基于预训练模型的中心分级燃烧室燃烧振荡预报方法

Combustion oscillation prediction method in centrally-staged combustors based on pre-training model

关键词:  燃气轮机  燃烧室  燃烧振荡  预训练模型  迁移学习  主成分分析  分层聚类

Key words:  Gas turbine  Combustor  Combustion instability  Pre-training model  Transfer learning  Principal component analysis  Hierarchical clustering

摘要:

为促进实现燃气轮机燃烧室中的燃烧振荡预报，提出一种结合预训练和迁移学习的研究思路。在预训练阶段，开展短火焰筒和长火焰筒下两类火焰图像的对比学习以完成编码器的自监督预训练。在迁移阶段，除了对特征编码构建线性分类器的直接迁移，本文还提出将工况参数作为先验条件的贝叶斯迁移学习。结果表明，在两种迁移学习方式下预训练模型相比传统监督学习模型具有4.6%左右的性能提升。同时基于贝叶斯推断的迁移学习相比直接迁移鲁棒性更好。通过主成分分析和分层聚类，验证预训练模型能够提取火焰图像更为通用的热声特征。

Abstract:

To promote the prediction of combustion oscillation in gas turbine combustors， a research approach combining pre-training and transfer learning is proposed. In the pre-training phase， contrastive learning using two types of flame images under short and long flame tubes is carried out to complete the self-supervised pre-training of the encoder. In the transfer phase， in addition to the direct transfer of constructing a linear classifier for feature encodings， Bayesian transfer learning with operating conditions parameters as prior conditions is proposed in this paper. The results show that the pre-trained model has a performance improvement of about 4.6% compared to traditional supervised learning models under two transfer learning methods. Meanwhile， transfer learning based on Bayesian inference exhibits better robustness compared to direct transfer. Through principal component analysis and hierarchical clustering， it is verified that the pre-trained model extracts more general thermoacoustic features from flame images.