Research On Aeroengine Fault Diagnosis Method Based On Deep Learning

As the "pearl on the industrial crown",the reliability of aeroengine is the key to ensure the safety of aircraft operation.If the potential faults existing in its operation cannot be found in time and the inspection and maintenance are ignored,it will not only aggravate the degree of damage and affect its normal operation,but also cause aircraft damage and human death.It has caused huge losses to the safety of life and property.In recent years,deep learning technology has shown great potential in the related fields of fault diagnosis,and has achieved very good results.However,in the actual production environment,aeroengine fault diagnosis still faces great challenges:（1）different from the constant state of rotating machinery such as bearings,the operating state of aeroengine is extremely complex,which operates in a variable state environment,and the equipment often experiences repeated acceleration and deceleration processes.Data sets no longer meet the characteristics of cyclic and stationary,and the existing methods are difficult to model them effectively.（2）In the actual production environment,it is difficult to collect enough fault samples,because the engine,as the core system of the aircraft,is not allowed to easily fail,and the progress of failure is slow.Each acquisition is a secondary damage to the equipment.（3）Supervised learning needs to produce enoug h data and labels.However,data annotation is often time-consuming and troublesome,especially for signal processing.Sometimes it involves frame annotation of data,which is difficult.Too many data annotations raise the threshold of deep learning.In order to solve the above problems,this paper proposes an aeroengine fault diagnosis method based on deep learning strategies.The work includes the following aspects:1.According to the characteristics of real-time multi state transformation input signal of aeroengine,an aeroengine fault diagnosis method based on time-frequency image analysis is proposed.The method adopts the combination of short time Fourier transform（STFT）and sedensenet（senet densenet）,which is divided into two stages: time-frequency analysis and image classification.Firstly,STFT is used for time-frequency analysis to generate a single channel gray-scale time-frequency domain image and obtain the significant features such as state transformation;Then,sedensenet network with channel attention mechanism is used to extract and classify the features of fault time-frequency images.2.Aiming at the problem of small sample size of aeroengine data,a small sample fault diagnosis method based on twin network is proposed.Different from the traditional deep learning to carry out feature modeling to fit the distribution of each fault category,this method constructs a twin network based on sample feature comparison,uses the network to learn the similarity between samples,and uses samples of the same or different categories to learn.Firstly,a basic time-frequency image feature extraction network is built,and the two input samples share the weight parameters of the image feature extraction network;Secondly,whether the input sample pairs belong to the same category is judged according to the extracted feature similarity distance.This method effectively improves the fault diagnosis performance under small samples,and can obtain 2 to 4 times higher performance advantages than the baseline method on small sample data sets of bearings and variable speed aeroengines.3.In order to reduce the complex workload caused by data annotation and reduce the use cost of deep learning,a fault diagnosis method based on self supervised comparative learning is proposed,which can complete the training with a small amount of label data.Firstly,the sedensenet encoder is pre trained using the simclr framework to migrate the trained sedensenet to the downstream fault classification task;In the downstream ta sk,the encoder is fine tuned through MLP（multilayer perceptron）.The characteristic of this method is that only a small amount of label data can be used to complete the training,and can achieve better results than supervised learning,effectively reduce the complex workload caused by data labeling and reduce the cost of in-depth learning.In conclusion,combined with time-frequency analysis technology,this paper proposes an aeroengine fault diagnosis method based on deep learning strategies.Through many tests on the variable state engine fault data set,the results show that this technology can effectively realize the fault diagnosis of aeroengine under variable state;The fault accuracy under small samples is improved;The effect of using a small amount of label data to train the fault diagnosis model is realized,and good fault diagnosis performance is obtained under a small amount of label samples,which provides an important reference for the research of aeroengine fault detection.