A hybrid neural network model based modelling methodology for the rubber bushing

Dai, Liangcheng; Chi, Maoru; Xu, Chuanbo; Gao, Hongxing; Sun, Jianfeng; Wu, Xingwen; Liang, Shulin

doi:10.1080/00423114.2021.1933090

Cited by 8 publications

(1 citation statement)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They calculated the radial stiffness of rubber bushes under different cross-sectional shapes and axial pre-compression conditions through simulation. Similarly, Dai et al [9] examined the dynamic performance of rubber bushes under different frequencies, amplitudes, and temperature excitations, and they developed a hybrid neural network model trained on performance data to accurately represent the dynamic performance of rubber bushes under diverse conditions. In addition, Dai et al [10] integrated a physical parameter model with a neural network model, considering structural parameters of the damper in the former and describing hydraulic damper characteristics in the latter.…”

Section: Introductionmentioning

confidence: 99%

Performance Prediction of the Elastic Support Structure of a Wind Turbine Based on Multi-Task Learning

Zhu,

Qi,

et al. 2024

Machines

View full text Add to dashboard Cite

The effectiveness of a wind turbine elastic support in reducing vibrations significantly impacts the unit’s lifespan. During the structural design process, it is necessary to consider the influence of structural design parameters on multiple performance indicators. While neural networks can fit the relationships between design parameters on multiple performance indicators, traditional modeling methods often isolate multiple tasks, hindering the learning on correlations between tasks and reducing efficiency. Moreover, acquiring training data through physical experiments is expensive and yields limited data, insufficient for effective model training. To address these challenges, this research introduces a data generation method using a digital twin model, simulating physical conditions to generate data at a lower cost. Building on this, a Multi-gate Mixture-of-Experts multi-task prediction model with Long Short-Term Memory (MMoE-LSTM) module is developed. LSTM enhances the model’s ability to extract nonlinear features from data, improving learning. Additionally, a dynamic weighting strategy, based on coefficient of variation weighting and ridge regression, is employed to automate loss weight adjustments and address imbalances in multi-task learning. The proposed model, validated on datasets created using the digital twin model, achieved over 95% predictive accuracy for multiple tasks, demonstrating that this method is effective.

show abstract

Section: Introductionmentioning

confidence: 99%