Title: NL-MambaXCT: Leveraging Self-Supervised Nested-Learning Mamba for Defect Classification in Nomex Honeycomb X-ray CT

Abstract:

While X-ray computed tomography (XCT) is a standard tool for the non-destructive evaluation of Nomex honeycomb structures within the aerospace sector, current industrial practices remain dependent on manual analysis and supervised algorithms that require substantial labeled datasets. To address these limitations, this study presents NL-MambaXCT, a novel framework grounded in the Mamba architecture. This system integrates self-supervised masked image modeling with a Nested Learning (NL) approach to enable automated defect classification from production XCT slices with high label efficiency.

The model’s architecture features a four-stage 2D encoder. The initial stages utilize RegNet convolutional blocks, whereas the deeper stages employ Mamba-based sequence mixing combined with attention mechanisms. Pretraining was conducted via masked image modeling on a dataset comprising 19,961 unlabeled industrial XCT slices. Subsequently, the model was fine-tuned using 2,000 relabeled Nomex XCT slices, organized according to production sequence.

The Nested Learning mechanism is realized through two-timescale parameter dynamics. This involves selected projections that preserve slow exponential-moving-average traces in parallel with fast weights, complemented by a deep-momentum optimizer that establishes an additional slow parameter-update trajectory.

Evaluation on a held-out test set demonstrates that the MIM-pretrained NL-MambaXCT model reaches an accuracy of 96.91% and a macro F1 score of 96.8%. These metrics surpass those of CNN, attention-based, and single-timescale Mamba baseline models by a margin of 3.11 to 10.31 percentage points in accuracy. The findings indicate that merging masked self-supervision with the fast/slow learning dynamics characteristic of NL represents a highly effective strategy for ensuring robust defect classification in Nomex honeycomb XCT inspections.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC