Fault diagnosis of aircraft starting circuit using Koopman operator and CWT

Abstract The aircraft starting circuit is a crucial system for ensuring safe takeoff, and its fault diagnosis plays an essential role in maintaining flight safety. Traditional diagnostic methods, which rely on predefined rules and thresholds, are often inadequate for handling complex nonlinear faults. Recent advancements in deep learning-based diagnostic methods have shown promise by automatically extracting features from data. However, these methods are still prone to noise interference, which complicates the differentiation of fault patterns. This paper proposes an aircraft starting circuit fault diagnosis method that integrates the Koopman operator, continuous wavelet transform (CWT), and a convolutional gated recurrent unit (Conv-GRU) network. Initially, a residual network based on the Koopman operator is employed to dynamically model the main characteristics of the signals, effectively reducing noise and enhancing the differences between signals of different faults. Next, CWT is applied to the extracted dynamic features to perform multiscale time–frequency analysis, generating time–frequency feature maps. Finally, a hybrid Conv-GRU model, which combines Convolutional neural networks and GRU, is designed to capture both spatial and temporal dependencies in the time–frequency features. Experimental results on a four-stage aircraft starting circuit fault dataset demonstrate that the proposed method significantly enhances fault diagnosis performance with a diagnostic accuracy of 99.81%.