Real-time and off-line detection system and method for weld defects

Abstract

The invention discloses a real-time and off-line detection system and method for weld defects. The system comprises a planar electromagnetic chromatography sensor array, a multi-channel excitation and acquisition unit, a data processing and imaging unit, a prototype guide segmentation network unit and an information processing platform, the data processing and imaging unit is in communication connection with the prototype guide segmentation network unit and can input the preprocessed image into the prototype guide segmentation network unit, features are extracted through an encoder, spatial resolution is recovered step by step through a decoder, a prototype similarity graph is calculated through prototype branches, then the prototype similarity graph is fused with the features of the decoder, and a prototype image is obtained. And outputting the defect probability graph P and processing the defect probability graph P through an information processing platform to realize automatic judgment of weld defects. The method has the advantages that high robustness, high sensitivity and low omission ratio can be realized on the surface of a complex welding seam, and the technical problems that the detection sensitivity of superficial, fine and inclined defects is insufficient, real-time imaging and high-precision off-line reconstruction are difficult to consider, the omission ratio is difficult to reduce and the like are solved.

Description

Technical Field

[0001] This invention relates to a real-time and offline detection system and method for weld defects based on planar electromagnetic tomography (EMT) and deep learning segmentation networks, belonging to the field of electromagnetic nondestructive testing and industrial intelligent inspection technology. Background Technology

[0002] Welds are critical connection points in pressure vessels, pipelines, rail transit, and large steel structures. Internal or surface defects (such as cracks, lack of fusion, porosity, and slag inclusions) can significantly reduce the structural load-bearing capacity and, in severe cases, lead to catastrophic accidents. Therefore, developing non-destructive testing technologies with high sensitivity and reliability under complex weld surface conditions is of significant engineering importance. Existing X-ray, ultrasonic, and array eddy current testing methods are widely used in weld inspection, but they still have the following shortcomings:

[0003] (1) Severe interference from uneven weld toe and rough surface: Weld surface usually has weld wave undulation, spatter and grinding marks, which makes ultrasonic testing highly sensitive to coupling conditions. The background signal of array eddy current testing can easily cover the shallow or tilted defect signal.

[0004] (2) Insufficient directional sensitivity: Conventional array eddy currents usually only use adjacent coil pairs for measurement, which has limited ability to distinguish cracks with different orientations;

[0005] (3) It is difficult to balance the real-time performance and accuracy of imaging algorithms: Non-iterative reconstruction algorithms such as linear back projection (LBP) and Tikhonov regularization have fast computation speed, but their imaging blur and noise suppression capabilities are limited; iterative algorithms such as Landweber (LW) iteration have higher imaging accuracy, but their computational load is large and they are difficult to use for online real-time detection.

[0006] (4) Defect regions are difficult to segment accurately: Even if electromagnetic tomography reconstruction images are obtained, traditional thresholding or simple image processing methods are difficult to accurately segment defect regions in low-contrast, small-size defect scenarios. Especially when the number of samples is limited, deep learning networks are prone to overfitting and have a high false negative rate.

[0007] Therefore, there is an urgent need for an integrated detection solution that can perform real-time coarse screening and high-precision offline analysis to achieve high robustness, high sensitivity and low false negative rate on complex weld surfaces. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a real-time and offline detection system and method for weld defects with real-time imaging, accurate segmentation and offline high-fidelity reconstruction capabilities, thereby achieving high robustness, high sensitivity and low false negative rate on complex weld surfaces.

[0009] To address the aforementioned technical problems, the present invention provides a real-time and offline weld defect detection system, comprising a planar electromagnetic tomography sensor array, a multi-channel excitation and acquisition unit for exciting and acquiring data from the planar electromagnetic tomography sensor array, and a data processing and imaging unit connected to the multi-channel excitation and acquisition unit for performing reference signal acquisition, differential voltage calculation, sensitivity matrix construction, real-time LBP reconstruction, offline LW reconstruction, and preprocessing of the reconstructed image. It also includes a prototype-guided segmentation network unit with an encoder, decoder, prototype branch module, and similarity fusion module, and an information processing platform communicatively connected to the prototype-guided segmentation network unit. The data processing and imaging unit is communicatively connected to the prototype-guided segmentation network unit and can input the preprocessed image into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Subsequently, the model is fused with the decoder features to output a defect probability map P, which is then processed by the information processing platform to achieve automatic weld defect determination.

