Fault diagnosis technology for the energy interconnection system of the large commercial aircraft C919

Due to the short operational period of the C919 aircraft, relatively limited operational data are available, and data collection in the event of faults is even more difficult. In addition, large aircraft systems have a complex structure, and the research on fault mechanisms remains undeveloped. Moreover, aircraft operation tasks and environmental conditions change frequently, and fault propagation characteristics remain unclear. Furthermore, the experimental and operational timeframes are relatively short, and a complete and effective fault diagnosis method has not yet been established. Therefore, fault diagnosis in large aircraft energy systems faces a series of challenges. Consequently, the flight safety of C919 hinges on developing efficient fault diagnosis techniques and promptly detecting system anomalies.

Because of low accuracy and poor real-time performance in fault diagnosis caused by insufficient fault data and the long-distance dependence of operating data features in the energy system of the C919 aircraft, a fault diagnosis technology combining an auxiliary classifier generative adversarial network (ACGAN) module and a bidirectional long short-term memory (BiLSTM) network is proposed. First, given the problem of limited operating time and scarce fault data in energy systems, a data augmentation module with an ACGAN network as the core is constructed to effectively generate operating data and balance data of different categories. Second, owing to the long-term dependence of fault data, a feature extraction network based on BiLSTM as its core is constructed to accurately characterize the feature dependency relationship during fault evolution and improve the accuracy of fault diagnosis. Concurrently, because it is difficult to efficiently process key features affecting fault diagnosis, an accelerated calculation module based on an attention mechanism is designed to focus on critical fault feature information and greatly reduce information overload. This module achieves rapid processing of fault features and improves the real-time performance of fault diagnosis.

To verify the effectiveness of the proposed method, comparative experiments were conducted with commonly used diagnostic models, including support vector machine, long short-term memory (LSTM) network, empirical mode decomposition, convolutional neural network, and the wavelet transform (WAVELET) model. To reduce statistical errors, each model was tested 30 times, and the average values were calculated to obtain diagnostic accuracy, recall, F1 score, and diagnostic time. The experiments showed that the proposed method achieved the best results in fault diagnosis accuracy, recall, and F1 score. Although the diagnostic time was lower than that of the LSTM and WAVELET methods due to model complexity, the proposed method still maintains high applicability in practical applications.

The proposed method achieves good diagnostic performance across various fault modes, with higher diagnostic accuracy and real-time performance than existing fault diagnosis methods. In-depth analysis of fault modes based on operational data enables more accurate and rapid identification of fault types and locations, providing useful references for fault diagnosis and offering strong support for the flight safety of large aircraft.