Fault Diagnosis Of Aviation Power Equipment Based On Deep Learning Algorithms

With the development of more-electric aircraft(MEA),the reliability and the safety of aviation power equipment become much more important.Reasonable maintenance and use of aviation power equipment is one of the foundations to ensure the continuous airworthiness of aircraft.Their Major faults and repeated faults are one of the supervision points of aircraft continuous airworthiness.As an important secondary power equipment on the aircraft,once the transformer rectifier units(TRUs)break down and cannot repair in time,which will lead to a serious failure of system,huge economic losses or casualties.In order to ensure the continued airworthiness of aircraft throughout the whole flight cycle,an efficient fault diagnosis of TRUs used to improve the reliability of aviation power system effectively,detect and handle the equipment failure timely.With the development of artificial intelligence technology,deep learning algorithm is applied to the fault diagnosis research of aviation TRUs.With the combination of the structural characteristics of TRUs,the research focuses on several problems existing in the actual diagnosis from data collection to the final fault diagnosis.The main contents are as follows:(1)Determine the failure modes of TRUs and obtain the datasets of different TRUs.Based on the study of TRU structures,TRUs were classified from transformer structure,voltage output symmetry and pulse number of rectifier bridge.Then,from the perspective of initial airworthiness,based on the FHA of aircraft power system,the FMEA and FTA of TRU were studied to determine the logical relationship between different fault modes of TRUs.MATLAB and test prototype of various TRUs were constructed to obtain and compare different electrical signals under different fault modes.(2)A hierarchical fault diagnosis method based on discrete time series convolutional neural network was developed.Firstly,the structural characteristics of the existing CNN-based fault diagnosis methods were studied,and the discrete time series convolution neural network(DTCNN)was designed.In DTCNN,large convolutional kernel was adopted in the first convolutional layer and Dropout with variable probability was added to improve the robustness of the network.Secondly,combined with the fault location and characteristics of TRUs,a hierarchical fault diagnosis method based on DTCNN was proposed.Finally,the DTCNN structure is determined and verified by TRU simulation and test data set.Compared with other CNN structures in terms of accuracy and network complexity,the effectiveness and advantages of the model were verified through network visualization.(3)The methods to deal with class imbalance of datasets were studied.Because the equipment is in normal condition for most of the time and the probability of different failure modes is different,the collected data sets often have class imbalance problem,which will greatly affect the diagnosis accuracy.The relevant parameters of three methods to deal with class imbalance problem which are suitable for CNN model were reset in this paper.A comparison of three methods and their combination algorithms,and different degree of sampling was made based on 9 TRUs datasets with different imbalance degrees.Processing methods applied to CNN models to deal with different kinds of imbalance degree problems were proposed.(4)For the confusing fault modes in TRU fault diagnosis,a progressive fault diagnosis method based on stack convolution neural network is developed.Firstly,the technical characteristics of the classic image recognition model based on CNN were compared and the stacked convolutional neural network(SDCNN)model was constructed.In SDCNN,the stack of small convolution kernels was used to improve the recognition ability of the model for the small features;the global average pool layer was used to realize the regularization of the model and reduce the total parameters of the model.Secondly,based on a variety of TRU electrical signals,a progressive fault diagnosis method based on multiple inputs was proposed to further improve the diagnosis accuracy of confusable faults.Finally,the performance of the method was verified by the datasets which contains 34 fault modes of 24 pulse TRUs;and the anti-noise property and robustness of the method were verified by the datasets containing noise signals of different degrees.(5)The rapid construction of different types of TRUs fault diagnosis models was studied.Lots of manpower and time are needed while establishing or training fault diagnosis models for each TRUs,since there are many kinds and structures of TRUs.Therefore,based on the idea of transfer learning,the TRU datasets which suit for the source domain of transfer learning and the influence factors of the transfer networks for different TRUs were analyzed with 13 datasets of 13 TRU structures.Fault diagnosis models for different TRUs were constructed based on transfer learning,which provide a method and a reference for the fast building of TRU fault diagnosis model.