| The power of the aircraft comes from the engine.Due to its complex structure and principle,it is difficult to intuitively model the corresponding relationship of its parameters.Therefore,the anomaly detection method of data-driven and machine learning shows great advantages in this task.There are some problems in the training of machine learning model:（1）There are few abnormal samples,the abnormal status is not clear.（2）Supervised learning is difficult due to the small number of label samples.The main work and research contents of this paper are listed as follows:（1）The possible abnormal states are analyzed and the temporary anomaly is selected to be the main type.The detection method for this data is designed based on self supervised learning and the characteristics of data source.The regressions include neural network model and random forest model.They use other sensor parameters to predict the measured value.The detectors use the D-value between the actual value and the predicted value and the difference value of the D-value to find abnormal state.（2）The principles and effects of convolution neural network,depth neural network,recurrent neural network,long short-term memory model and random forest model are analyzed.Convolution neural network,depth neural network and random forest model are used as the base model and fused to judge the abnormal state.（3）Semi supervised learning is used to optimize the performance of the base model.An improved Tri-training method of regression is designed to meet the needs of this study,the initial parameters and judgment conditions have been modified.The training effect of this method is verified on the flight parameter data.The optimized base model is fused by hard voting to form the anomaly detection model.Finally,a new engine anomaly detection method is designed,which is based on semi-supervised learning,its can effect is similar to supervised learning.This method reduces the pressure of manual marking. |
PDF Full Text Request