Biologically inspired information integration pooling module of spiking neural networks for rolling bearing fault diagnosis
•The research establishes a mathematical model for the application of rolling bearing fault diagnosis, with theoretical analysis of ANN-to-SNN.•We develop IIPooling, a novel and effective module for information integration of ANN-to-SNN.•IIPooling includes the promotion, lateral inhibition, and long...
Uloženo v:
| Vydáno v: | Expert systems with applications Ročník 286; s. 128032 |
|---|---|
| Hlavní autoři: | , , , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
Elsevier Ltd
15.08.2025
|
| Témata: | |
| ISSN: | 0957-4174 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | •The research establishes a mathematical model for the application of rolling bearing fault diagnosis, with theoretical analysis of ANN-to-SNN.•We develop IIPooling, a novel and effective module for information integration of ANN-to-SNN.•IIPooling includes the promotion, lateral inhibition, and longitudinal inhibition processes.•IIPooling incorporates negative feedback mechanism through inductive bias, such as surrogate gradient and spatiotemporal information.•IIPooling adjusts both promotion and lateral inhibition parameters adaptively.
Rolling bearings constitute critical components in mechanical systems, and their effective fault diagnosis plays a vital role in ensuring operational safety and reliability. Despite substantial advancements in diagnostic methodologies, achieving high-precision, low-latency fault detection in large-scale rolling bearing datasets remains a persistent challenge. Artificial neural networks to spiking neural networks (ANN-to-SNN) conversion algorithms offer promising solutions for reducing computational costs and hardware adaptation barriers compared to conventional approaches. However, inherent information degradation during conversion processes has been demonstrated to significantly undermine model performance in terms of accuracy, stability, and interpretability. To overcome these limitations, we propose IIPooling, a novel Information-Integration biological neural mechanism-inspired Pooling module designed to enhance global structural feature extraction. The proposed framework incorporates three core neurobiological information processing principles: (1) promotion, (2) lateral inhibition, and (3) longitudinal inhibition, effectively simulating biological neuronal spiking patterns while optimally utilizing ANN activation characteristics. Additionally, we develop two adaptive learning mechanisms — Surrogate Gradient Adaptive Learning and Spiking Spatiotemporal Adaptive Learning — to implement negative feedback mechanism and adaptive parameter adjustment, thereby improving cross-conditional generalization capabilities. Comprehensive evaluations conducted on five bearing datasets (CWRU: DE/FE, PU, and Lab-collected: 6025/N205EW) demonstrate superior performance across four metrics (ACC, AUC, AUPRC, Mean-ACC). Our method achieves state-of-the-art results, notably attaining ACC values of 0.959 (CWRU-DE) and 0.861 (Lab-N205EW) in IIPooling-ST configurations, confirming its effectiveness for real-time, high-accuracy fault detection in complex operational environments. |
|---|---|
| ISSN: | 0957-4174 |
| DOI: | 10.1016/j.eswa.2025.128032 |