A Sparse Batch-Wise Deformable Attention Module for Damage Detection under Imbalanced Data

Jitong Ma

doi:10.1109/jiot.2026.3726692

A Sparse Batch-Wise Deformable Attention Module for Damage Detection under Imbalanced Data

2026-08-24

0

PRE

AI

DOI：10.1109/jiot.2026.3726692

Original

Original request for help

Save

Abstract

En 中文

Ultrasonic guided wave (UGW) sensing is widely used for structural health monitoring of Internet of Things-enabled composite structures. However, data-driven UGW damage detection remains challenging because damage samples are scarce and the collected data are often severely imbalanced. To address this issue, this paper proposes a sparse batch-wise deformable attention (SBDA) module for imbalanced damage detection. SBDA combines response-guided Top-k sparse channel selection with batch-wise deformable interaction, thereby reducing redundant channel participation and adaptively modeling inter-sample relationships within mini-batches. As a lightweight plug-and-play module, SBDA can be integrated into CNN- and Transformer-based monitoring models. Experimental results on imbalanced composite datasets demonstrate that SBDA improves damage detection performance with low computational overhead.

Keywords:

Internet of Things (IoT)

Imbalanced data

Attention

Damage detection

Ultrasonic guided waves

A Sparse Batch-Wise Deformable Attention Module for Damage Detection under Imbalanced Data

Abstract

Journal

Organization

Cited Papers