Research On Intelligent Diagnosis Of Aircraft Turboprop Engine Gas Path Fault Based On Neural Networks

With the increasing dependence of human beings on aeronautical aircraft,the safety of aircraft has been paid more and more attention,and aviation engine is the core of determining the safety of aircraft.Turboprop aero engines can easily cause airpath failures due to the extremely high temperature and pressure generated by the internal combustion chamber.However,due to the lack of efficient turboprop aviation engine fault diagnosis technology,the current engine maintenance mainly depends on the traditional manual regular repair,this way usually consume a lot of resources and it’s very inefficient,so in order to improve the efficiency of turboprop aviation engine fault diagnosis,as early as possible to find the possible fault and ensure the safety of turboprop engine aircraft,the intelligent air fault diagnosis technology of turboprop aviation engine is very necessary.Therefore,this paper studies the application of onedimensional convolutional neural network,Short and Long-Term Memory Network and other deep learning techniques to the intelligent air path fault diagnosis of turboprop aviation engines.The data of turboprop engine air path fault belongs multi-dimensional time series data,so the convolutional neural network and Short and Long-Term Memory Network are considered in the study of this paper.Though the traditional convolutional neural network has powerful feature extraction ability,it cannot be applied directly to the air path fault data in this paper.In this paper,a lightweight one-dimensional convolutional neural network model called L1 Net is proposed,which uses a very simple architecture.L1 Net just includes a fixed-size 1D convolution kernel and a single convolutional layer,that can greatly reduce the number of parameters of the model.Besides,L1 Net almost prevents the model from overfitting during training,and meets real-time requirements.L1Net’s accuracy can achieve a 76.3% in our test dataset,then we further optimize L1 Net and propose other two variants of L1Net-g and L1Net-m based on L1 Net.For L1Net-g,we adopt the design of group convolution on the basis of L1 Net,which uses different size of convolutional kernels to extract features.We improve the accuracy of the L1Net-g to 77.9% in this paper eventually.For L1Net-m,we use a larger convolution sliding step and increase the number of network layers,so that more global features be extracted.L1Net-m’s accuracy can reach 81.2% in our test dataset.On the other hand,this paper applies Long and Short-Term Memory network to extract air-path fault data,then through the fully connected neural network to classify the extracted characteristics to achieve gas-path fault diagnosis.In addition,the effect of Long and Short-Term Memory network models with different internal hidden layer nodes on the efficiency of gas path fault diagnosis is also studied.Finally,the above several air path fault diagnosis models are modularly packaged,and actually are deployed to a certain type of turboprop aviation engine air path fault diagnosis platform,which achieve real-time diagnosis of air path fault.