⚡ electrical engineering

Dual-Stream Network for Width Inversion of Composite Defects and Experimental Validation

This paper proposes a dual-stream multiscale attention network (MS-BiGRU-Attn) that processes axial and normal magnetic flux leakage signals independently to accurately invert the width of composite pipeline defects under nonlinear magnetic interference, achieving superior performance on both out-of-distribution simulations and experimental validation compared to baseline models.

Original authors: Yong Hon, Daohao Pan, Jianghui Cheng, Changhui Li, Huajian Liu

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yong Hon, Daohao Pan, Jianghui Cheng, Changhui Li, Huajian Liu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Oil and gas pipelines are the hidden arteries of modern energy security, carrying vast quantities of fuel across continents. Keeping these pipes safe requires knowing exactly what is wrong with them when they develop flaws. Over time, the metal walls of these pipes can corrode, forming pits and grooves that weaken the structure. To find these hidden dangers without cutting the pipe open, inspectors use a technique called magnetic flux leakage. Imagine running a strong magnet along the outside of a steel pipe; the magnetic field flows smoothly through the healthy metal. But when the magnet passes over a defect, some of that magnetic field leaks out into the air, creating a unique signal that sensors can detect. By analyzing these signals, engineers can estimate the size and shape of the damage, which is critical for deciding if a pipe needs immediate repair or can continue operating safely.

The challenge becomes much harder when the damage is not a single, neat pit, but a complex cluster of defects nested inside one another. In the real world, corrosion often creates a shallow, wide area of damage with a deep, narrow hole sitting right in the middle. This "composite" defect creates a chaotic magnetic environment where the signals from the different parts of the damage mix together and distort each other. Traditional methods for measuring the width of these defects often fail in this situation because the messy signals confuse the measurement tools. A team of researchers at Shanghai Dianji University and the Wuhu Special Equipment Inspection Institute has developed a new approach to solve this specific problem, using a type of artificial intelligence designed to untangle these mixed signals.

The researchers focused on two specific types of magnetic signals that sensors pick up: one that runs along the length of the pipe and another that points straight out from the surface. In a perfect, simple defect, these two signals behave predictably. However, when defects are nested, the signal pointing outward becomes extremely noisy and erratic, and this noise tends to corrupt the signal running along the length, making it impossible to tell where the defect actually begins and ends. Most existing computer models try to feed both of these messy signals into the system at the same time, hoping the computer can figure out the pattern. The researchers found that this approach causes the computer to get confused, mixing up the noisy parts with the useful parts, leading to wildly inaccurate measurements.

To fix this, the team built a new kind of neural network, which is a computer program modeled after the human brain, that treats the two signals completely separately at first. Instead of forcing the signals to mix immediately, the system sends the length-wise signal down one path and the outward-pointing signal down a different, independent path. Each path has its own specialized tools to analyze the data, looking for specific patterns like sharp edges or gradual slopes. Only after each path has done its own deep analysis does the system bring the two results together to make a final decision. This design acts like a filter, preventing the chaotic noise from one signal from ruining the clear information in the other. The researchers trained this system using thousands of simulated examples of simple, single defects, never showing it any of the complex, nested defects during the learning phase.

When tested on these complex, nested defects that the system had never seen before, the new method proved remarkably effective. While older models that mixed the signals together failed completely, producing errors that were too large to be useful, the new dual-path system accurately predicted the width of the damage in most cases. In a series of tests involving 150 different complex defect scenarios, the new model correctly identified the width with a high degree of reliability, whereas the older methods failed to find a consistent pattern. The researchers also built physical steel plates with real, nested defects and ran their sensors over them to see if the computer predictions held up in the real world. The results were consistent with the simulations: the new system successfully tracked the overall size of the damage, while the older systems struggled significantly, often guessing the wrong size by several millimeters.

However, the study also revealed a hard limit to what can be achieved with this technology. While the system became very good at measuring the width of the defects, it could not accurately determine their depth. No matter how the computer was adjusted, it consistently failed to guess how deep the holes were in the nested defects. The researchers concluded that this is not a flaw in the computer program, but a fundamental physical limitation. The magnetic signals simply do not carry enough clear information about the depth when the defects are nested in this specific way. This finding is important because it tells engineers exactly what they can and cannot expect from this technology, preventing them from relying on it for measurements that are physically impossible to make with current sensors.

The success of this approach lies in its ability to respect the physical nature of the problem. By keeping the different types of magnetic data separate until they have been thoroughly analyzed, the system avoids the confusion that plagues other methods. This work suggests that for complex industrial problems, the best solution is not always to throw more data at a single computer model, but to design the model to handle different types of information in the most logical way possible. The researchers plan to continue this work by testing the system on full-scale pipeline inspection rigs and exploring ways to combine data from multiple sensors to overcome the depth measurement limit. For now, they have provided a robust tool that can reliably measure the width of dangerous, complex defects, offering a clearer view of the health of our critical energy infrastructure.

