal from multiple angles to obtain ultrasonic signal echoes, resulting in a large amount of training data. Full-matrix capture technology is then used to acquire complete ultrasonic signal data. The output of this stage is a full-matrix dataset containing complete signal information and corresponding defect-annotated images.

[0073] Next, we proceed to the image reconstruction stage. The acquired full-matrix data is processed using a full-focusing method, which involves two key steps. First, a Hilbert transform is performed on each ultrasound signal to extract the signal envelope and obtain smoother signal features. Then, signal amplitude superposition calculations are performed at each pixel in the imaging region, and a clear defect image is reconstructed using the synthetic aperture focusing principle. Finally, the reconstructed image is normalized and converted into a grayscale image format suitable for neural network processing.

[0074] The third stage is probabilistic neural network segmentation and inference. Normalized, fully focused images are input into a specially designed probabilistic U-Net network for training and inference. This network structure adds a probabilistic inference mechanism to the traditional U-Net, enabling it to learn the latent distribution characteristics of defect segmentation. During inference, the same input image is sampled multiple times to obtain a set of possible segmentation results, reflecting the model's understanding of the uncertainty in defect boundaries and shapes.

[0075] Finally, uncertainty quantification and result analysis are performed. Statistical analysis is conducted on multiple segmentation samples output by the neural network to calculate the average probability value of each pixel, forming the final segmentation result. Simultaneously, the uncertainty measure of each pixel is calculated based on information entropy theory. By setting an appropriate threshold, the probability map is converted into a binary segmentation result, which is then combined with the uncertainty map for comprehensive evaluation. Special attention is paid to the correlation between high-uncertainty regions and physical phenomena, such as typical cases like defect shadow areas, multiple reflection signals, and small-sized defects.

[0076] This method achieves accurate defect segmentation and reliability assessment through a complete workflow, providing more comprehensive decision support for nondestructive testing. The final output includes not only the location and shape information of the defect, but also a reliable quantification of the segmentation results, significantly improving the reliability and interpretability of the detection results.

[0077] like Figure 4 The diagram shows a schematic representation of an embodiment of a cable lead sealing defect identification and uncertainty assessment system provided by the present invention. In this embodiment, the system 1 includes at least: Data acquisition module 10 is used to acquire full-matrix capture ultrasound data of cable lead sealing through a phased array two-dimensional array probe; Image reconstruction module 11 is used to process the full matrix capture data using the full focusing method to reconstruct a full focusing image of the cable lead seal; The probabilistic segmentation module 12 is used to input the fully focused image into the probabilistic image segmentation neural network model, and to sample the same input image multiple times to obtain multiple segmentation samples; Uncertainty quantification module 13 is used to calculate pixel-level uncertainty quantification index based on the multiple segmentation samples and generate an uncertainty map; the uncertainty map identifies the uncertain regions of the defect segmentation results in the fully focused image; The result output module 14 is used to output the final defect identification and uncertainty assessment results of the cable lead sealing based on the statistical results of the segmented samples and the uncertainty quantification index.

[0078] The probabilistic image segmentation neural network model used by the probabilistic segmentation module 12 is any one of the following: The probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generates the multiple segmentation samples by sampling from the latent space; or An ensemble U-Net model generates the multiple segmentation samples by integrating the predictions from multiple independently trained deterministic U-Net models; or The U-Net model with Monte Carlo dropout enables a random dropout layer during the model inference phase, generating the multiple segmentation samples through multiple forward propagations.

[0079] The uncertainty quantification module 13 calculates the information entropy-based metric H using the following formula:

[0080] in, This represents the probability of predicting a value of class c given the training data and the current pixel, where C is the set of all classes.

[0081] In a specific embodiment, system 1 further includes: The model training module 15 is used to train the probabilistic image segmentation neural network model using synthetic data, the synthetic data including at least one of circular hole defects in the lead seal, complex shape defects, and circular defects in the microstructure scattering noise of the lead seal.

[0082] In a specific embodiment, system 1 further includes: The defect size performance analysis module 16 is used to calculate the segmentation evaluation index F1 score corresponding to defects of different sizes; analyze the change curve of F1 score with defect size; and evaluate the model's ability to identify defects smaller than half the wavelength of ultrasound based on the change curve.

[0083] The correlation analysis module 17 is used to perform correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

[0084] For more details, please refer to and combine with the above. Figures 1 to 3 The description of that will not be repeated here.

[0085] Implementing this embodiment has the following beneficial effects: This invention provides a method and system for identifying defects in cable lead sealing and assessing uncertainty. It generates an uncertainty map by outputting multiple segmented samples through a probabilistic neural network and combining information entropy calculation, thereby improving the interpretability of the detection results. In embodiments of the present invention, probabilistic segmentation and multiple sampling are used to improve the accuracy and reliability of identifying defects smaller than λ / 2.

[0086] In embodiments of the present invention, defect segmentation and reliability assessment can be integrated: the output includes not only the defect location and shape, but also the reliability of the segmentation results, supporting more comprehensive non-destructive testing decisions.