[0010] The planar electromagnetic tomography sensor array consists of multiple planar electromagnetic coils arranged in two staggered rows and a magnetic core, with each coil wound on the magnetic core. The planar electromagnetic tomography sensor is installed above the weld to be measured, and the axis of the planar electromagnetic coil is perpendicular to the weld surface. The multi-channel excitation and acquisition unit sequentially selects each coil as the excitation coil and uses the remaining coils as receiving coils to perform measurements. It can also transmit the measurement data to the data processing and imaging unit via UDP or bus.

[0011] The data processing and imaging unit includes a reference voltage vector V for acquiring the reference voltage vector V under defect-free weld or plate conditions. ref The reference signal acquisition module is used to acquire the voltage vector V under the measured state of the weld. sample And calculate the differential signal ΔV = V sample -V ref The differential voltage calculation module, based on finite element simulation, constructs a sensitivity matrix S containing only adjacent and next-nearest coil pairs, with corresponding coil pair spacings of 1 and 2 times the coil center spacing. The simplified sensitivity matrix construction module for long-distance coil pairs is discarded. A linear back-projection algorithm σ = S is employed. T A real-time LBP reconstruction module that maps differential voltage to a two-dimensional weld cross-section conductivity / permeability perturbation image, and a Landweber iterative algorithm σ for stored detection data. (k+1) =σ (k) +αS T (ΔV-Sσ (k) An offline LW reconstruction module that performs multiple iterations to obtain reconstructed images with higher spatial resolution and contrast.

[0012] The encoder extracts multi-scale deep features based on a residual network; the decoder uses UNet-style skip connections to upsample multi-scale features step by step and restore spatial resolution; the prototype branch module globally aggregates the high-level feature maps of the encoder to obtain class prototype vectors for the defect class and the background class, calculates the similarity between each pixel location feature and the class prototype using cosine similarity, generates a prototype similarity map, and the prototype uses an exponential moving average (EMA) update strategy to continuously update the running prototype during training and inference; the similarity fusion module concatenates the prototype similarity map and the decoder output features in the channel dimension, and fuses them through 1×1 convolution to generate the final defect probability map.

[0013] The information processing platform is also connected to a visualization and interaction unit.

[0014] The visualization and interaction unit can provide a multi-window interface to display I / Q waveforms, real-time LBP reconstruction maps, defect probability heatmaps and binary segmentation results in real time, and supports user adjustment of thresholds, display of historical frame sequences, and switching between real-time and offline modes.

[0015] A detection method using the above-mentioned real-time and offline weld defect detection system includes the following steps:

[0016] S1: Reference Data Acquisition

[0017] Under defect-free weld or reference specimen conditions, the multi-channel excitation and acquisition unit sequentially excites each coil, acquires the voltage vector of all selected coil pairs, and calculates and stores the reference voltage V. ref .

[0018] S2: Data acquisition and differential analysis of the weld under test

[0019] A planar electromagnetic tomography sensor array is arranged above the weld to be measured, and the excitation and acquisition process in step S1 is repeated to obtain the measured voltage vector V. sample Calculate the differential signal ΔV = V sample -V ref ;

[0020] S3: Real-time LBP imaging based on a simplified sensitivity matrix

[0021] Using a simplified sensitivity matrix S containing only adjacent and next-nearest coil pairs, the two-dimensional perturbation image σ = S is calculated via linear backprojection. T ΔV is calculated and mapped to a predefined grid to obtain a real-time weld cross-section image;

[0022] S4: Image Reconstruction Preprocessing

[0023] The real-time reconstructed image is normalized, interpolated and scaled to a specified size, and lightweight morphological operations and denoising are performed as input to the segmentation network.

[0024] S5: Prototype-guided segmentation network inference

[0025] The preprocessed image is input into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores the spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Then, it is fused with the decoder features to output the defect probability map P.

