Combustion Instability Diagnosis of OH* Chemiluminescence Based on a Self-Supervised Learning Method

Abstract

Combustion instability has many detrimental effects on the dynamics and structural reliability of areo-engines and gas turbines. Hence, combustion diagnosis plays an important role on engine and turbine health monitoring and prognostics. In this paper, premixed combustion experiments were designed to obtain flame stability and instability data by varying different equivalence ratios and verified the ability of four backbones extracting features on combustion instability in a supervised-learning way. In order to address the problem that the performance of the model is directly related to the labeling of the data, especially for the case where the label classification criterion is ambiguous, a method was proposed based on the SimCLR, a self-supervised method, involving pre-training the model on a large unlabeled dataset and fine-tuning on a smaller labeled dataset for the downstream task. The experimental results show that the method can achieve remarkable enhancement in the models' expressiveness, pushing their upper limit to unprecedented levels by leveraging the self-supervised learning, though fine-tuning only with a small number of labels.