Aero-engine Anomaly Detection Based On Deep Learning

Efficient and accurate anomaly detection is critical to safeguarding the safety of aero-engine operations and improving the economy of the airline.At present,the anomaly detection methods that are actually applied are single and have low efficiency.Therefore,it is of great theoretical value and practical valu e to study the anomaly detection method of aero-engines.Because of its multi-layered structure,deep learning can learn effective features from a large number of input data.The learned high-order features contain many structural information of input data,which is of great help in improving the effect of anomaly detection.Therefore,this paper combines the characteristics of aero-engine monitoring data and incorporates the idea of deep learning,and proposes several aero-engine anomaly detection methods based on deep learning.In view of the characteristics of aero-engine point anomalies,and taking into account that the Stack Denoiseing Auto Encoder(SDAE)can unsupervisedly extract features,an aero-engine point anomaly detection method based on SDAE-BP is proposed.The method includes two main parts: feature extraction based on SDAE and anomaly detection based on BP neural network.Aiming at the difficulty of determining the structure of SDAE,a method for determining the SDAE structure based on correlation analysis was proposed.In order to realize efficient aero-engine time series anomaly detection,Aiming at the anomalous data characteristics of the aero-engine time series,inspired by the convolutional neural network(CNN)is commonly used in recognition image in the computer vision field,the convolutional neural network was introduced into field of aero-engine time series anomaly detection,and an aero-engine time series anomaly detection method based on CNN-BP is proposed.The methods include: optimization of ranking aero-engine monitoring parameters,feature extraction based on CNN and anomaly detection based on BP neural network.Among them,in order to solve the problem of sequencing optimization of aero-engine monitoring parameters,a parameter scheduling optimization method based on dynamic programming and a parameter ranking optimization method based on heuristic greedy algorithm are proposed.In order to improve the effect of anomaly detection,an aero-engine abnormal detection algorithm based on ensemble learning is proposed.According to the different basic classifiers,it is divided into three methods: an ensemble learning point anomaly detection method based on BP neural network,an ensemble learning point anomaly detection method based on SDAE-BP,and an ensemble learning time series anomaly detection method based on CNN-BP.Finally,based on several aero-engine anomaly detection methods studied in this paper,combined with the needs of aero-engines,a prototype system for aero-engine anomaly detection was designed and developed,and it can provide technical support for anomaly detection of aero-engine monitoring parameters.