[0087] In embodiments of the present invention, the generalization ability and robustness of the model can be improved: by probabilistic modeling and training with synthetic data, the model's adaptability to complex defects and noise can be enhanced.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for identifying lead sealing defects and assessing uncertainties in cables, characterized in that, Includes the following steps: Step S1: Acquire full-matrix capture ultrasound data of the cable lead seal using a phased array two-dimensional array probe; Step S2: Process the full matrix capture data using the full focusing method to reconstruct the full focusing image of the cable lead seal; Step S3: Input the fully focused image into a pre-trained probabilistic image segmentation neural network model, and sample the same input image multiple times to obtain multiple segmentation samples; Step S4: Based on the multiple segmentation samples, calculate the pixel-level uncertainty quantification index and generate an uncertainty map, in which the uncertainty region of the defect segmentation result in the full-focus image is identified; Step S5: Based on the statistical results of the segmented samples and the uncertainty quantification index, output the final defect identification and uncertainty assessment results for the cable lead sealing.

2. The method according to claim 1, characterized in that, In step S3: The probabilistic image segmentation neural network model is a probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generating the multiple segmentation samples by sampling from the latent space; or The probabilistic image segmentation neural network model is an ensemble U-Net model, which generates the multiple segmentation samples by integrating the prediction results of multiple independently trained deterministic U-Net models; or The probabilistic image segmentation neural network model is a U-Net model with Monte Carlo dropout, which enables a random dropout layer during the model inference phase and generates the multiple segmentation samples through multiple forward propagations.

3. The method according to claim 1, characterized in that, The uncertainty quantification index is a measure H based on information entropy, and its calculation formula is as follows: in, This represents the probability of predicting a class c given the training data and the current pixel, where C is the set of all classes.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes a model training step, wherein the dataset used to train the probabilistic image segmentation neural network model is synthetic data generated by a numerical simulation method, and the dataset includes at least one of circular hole defects in lead seal, complex shape defects in lead seal, and circular defects in the microstructure scattering noise of lead seal. The complex shape defect is defined using a randomly generated Bézier curve.

5. The method according to claim 4, characterized in that, in, The full-focusing method for processing the full matrix capture data includes: Perform a Hilbert transform on each A-scan signal in the full matrix data to form an analytic signal; Based on the analytical signal, the signal amplitude is coherently superimposed at each pixel in the imaging area; The method further includes a defect size performance analysis step: Calculate the F1 score of the segmentation evaluation index corresponding to defects of different sizes; Analyze the curve of F1 score as a function of defect size; The evaluation model's ability to identify defects smaller than half the wavelength of ultrasound is based on the aforementioned variation curve, as well as its uncertainty. The method further includes: performing correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

6. A cable lead sealing defect identification and uncertainty assessment system, characterized in that, include: The data acquisition module is used to acquire full-matrix capture ultrasound data of the cable lead seal using a phased array two-dimensional array probe. The image reconstruction module is used to process the full matrix capture data using the full focusing method to reconstruct a full focusing image of the cable lead seal; The probabilistic segmentation module is used to input the fully focused image into a probabilistic image segmentation neural network model, and to sample the same input image multiple times to obtain multiple segmentation samples; The uncertainty quantification module is used to calculate pixel-level uncertainty quantification indicators based on the multiple segmented samples and generate an uncertainty map. The uncertainty map identifies the uncertain regions in the defect segmentation results of the fully focused image; The result output module is used to output the final defect identification and uncertainty assessment results of the cable lead sealing based on the statistical results of the segmented samples and the uncertainty quantification index.

7. The system according to claim 6, characterized in that, The probabilistic image segmentation neural network model used by the probabilistic segmentation module is any one of the following: The probabilistic U-Net model, which combines the U-Net encoder-decoder structure with a conditional variational autoencoder, generates the multiple segmentation samples by sampling from the latent space; or An integrated U-Net model is used to generate the multiple segmentation samples by integrating the prediction results of multiple independently trained deterministic U-Net models. or The U-Net model with Monte Carlo dropout enables a random dropout layer during the model inference phase, generating the multiple segmentation samples through multiple forward propagations.

8. The system according to claim 6, characterized in that, The uncertainty quantification module calculates the information entropy-based metric H using the following formula: in, This represents the probability of predicting a class c given the training data and the current pixel, where C is the set of all classes.

9. The system according to any one of claims 6 to 8, characterized in that, The system also includes: The model training module is used to train the probabilistic image segmentation neural network model using synthetic data, the synthetic data including at least one of circular hole defects in lead seal, complex shape defects, and circular defects in the microstructure scattering noise of lead seal.

10. The system according to claim 6, characterized in that, The system also includes: The defect size performance analysis module is used to calculate the F1 score, a segmentation evaluation index, for defects of different sizes; analyze the change curve of the F1 score with defect size; and evaluate the model's ability to identify defects smaller than half the wavelength of ultrasound based on the change curve. The correlation analysis module is used to perform correlation analysis between high uncertainty areas in the uncertainty map and preset physical phenomenon types; the physical phenomenon types include at least one of defect shadow areas, multiple reflection signals, and small-sized defects.

Images