A MOE-attention-based reinforcement learning for gearbox fault diagnosis with unbalanced data and computation complexity

•Uniform sampling minimizing bias in imbalanced datasets.•Optimal reinforcement learning removing redundant information from vision Transformer.•Effective mixture of experts in attention mechanisms. In fault diagnosis, imbalanced data poses a challenge to accuracy, whereas computational complexity e...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Measurement : journal of the International Measurement Confederation Ročník 254; s. 117917
Hlavní autori: Liu, Weimin, Liu, Jiaxin, Han, Bin, Zheng, Aiyun, Zheng, Zhi
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier Ltd 01.10.2025
Predmet:
Fault diagnosis
MOE-attention
Reinforcement learning
Vision transformer
LN
SwinT
Imba – DA
PLC
DL
CoAtNet
SVM
T2T – ViT
DeiT
CNN-Transformer
PPO
CNNs
FFN
Fault diagnosis
MLP
BiGRU
SMOTE
LBM
SNR
MDP
OWA
PiT
ViTs
MOE-attention
VTARL
CvT
DQN
MOE
cmor
LSTM
Reinforcement learning
CrossViT
CNN – Attention
Vision transformer
MobileViT
ReLU
RL
fc
TCN
DRL
ISSN:0263-2241
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:•Uniform sampling minimizing bias in imbalanced datasets.•Optimal reinforcement learning removing redundant information from vision Transformer.•Effective mixture of experts in attention mechanisms. In fault diagnosis, imbalanced data poses a challenge to accuracy, whereas computational complexity exacerbates inefficiencies in real-world deployment. To address those, a vision transformer assisted by reinforcement learning (VTARL) was proposed for gearbox fault diagnosis. As a start, four kinds of datasets with noised data were artificially created. Subsequently, a strategy based on uniform sampling was raised with a pilot model to balance positive and negative samples. Then, an agent based on reinforcement learning (RL) was developed to optimize model through continuous interaction, instead of traditional form between agent and environment. Furthermore, a sparse attention was derived from mixture of experts (MOE) to reduce the complexity of multi-head attention by selectively activating experts. Finally, VTARL was vivified through three parallel channels and the performance comparisons among them. The results indicate that both the positive and negative data is well balanced with pilot model. Then, the length of feature vector sequence is reduced from over 500 down to single digit by agent. Additionally, MOE-attention compresses the computational complexity from O((seq_length)2 * hidden_size) to O(seq_length * expert output * hidden_size). Accuracy rates for diagnostic tasks under D2 dataset were recorded at 99.18%, 100%, 99.41%, and 98.52%, respectively, which consequently underscore the efficacy of the VTARL approach to imbalance data and computation complexity.
ISSN:0263-2241
DOI:10.1016/j.measurement.2025.117917