Resformer: An end-to-end framework for fault diagnosis of governor valve actuator in the coupled scenario of data scarcity and high noise

Abstract

As the actuator of the turbine speed control system, the performance and response characteristics of the speed control valve actuator directly affect the operational economy, maneuverability, and reliability of the turbine unit. When faults occur in scenarios where data scarcity is coupled with high noise levels, existing deep neural network models are limited by their inability to extract key discriminative features from noisy signals and by the lack of sufficient training information. This limitation hinders the development and application of highly reliable fault diagnosis systems. We propose a novel fault diagnosis framework, Resformer, which is designed to address the challenges posed by data scarcity and high noise coupling, as well as the highly coupled and complex fault modes in electro-hydraulic systems. The Resformer framework offers a highly interpretable feature selection and fusion strategy to identify key features. It also integrates the Local Binary Pattern algorithm to extract local features from grayscale images of multi-sensor data, significantly enhancing the representativeness and noise resistance of the dataset. Moreover, to strengthen the Resformer’s multi-scale feature extraction capability and noise robustness, a multi-kernel dilated convolutional residual network architecture is introduced, enabling the discovery of critical discriminative features under conditions of data scarcity and high noise coupling. The proposed efficient multi-scale self-attention mechanism effectively extracts important features at different scales, further improving the performance of Resformer. Experiments conducted on the GVA testbed have validated the effectiveness and robustness of Resformer.