Technical Summary: Dual-Stream Network for Width Inversion of Composite Defects

Problem Statement
Magnetic Flux Leakage (MFL) testing is a standard nondestructive evaluation method for pipeline integrity. While quantitative inversion of isolated, regular defects is well-established, characterizing composite defects (nested, overlapping corrosion morphologies) remains a significant challenge. In these scenarios, nonlinear magnetic field interference causes signal distortion, where the leakage fields of adjacent defects superimpose and reconfigure topologically. Specifically, the high-frequency distortion in the normal magnetic component contaminates the width-associated features in the axial component. Conventional deep learning models typically employ an early-fusion strategy, rigidly concatenating multi-axis signals at the input layer. This approach fails under composite conditions, leading to cross-dimensional feature cross-talk, representation distortion, and severe inaccuracies in width inversion. Furthermore, the depth dimension of composite defects presents a physically ill-posed problem due to magnetic field diffusion and saturation, making precise depth quantification from one-dimensional signals inherently difficult.

Methodology
To address the issue of nonlinear interference and feature cross-talk, the authors propose a Dual-Stream Multiscale Attention Network (MS-BiGRU-Attn). The core architectural innovation is physical isolation:

Dual-Stream Architecture: Instead of early fusion, the network processes the axial ( $B_{ax}$ ) and normal ( $B_{no}$ ) signal components in two independent, parallel branches with unshared weights. This prevents high-frequency noise from the normal component from corrupting the width-boundary information in the axial component at the shallow layers.
Spatiotemporal Feature Extraction: Each stream utilizes a Dilated Multi-Scale CNN (MS-CNN) to capture micro-edge gradients and macroscopic leakage contours across different receptive fields. This is followed by Bidirectional Gated Recurrent Units (BiGRU) to model temporal dependencies and spatial sequences.
Attention Mechanism: A temporal attention mechanism is integrated into each stream to adaptively weight high-gradient regions (e.g., bimodal peaks or zero-crossings) while suppressing background noise and interference distortion.
Late Fusion: Features from the two isolated streams are fused only at deep layers (Stage 3) via concatenation, allowing interaction within a high-dimensional nonlinear semantic space rather than at the raw signal level.
Multi-Task Learning: The network performs joint optimization for defect morphology classification (Pitting, Grooving, General Corrosion) and 3D size regression (Length, Width, Depth).

The model was trained exclusively on single-defect data (2,000 samples of rectangular and cylindrical defects) generated via finite-element modeling. Its generalization was tested on an Out-of-Distribution (OOD) set comprising 150 nested composite defects and validated against real-world physical specimens.

Key Results

OOD Generalization (Simulation): On the OOD composite defect test set, the proposed MS-BiGRU-Attn achieved a width coefficient of determination ( $R_W^2$ ) of 0.883. This significantly outperformed the early-fusion baseline ( $R_W^2 = 0.178$ ) and the single-stream BiGRU ( $R_W^2 = 0.648$ ). The early-fusion model exhibited severe error long-tails and systematic underestimation, whereas the dual-stream model maintained compact error distributions.
Physical Validation (Experiment): In experiments using seven real-world composite defect specimens (C1–C7) fabricated on Q235 steel plates, the model demonstrated robust zero-shot generalization. The mean $R_W^2$ across five random seeds was 0.807, with predictions closely following the ideal diagonal. In contrast, the early-fusion model collapsed with a mean $R_W^2$ of 0.282, and the Vanilla-BiGRU showed instability with some seeds dropping below $R_W^2 = 0.66$ .
Depth Inversion Limitation: Across both simulated OOD tests and real physical experiments, all models failed to accurately invert the depth dimension ( $R_D^2 < 0$ ). The predicted depths clustered in a narrow range regardless of the true depth variation, confirming that depth quantification of composite defects from 1D MFL signals is a physically ill-posed problem that cannot be solved solely by data-driven methods.
Ablation Studies: Removing the multi-scale module or the attention mechanism significantly degraded performance ( $R_W^2$ dropped to 0.828 and 0.542, respectively). Early and mid-fusion strategies also failed, validating the necessity of deep late fusion.

Significance and Claims
The paper claims that the proposed dual-stream architecture provides a robust solution for width inversion under nonlinear magnetic field interference by enforcing physical isolation of signal components. The study demonstrates that sacrificing a marginal amount of accuracy on clean, single-defect data (a "structural regularization cost") yields substantial gains in robustness against complex, nested distortions. The authors assert that their approach effectively cuts off cross-dimensional noise propagation at the architectural source, preserving critical width-associated features.

Crucially, the paper modestly concludes that while width inversion is achievable through this method, the depth inversion of composite defects remains an insurmountable physical ill-posedness when relying solely on one-dimensional MFL signals. The authors emphasize that future work must move beyond single-stream signal inversion toward multi-dimensional sensor fusion and physics-guided strategies to address the depth limitation. The current experimental validation is limited by small specimen sizes and the reduction of array data to 1D sequences, suggesting that full-scale pipeline testing is required for further verification.

Technical Summary: Dual-Stream Network for Width Inversion of Composite Defects

More like this