Research On Fault Diagnosis Technology Of Fiber Optic Gyro Based On Neural Network

Fiber optic gyro is a medium high precision angular velocity sensor commonly used in various inertial navigation systems.It is one of the core components of inertial navigation systems,especially in military and aerospace applications.With the continuous development of China's military modernization,it is necessary to ensure the high performance of fiber optic gyroscopes with good stability.Therefore,the fault diagnosis of fiber optic gyroscopes has very great significance.Fiber optic gyroscopes generate a large amount of data in the process of production,testing,and equiping.These data often contain a large amount of unexcavated information with value.Proper use of these data can improve the level of fiber gyro fault diagnosis technology.Neural networks are currently a hot area of global research and are widely used in all social circles.In this theis,the fault diagnosis technology of fiber optic gyroscope based on neural network is researched.The fiber optic gyroscope fault diagnosis system based on neural network is realized.In this thesis,the fiber optic gyroscope is taken as the research object.The fiber optic gyroscope's physical model and the random noise existing in the fiber optic gyroscope work are researched.The digital model of the fiber optic gyroscope is established and the fiber optic gyroscope fault output is analyzed.This thesis analyzes five types of signals and generate simulation data of fiber optic gyroscope failure,which solves the problem of data set acquisition based on the digital model of fiber optic gyroscope.Aiming at the problem of feature extraction of fiber optic gyroscope fault data,wavelet transform and wavelet packet decomposition and reconstruction algorithm are researched.According to the characteristics of fiber optic gyroscope fault signals,db4 wavelet is selected as the wavelet basis function to decompose the fault signal into three layers of wavelet packet to extract different frequency bands.Energy information is used as a feature vector.Aiming at the problem that the classification of various fault types is not high enough after wavelet packet decomposition,an improved method based on mean proportional amplification is proposed,which increases the discrimination of eigenvectors of different fault data and solves the preprocessing problem of fiber optic gyroscope fault data.Aiming at the effective training problem of multi-layer neural networks,the error back propagation algorithm is researched,and the advantages and disadvantages of different loss function,objective function,activation function and various hyperparameter optimization methods are analyzed.The principle of convolutional neural network is researched.According to the characteristics of fiber optic gyroscope fault signal,the basis of choice of deep feedforward neural network and convolutional neural network is proposed,and a suitable hyperparameter optimization algorithm is selected.Aiming at the fault diagnosis of fiber optic gyroscope,the fault diagnosis algorithm based on deep feedforward neural network is designed and implemented.The reason why the sigmoid activation function causes the neural network overfitting is analyzed.The Dropout method is used to improve the generalization performance of the algorithm and gradient vanishing is solved.Aiming at the problem of partial fault data being diagnosed incorrectly,an improved method of adding convolutional neural network is proposed and implemented by using Tensorflow.It is proved that the addition of convolutional neural network can effectively improve the accuracy of fault diagnosis.