[0026] S6: Post-segmentation processing and defect assessment

[0027] The defect probability map is binarized using an adjustable threshold τ to obtain a defect region mask. Isolated noise is removed and small holes are filled through morphological opening and closing operations. The area, length, position and orientation of the defect are calculated to achieve automatic identification and visualization of weld defects.

[0028] S7: Offline high-precision reconstruction (optional):

[0029] For the stored detection data, the Landweber iterative algorithm combined with the full sensitivity matrix is ​​used to perform offline high-precision reconstruction as needed, so as to obtain weld imaging results with higher signal-to-noise ratio and spatial resolution. Then, the data is input into the prototype guided segmentation network unit for offline high-precision segmentation and analysis.

[0030] The advantages of this invention are:

[0031] By designing a 16-channel staggered planar EMT coil array and constructing a simplified sensitivity matrix based on "adjacent and next-nearest coil pairs," redundant measurements are reduced and real-time performance is improved. Furthermore, the LBP fast reconstruction algorithm is combined for online real-time monitoring, and Landweber iteration is used for offline high-precision imaging. Additionally, a prototype-guided similarity fusion UNet (ProSiF-UNet) segmentation network is employed to fuse EMT reconstruction priors with encoder deep features, reducing the false negative rate and enhancing the segmentation capability for slender, low-contrast defects. Thus, by combining a planar electromagnetic tomography sensor, an optimized simplified sensitivity matrix for coil pair selection, and a deep learning segmentation network incorporating a prototype-guided mechanism, real-time imaging, accurate segmentation, and offline high-fidelity reconstruction capabilities are achieved. This enables high robustness, high sensitivity, and low false negative rates on complex weld surfaces, effectively solving technical challenges such as insufficient sensitivity for shallow, small, and tilted defects, the difficulty in balancing real-time imaging and high-precision offline reconstruction, and the difficulty of reducing false negative rates in deep learning segmentation under small sample conditions. Attached Figure Description

[0032] Figure 1This is a schematic diagram of the planar EMT coil array and the weld seam being measured according to the present invention.

[0033] Figure 2 A schematic diagram of the sensitivity matrix construction (showing the coil pair arrangement and spatial grid division);

[0034] Figure 3 To simplify the schematic diagram of the distribution of second-neighbor coil pairs in the sensitivity matrix;

[0035] Figure 4 This is a schematic diagram of the real-time weld defect reconstruction results based on LBP.

[0036] Figure 5 This is a schematic diagram of the offline high-precision weld reconstruction results based on Landweber iteration;

[0037] Figure 6 A structural block diagram of a prototype-guided segmentation network unit;

[0038] Figure 7 This is a schematic diagram of the software interface of the real-time weld defect detection system of the present invention. Detailed Implementation

[0039] The real-time and offline detection system and method for weld defects of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1:

[0041] like Figure 1 As shown, the real-time and offline weld defect detection system of this embodiment includes a planar electromagnetic tomography sensor array, a multi-channel excitation and acquisition unit for exciting and acquiring data from the planar electromagnetic tomography sensor array, and a data processing and imaging unit connected to the multi-channel excitation and acquisition unit for performing reference signal acquisition, differential voltage calculation, sensitivity matrix construction, real-time LBP reconstruction, offline LW reconstruction, and preprocessing of the reconstructed image from the data from the multi-channel excitation and acquisition unit. The planar electromagnetic tomography sensor array includes 16 planar electromagnetic coils arranged in two staggered rows with equal center-to-center spacing. Each coil... The coils are wound on a magnetic core, with strict tolerance control over coil position and core height to ensure array consistency. The sensor is mounted above the weld to be measured, with the coil axis substantially perpendicular to the weld surface. The multi-channel excitation and acquisition unit sequentially selects each coil as the excitation coil and uses the remaining coils as receiving coils for measurement. The preferred excitation frequency is 40kHz, and the excitation current is 1A. It supports the acquisition of multi-channel complex voltage signals (I / Q components) and transmits the measurement data to the data processing and imaging unit via UDP or bus. The data processing and imaging unit includes a reference signal acquisition module: acquiring the reference voltage vector V under defect-free weld or plate conditions.ref Differential voltage calculation module: Acquires voltage vector V under the measured state of the weld. sample And calculate the differential signal ΔV = V sample -V ref Simplified Sensitivity Matrix Construction Module: Based on finite element simulation, a sensitivity matrix S is constructed that contains only adjacent and second-to-nearest coil pairs, with corresponding coil pair spacings of 1 and 2 times the coil center-to-center spacing, discarding distant coil pairs; Real-time LBP Reconstruction Module: Employs the linear back-projection algorithm σ = S T ΔV maps the differential voltage to a two-dimensional weld cross-section conductivity / permeability perturbation image; Offline LW reconstruction module: For the stored detection data, the Landweber iterative algorithm σ is used. (k+1) =σ (k) +αS T (ΔV-Sσ (k) The process involves multiple iterations to obtain a reconstructed image with higher spatial resolution and contrast; the preprocessing module normalizes, denoises, and smooths the reconstructed image, and crops and scales it proportionally according to the imaging area.

