Augmenting gas turbine health monitoring: invariant feature extraction from unlabeled blade temperature profiles via GAF and knowledge distillation

Abstract Gas turbine health monitoring fundamentally relies on accurate measurement and interpretation of physical signals. However, the detection and diagnosis of abnormal states of gas turbines predominantly rely on indirect methods using data such as vibration and exhaust temperature currently. To enrich the measurement modalities for existing gas turbine monitoring systems, this paper presents a novel method to construct individual blade-level invariant feature by leveraging an augmented deep learning framework. Firstly, the new turbine blades infrared fiber optic temperature measurement system has been developed. Secondly, to compensate for the absence of labeled fault data, an identity recognition proxy task is designed, enabling the model to learn inherent feature distributions from healthy turbines in a self-supervised manner. Thirdly, to mitigate the high sensitivity of raw signals to operating conditions, a multi-channel gramian angular field encoding with cyclic phase augmentation is introduced, which enhances spatial perception and eliminates condition-induced variations. Finally, to extract invariant thermal features from healthy blades, a knowledge distillation architecture is employed, ensuring both accurate feature representation and lightweight deployability for subsequent health monitoring. Experimental results on multi-condition and multi-cycle datasets show that the drop in classification accuracy serves as a potential indicator for health degradation. This approach provides a practical measurement-based complementary method to existing gas turbine monitoring systems, facilitating the establishment of health baselines without the need for prior fault labels.