A tri-axial MFL GADF representation and axis-weighted Swin transformer approach for pipeline defect identification

Magnetic flux leakage (MFL) measurements contain coupled directional responses that are affected by defect morphology, magnetisation conditions, sensor position, and measurement noise. Consequently, conventional handcrafted-feature methods and single-axis deep-learning models may not fully exploit the complementary information contained in tri-axial MFL signals. This study proposes AW-SwinT-Cal, a pipeline-defect classification framework combining per-axis Gramian Angular Difference Field (GADF) encoding, a shared Swin Transformer backbone, global axis-weighted feature fusion, test-time augmentation (TTA), and validation-guided class-wise logit adjustment. Each synchronised MFL component is independently encoded as a two-dimensional GADF image. A shared Swin-T backbone then extracts axis-specific features, which are aggregated using three softmax-normalised global coefficients. TTA and class-wise additive logit adjustment are applied during inference to reduce prediction variability and modify class decision boundaries. Experiments were conducted on a six-class simulation-derived dataset containing 7,307 complete tri-axial samples and 21,921 GADF images. An independent data-integrity audit confirmed that no source CSV file or raw signal window was shared across the training, validation, and test subsets. AW-SwinT-Cal achieved 95.31% Accuracy, 95.05% Macro-F1, and 95.29% Weighted-F1 on the fixed test set. In direct raw-signal comparisons, an RBF-SVM using 69 handcrafted descriptors and a four-block 1D-CNN achieved 73.52% and 78.07% Accuracy, respectively. Temperature scaling improved probability calibration without changing the predicted labels, whereas the proposed class-wise logit adjustment increased Accuracy by 0.55 percentage points and Macro-F1 by 0.58 percentage points relative to the unadjusted TTA output. Source-cluster bootstrap analysis yielded 95% confidence intervals of [0.00, 1.27] and [0.00, 1.40] percentage points for these improvements, respectively. However, the exact McNemar test was not significant (p = 0.125). Across three random seeds, the method achieved an Accuracy of 0.9503 ± 0.0048 and a Macro-F1 of 0.9477 ± 0.0049. These results support the effectiveness of the complete tri-axial GADF framework under the evaluated simulation conditions, while indicating that the modest numerical gain from class-wise logit adjustment requires confirmation using larger independent datasets.