[0042] It also includes a prototype-guided segmentation network unit with an encoder, decoder, prototype branch module, and similarity fusion module, as well as an information processing platform communicating with the prototype-guided segmentation network unit. In the prototype-guided segmentation network unit (ProSiF-UNet), the encoder extracts multi-scale deep features based on a residual network (such as ResNet-34); the decoder uses UNet-style skip connections to progressively upsample the multi-scale features and restore spatial resolution; the prototype branch module globally aggregates the high-level feature maps of the encoder to obtain class prototype vectors for the defect class and the background class; and cosine similarity is used to calculate the feature at each pixel location and... The similarity of class prototypes is used to generate a prototype similarity map. The prototypes are updated using an exponential moving average (EMA) update strategy, continuously updating the running prototypes during training and inference. The similarity fusion module concatenates the prototype similarity map with the decoder output features along the channel dimension, and then fuses them through a 1×1 convolution to generate the final defect probability map. The loss function uses a combination of Tversky loss and Focal binary cross-entropy to enhance the recall rate for small, sparse defects and reduce the false negative rate. (During network training, Tversky loss and Focal binary cross-entropy loss are first calculated based on the predicted probability map and the ground truth mask, respectively, and then calculated according to L = λL...) Tv +(1-λ)L focalThe total loss is calculated by weighted summation. Tversky loss enhances recall of small, sparse defect pixels by increasing the penalty for FN; Focal loss suppresses the weight of easily distinguishable samples and highlights difficult-to-distinguish defect regions, thus working together to reduce the false negative rate. The data processing and imaging unit communicates with the prototype-guided segmentation network unit and can input the preprocessed image into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores the spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Then, it is fused with the decoder features to output the defect probability map P. After processing by the information processing platform, the weld defect is automatically determined.

[0043] Furthermore, the information processing platform is also connected to a visualization and interaction unit, which can provide a multi-window interface to display I / Q waveforms, real-time LBP reconstruction maps, defect probability heatmaps and binary segmentation results in real time, and supports users to adjust thresholds, display historical frame sequences, and switch between real-time and offline modes.

[0044] Example 2:

[0045] This embodiment describes a detection method using the real-time and offline weld defect detection system described in Embodiment 1, including the following steps:

[0046] S1: Reference Data Acquisition

[0047] Under defect-free weld or reference specimen conditions, the multi-channel excitation and acquisition unit sequentially excites each coil, acquires the voltage vector of all selected coil pairs, and calculates and stores the reference voltage V. ref .

[0048] S2: Data acquisition and differential analysis of the weld under test

[0049] A planar electromagnetic tomography sensor array is arranged above the weld to be measured, and the excitation and acquisition process in step S1 is repeated to obtain the measured voltage vector V. sample Calculate the differential signal ΔV = V sample -V ref ;

[0050] S3: Real-time LBP imaging based on a simplified sensitivity matrix

[0051] Using a simplified sensitivity matrix S containing only adjacent and next-nearest coil pairs, the two-dimensional perturbation image σ = S is calculated via linear backprojection. T ΔV is calculated and mapped to a predefined grid to obtain a real-time weld cross-section image;

[0052] S4: Image Reconstruction Preprocessing

[0053] The real-time reconstructed image is normalized, interpolated and scaled to a specified size, and lightweight morphological operations and denoising are performed as input to the segmentation network.

