LDS-former: A lightweight dual-stream transformer for real-time acoustic emission monitoring of crack evolution in offshore steel structures
Lu, Y.; Zhu, Z.; Li, Y.; Liu, H.; Hou, J.; Zhou, C.; Wahab, M.A. (2026). LDS-former: A lightweight dual-stream transformer for real-time acoustic emission monitoring of crack evolution in offshore steel structures. Advanced Engineering Informatics 74: 104635. https://dx.doi.org/10.1016/j.aei.2026.104635
In: Advanced Engineering Informatics. Elsevier: Oxford. ISSN 1474-0346; e-ISSN 1873-5320, more
|  |
| Author keywords |
Structure health monitoring;Acoustic emission;Fatigue crack detection;Lightweight transformer |
| Authors | | Top |
- Lu, Y.
- Zhu, Z.
- Li, Y.
- Liu, H.
|
- Hou, J.
- Zhou, C.
- Wahab, M.A., more
|
|
| Abstract |
Fatigue damage under cyclic loading remains a critical failure driver in offshore energy infrastructures, where dynamic stresses accelerate crack initiation and propagation, necessitating real-time diagnostic technologies to ensure service safety while limited computational capacity at edge nodes constrains monitoring system deployment. Acoustic emission (AE) technology achieves damage monitoring by capturing stress wave release during material fracture processes, yet AE signals exhibit transient-evolutionary duality where high-frequency bursts encode instantaneous fracture events while low-frequency trends reflect gradual damage accumulation, demanding intelligent diagnostic models to simultaneously capture heterogeneous temporal patterns while existing deep learning methods struggle to jointly resolve this dual characteristic and their computational demands prohibit edge deployment. To address this challenge, this study introduces a mechanism-aligned lightweight dual-stream transformer (LDS-Former) that matches the physical duality of AE signals through dualstream architecture, where the local stream preserves transient waveform details while the global stream models long-range dependencies, with a gated fusion mechanism adaptively weighting contributions to optimize feature representation. LDS-Former achieves 95.05% accuracy with only 0.199 M parameters and 5.215 M FLOPs, accomplishing 97.2% parameter reduction relative to deep architectures and 6.2% accuracy improvement over state-of-the-art lightweight methods. LDS-Former retains 86.24% accuracy with only 20% training data and maintains over 84% accuracy under extreme noise conditions. Edge deployment validation shows that LDS-Former achieves 2.086 ms inference latency on the STM32H7B3I-DK microcontroller with flash and RAM occupation of only 845.0 KB and 100.3 KB respectively. These findings confirm that LDS-Former, through synergistic integration of signal physical mechanisms and deep learning, provides an intelligent solution combining diagnostic precision with computational efficiency for offshore structural health monitoring, advancing practical deployment of edge intelligence in industrial safety monitoring. |
|
