Research And Implementation Of Abnormal Detection In Aircraft Agent Based On Machine Learning

The development of aircraft is gradually turning to intelligence and informatization,and aircraft malfunction often causes inestimable losses.In order to improve maintenance efficiency and ensure flight safety,it is of great significance to monitor and analyze the system data of the aircraft.The existing aircraft sensor data has a high degree of imbalance,high dimensionality,and complexity,which makes it difficult to effectively carry out feature retention and dimensionality reduction and simplification in feature learning.Secondly,the lack of samples and priors in data collection makes it difficult for traditional supervised learning to perform effective detection,while pure unsupervised learning,such as Auto-Encoder,cannot identify local errors and fit the data distribution.For the above problems,related research work is carried out to solve them in this thesis.First,an intelligent anomaly detection model based on variational long short-term memory（VTLSTM）is proposed to learn the low-dimensional feature representation from high-dimensional raw data,efficiently map the data,and constrain the reconstructed hidden variable.A lightweight estimation network then identifies anomalies.The proposed VTLSTM model can efficiently cope with the imbalance and high-dimensional issues in the complex high-dimensional data,and significantly improve the accuracy and reduce the false rate in anomaly detection.Second,an unsupervised anomaly detection model（MTGAN）based on generative confrontation is proposed,which designs an encoder-decoder network as a generator for time series data,and two discriminators to discriminate the network reconstruction ability and map to latent variables.The Wasserstein loss and the cycle consistency loss are associated with two discriminators to maximize the original distribution of the learned data and reduce overfitting.The reconstruction error and the discriminator output are finally combined to calculate the anomaly score.The proposed MTGAN model avoids the overfitting of the network while reconstructing the data,effectively enhancing its performance in reconstructing the data and detecting anomalies.Third,to demonstrate the performance of the approach,the method is tested with other representative methods at present.At the same time,the proposed method is applied to the anomaly verification platform of the aircraft system,and it is proved that the proposed method can effectively detect anomalies and has the highest averaged F1 score,which is superior to other methods in most cases.