[0054] S5: Prototype-guided segmentation network inference

[0055] The preprocessed image is input into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores the spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Then, it is fused with the decoder features to output the defect probability map P.

[0056] S6: Post-segmentation processing and defect assessment

[0057] The defect probability map is binarized using an adjustable threshold τ to obtain a defect region mask. Isolated noise is removed and small holes are filled through morphological opening and closing operations. The area, length, position and orientation of the defect are calculated to achieve automatic identification and visualization of weld defects.

[0058] S7: Offline high-precision reconstruction (optional)

[0059] For the stored detection data, the Landweber iterative algorithm combined with the full sensitivity matrix is ​​used to perform offline high-precision reconstruction as needed, so as to obtain weld imaging results with higher signal-to-noise ratio and spatial resolution. Then, the data is input into the prototype guided segmentation network unit for offline high-precision segmentation and analysis.

[0060] The specific operation and application steps are as follows:

[0061] 1. Hardware Configuration

[0062] The planar EMT sensor uses 16 planar coils arranged in two staggered rows, with a center distance of 5mm between adjacent coils, an inner and outer radius of 1mm and 2mm respectively, and 500 coil turns.

[0063] Each coil has a ferrite core inserted inside to improve magnetic field concentration and signal-to-noise ratio;

[0064] Below the coil array is a 4mm thick metal welding plate. The weld area has an uneven surface and contains artificial defects with depths of 0.5mm and 1mm.

[0065] 2. Sensitivity Matrix Construction

[0066] A model of the coil array and weld plate was established using 3D finite element software. In each simulation, one coil was selected as the excitation coil and the rest were selected as the receiving coils.

[0067] Define a 44×10 voxel grid below the weld and record the magnetic field components of each coil pair at each grid point;

[0068] according to Calculate the sensitivity values ​​to obtain the complete sensitivity matrix;

[0069] A simplified sensitivity matrix is ​​constructed by selecting coil pairs with a distance of 1 times and 2 times the coil center distance from the complete matrix, which is used for real-time reconstruction.

[0070] 3. Real-time detection process

[0071] First, the reference voltage V was acquired under defect-free reference plate conditions. ref ;

[0072] In actual testing, the sensor is placed stably on the weld surface, and 16 coils are sequentially excited using a multi-channel instrument. The complex voltage data of the corresponding coil pairs are collected to obtain V. sample ;

[0073] The data processing unit calculates the difference ΔV = V sample -V ref LBP reconstruction was performed using a simplified sensitivity matrix to obtain a 44×10 two-dimensional perturbation image;

[0074] The perturbation image is interpolated to 512×512, normalized, and then input into the ProSiF-UNet segmentation network to output a defect probability map, thereby obtaining a binary defect mask.

[0075] 4. ProSiF-UNet Training and Inference

[0076] The network encoder uses a pre-trained residual network, and the decoder uses a standard UNet structure;

[0077] During the training phase, manually labeled weld defect masks were used as supervision signals, and Tversky+Focal composite loss was used for training for 200 epochs.

[0078] During training, data augmentation techniques such as multi-scale (0.75, 1.0, 1.25) scaling, flipping, and slight elastic deformation are used to improve the model's generalization ability.

[0079] The prototype branch constructs a class prototype by averaging the features from the training batch and updates the running prototype using the EMA strategy.

[0080] During the inference stage, a defect probability map can be obtained by inputting a single-frame reconstructed image, and the defect region can be obtained through thresholding and morphological processing.

[0081] Performance

[0082] Through simulation and experimental comparison, LBP imaging based on simplified sensitivity matrix is ​​highly consistent with the full matrix Landweber reconstruction results in terms of defect localization and contour preservation, but the single-frame reconstruction time is significantly shortened, meeting the requirements of real-time detection.

