Research on fault diagnosis methods for certain civil aircraft based on human-machine interaction data enhancement

The structures of modern civil aircraft are becoming more and more complex, particularly with regard to aeroengines, where faults are often challenging to detect in the early stages. The implementation of Engine Control Units (ECU) in aviation provides an avenue for storing a substantial amount of historical operational data, forming the basis for data-driven fault diagnosis of aeroengines. In order to conduct research on defect diagnosis, this paper gathers the historical operation data of a certain type of civil aircraft. However, the majority of historical data consists of normal operation data, creating an imbalance between data classes, especially in terms of fault data. In addition, the complexity of engine operation modes, with dynamic operations like acceleration and takeoff having shorter durations and less data compared to horizontal flight modes, exacerbates the imbalance in data classes. The imbalance between data classes and imbalance in data classes result in missing and false alarms of many diagnosis results of this type of civil aircraft by using existing fault diagnosis technique. Furthermore, autocorrelation between variables and distinct dynamic characteristics further complicate the issue. To address these challenges, a fault diagnosis model based on a Long ShortTerm Memory (LSTM) network was developed to extract dynamic characteristics from the system. Human-machine interaction was employed to identify areas of missing and false alarms, addressing the imbalance in two types of samples. Imbalanced data was then enhanced using the Generative Adversarial Network (GAN), and the training process was manually reinforced. Experimental results demonstrated that the fault diagnosis model, augmented through human-machine interaction data enhancement, effectively captured dynamic characteristics in the data of a specific civil engine. It enhanced dynamic mode data samples and reduced the rates of missing and false alarms, validating the feasibility of the model.