Research On Fault Diagnosis Of Aero-engine Accessory Case Based On Deep Learning

The heart of the transmission system is the aero-engine accessory casing.With the prolongation of the service time of the casing,the performance slowly degrades,and the risk of accidents in the casing gradually increases.Therefore,how to accurately diagnose the failure of the aviation engine and accessories is very important for the operation and maintenance of military and civil aircraft.Traditional fault diagnosis uses manual observation,which has the characteristics of low fault recognition rate and large amount of work.In order to solve these problems,intelligent diagnosis technology came into being,which can diagnose whether the machine is normal or not according to the vibration signal generated when the machine is running.In recent years,deep learning technology has performed prominently in the field of machine fault diagnosis.However,due to the effect of a lot of elements on the experimental data collection site of the aviation engine case,the acquired fault data or information is insufficient,resulting in the existing deep learning network unable to the actual diagnostic requirements.In this paper,we propose a fault diagnosis based on multiple deep learning methods.The work content includes the following directions:(1)A multi-channel and multi-scale image fault diagnosis integration method is proposed,to solves the problem of low fault recognition rate caused by the large body of the aviation engine case and the lack of singlechannel information.Resonance with the machine causes the problem of random noise in the data.In this paper,a multi-directional acceleration sensor is deployed on the hair-attached receiver to obtain multi-channel fault data.Time-frequency analysis is performed on the vibration signal of each channel to obtain the characteristic representation of the combination of time domain and frequency domain.Finally,ensemble learning is used to synthesize the model results of each channel to improve the probability of failure prediction.The neural network based on attention mechanism and multi-scale convolution kernel is used to capture the range of high and low frequency features,reduce the noise in vibration data,and enhance the weight of fault features.(2)A few-shot fault diagnosis method based on graph relationship and meta-learning is proposed to solve the problems of low model accuracy caused by insufficient fault data,poor generalization ability of the model in transfer learning and repeated fine-tuning.The model training strategy is to use meta-learning to extract general meta-knowledge from source domain data,so that the model can quickly converge when dealing with new tasks.The network model includes feature extraction network and graph learning network.Firstly,the fault features are extracted,and then a graph is constructed in units of meta-tasks,and the fault category of unknown samples is predicted by using the message passing mechanism of the graph neural network.The feature extraction network parameters and graph representations learned from the source domain are transferred to new tasks to improve the fault recognition rate in the case of small samples.(3)A personalized federated learning approach for fault diagnosis is proposed,to solve the data privacy and security issues of aviation engine accessories.This method contains multiple clients and a server.Each client model trains a local model using data augmentation and fault diagnosis methods.The server aggregates all local models into a global model.This paper proposes a new personalized aggregation algorithm,which customizes the weight of model parameter aggregation for each client and improves the recognition rate of the global model on each client.After multiple rounds of training,the optimal global model is obtained,and the data is not shared during the process,only the model parameters are shared.To sum up,this paper starts from the lack of fault information of the aviation engine case,and aims at three practical problems,and proposes a fault diagnosis strategy for the aviation engine and accessories based on deep learning.The experimental results show that the method in this paper outperforms similar methods in the corresponding tasks.