[0083] Compared to UNet, DeepLabv3+, and PSPNet, ProSiF-UNet achieves higher IoU and DICE, and significantly reduces the false negative rate of thin crack defects. Its single-frame inference speed can reach over 200 frames / second, and it can be overlaid on the reconstructed image in real time.

[0084] Verification has shown that the present invention achieves the following beneficial effects through the above-described solution:

[0085] (1) Highly robust testing adaptable to uneven weld surfaces

[0086] By constructing a three-dimensional electromagnetic field sensitivity matrix using a planar EMT sensor array, compared to traditional array eddy currents that only utilize adjacent coil pairs, this invention uses a combination of multiple coil pairs and optimizes through physical modeling and simulation, enabling the system to maintain high-sensitivity imaging capability for defects even under interference from weld toe undulations, irregular weld waves, and grinding marks.

[0087] (2) Simplify the sensitivity matrix to balance real-time performance and imaging quality

[0088] By selecting adjacent and second-nearer coil pairs to construct a simplified sensitivity matrix, distant coil pairs with limited contribution to local defects are effectively eliminated, significantly reducing the amount of data and computation. Simulation and experimental results show that the imaging quality is close to that of the full coil pair scheme under the d=2 (second-nearer coil pair) configuration, achieving a balance between real-time performance and imaging accuracy.

[0089] (3) Prototype-guided segmentation network significantly reduces false negative rate

[0090] By introducing class prototypes and cosine similarity maps into the segmentation network and fusing them with decoder features, the network can enhance the expression of hard-to-detect regions by utilizing "defect appearance priors" in scenarios with small samples, low contrast, and thin cracks. Compared with traditional baseline networks such as UNet and DeepLabv3+, it achieves higher IoU and Dice, while significantly reducing the false negative (missed detection) rate.

[0091] (4) Real-time / offline integrated weld inspection frame

[0092] It supports both online rapid monitoring based on LBP and offline high-precision analysis based on Landweber iteration, making it suitable for two application scenarios: real-time screening on production lines and fine evaluation in laboratories. It has high engineering promotion value.

[0093] (5) The system has a clear structure and can be easily implemented in both hardware and software.

[0094] Sensor arrays, multi-channel acquisition, data processing, imaging, and segmentation networks can be implemented on general-purpose GPU platforms, facilitating integration with industrial automation systems and robotic inspection platforms, and enabling easy engineering and large-scale applications.

[0095] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A real-time and offline detection system for weld defects, characterized in that: The system includes a planar electromagnetic tomography sensor array, a multi-channel excitation and acquisition unit for exciting and acquiring data from the planar electromagnetic tomography sensor array, and a data processing and imaging unit connected to the multi-channel excitation and acquisition unit for performing reference signal acquisition, differential voltage calculation, sensitivity matrix construction, real-time LBP reconstruction, offline LW reconstruction, and preprocessing of the reconstructed image. It also includes a prototype-guided segmentation network unit with an encoder, decoder, prototype branch module, and similarity fusion module, and an information processing platform communicatively connected to the prototype-guided segmentation network unit. The data processing and imaging unit communicates with the prototype-guided segmentation network unit and can input the preprocessed image into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Subsequently, the features are fused with the decoder features to output a defect probability map P, which is then processed by the information processing platform to achieve automatic weld defect determination.

2. The real-time and offline detection system for weld defects according to claim 1, characterized in that: The planar electromagnetic tomography sensor array consists of multiple planar electromagnetic coils arranged in two staggered rows and a magnetic core, with each coil wound on the magnetic core. The planar electromagnetic tomography sensor is installed above the weld to be measured, and the axis of the planar electromagnetic coil is perpendicular to the weld surface. The multi-channel excitation and acquisition unit sequentially selects each coil as the excitation coil and uses the remaining coils as receiving coils to perform measurements. It can also transmit the measurement data to the data processing and imaging unit via UDP or bus.

3. The real-time and offline detection system for weld defects according to claim 1 or 2, characterized in that: The data processing and imaging unit includes a reference voltage vector V for acquiring the reference voltage vector V under defect-free weld or plate conditions. ref The reference signal acquisition module is used to acquire the voltage vector V under the measured state of the weld. sample And calculate the differential signal ΔV = V sample -V ref The differential voltage calculation module, based on finite element simulation, constructs a sensitivity matrix S containing only adjacent and next-nearest coil pairs, with corresponding coil pair spacings of 1 and 2 times the coil center spacing. The simplified sensitivity matrix construction module for long-distance coil pairs is discarded. A linear back-projection algorithm σ = S is employed. T A real-time LBP reconstruction module that maps differential voltage to a two-dimensional weld cross-section conductivity / permeability perturbation image, and a Landweber iterative algorithm σ for stored detection data. (k+1) =σ (k) +αS T (ΔV-Sσ (k) An offline LW reconstruction module that performs multiple iterations to obtain reconstructed images with higher spatial resolution and contrast.

4. The real-time and offline detection system for weld defects according to claim 3, characterized in that: The encoder extracts multi-scale deep features based on a residual network; the decoder uses UNet-style skip connections to upsample multi-scale features step by step and restore spatial resolution; the prototype branch module globally aggregates the high-level feature maps of the encoder to obtain class prototype vectors for the defect class and the background class, calculates the similarity between each pixel location feature and the class prototype using cosine similarity, generates a prototype similarity map, and the prototype uses an exponential moving average (EMA) update strategy to continuously update the running prototype during training and inference; the similarity fusion module concatenates the prototype similarity map and the decoder output features in the channel dimension, and fuses them through 1×1 convolution to generate the final defect probability map.

5. The real-time and offline detection system for weld defects according to claim 1, 2 or 4, characterized in that: The information processing platform is also connected to a visualization and interaction unit.

6. The real-time and offline detection system for weld defects according to claim 5, characterized in that: The visualization and interaction unit can provide a multi-window interface to display I / Q waveforms, real-time LBP reconstruction maps, defect probability heatmaps and binary segmentation results in real time, and supports user adjustment of thresholds, display of historical frame sequences, and switching between real-time and offline modes.

7. A detection method using the real-time and offline detection system for weld defects as described in any one of claims 1-6, characterized in that: Includes the following steps: S1: Reference Data Acquisition Under defect-free weld or reference specimen conditions, the multi-channel excitation and acquisition unit sequentially excites each coil, acquires the voltage vector of all selected coil pairs, and calculates and stores the reference voltage V. ref . S2: Data acquisition and differential analysis of the weld under test A planar electromagnetic tomography sensor array is arranged above the weld to be measured, and the excitation and acquisition process in step S1 is repeated to obtain the measured voltage vector V. sample Calculate the differential signal ΔV = V sample -V ref ; S3: Real-time LBP imaging based on a simplified sensitivity matrix Using a simplified sensitivity matrix S containing only adjacent and next-nearest coil pairs, the two-dimensional perturbation image σ = S is calculated via linear backprojection. T ΔV is calculated and mapped to a predefined grid to obtain a real-time weld cross-section image; S4: Image Reconstruction Preprocessing The real-time reconstructed image is normalized, interpolated and scaled to a specified size, and lightweight morphological operations and denoising are performed as input to the segmentation network. S5: Prototype-guided segmentation network inference The preprocessed image is input into the prototype-guided segmentation network unit. The encoder extracts features, the decoder restores the spatial resolution step by step, and the prototype similarity map is calculated using the prototype branch. Then, it is fused with the decoder features to output the defect probability map P. S6: Post-segmentation processing and defect assessment The defect probability map is binarized using an adjustable threshold τ to obtain a defect region mask. Isolated noise is removed and small holes are filled through morphological opening and closing operations. The area, length, position and orientation of the defect are calculated to achieve automatic identification and visualization of weld defects.

8. The real-time and offline detection method for weld defects according to claim 7, characterized in that: After step S6, an offline high-precision reconstruction step is also included. In this step, for the stored detection data, the Landweber iterative algorithm is used in combination with the full sensitivity matrix to perform offline high-precision reconstruction as needed, so as to obtain weld imaging results with higher signal-to-noise ratio and spatial resolution. Then, the results are input into the prototype guided segmentation network unit for offline high-precision segmentation and analysis.

Images