Fault Diagnosis Of Underwater Vehicle Propulsion System Based On Deep Learning

With the development of ocean exploration strategy,the importance of underwater vehicle which is the main carrier of underwater equipment has been widely recognized.The stability of the underwater vehicle under complex working conditions is what people concerned with.Many research focus on how to detect the failure of the underwater vehicle in time,meanwhile,as the most basic movement performance guarantee of the underwater vehicle,the fault diagnosis of the propulsion system is one of the most important things.The underwater vehicle has been in the underwater environment with multiple degrees of freedom,high pressure and unknown complex conditions.And the noise sources are complex,the working environment is diversified and unstable,the fault forms are diversified,the cabin capacity and power limitation are strict(it has little hardware redundancy),and the performance and types of sensors of different test bed are also varied.Because the acquisition of underwater experimental data is difficult,especially the simulation of underwater operation fault situation,it is risky,so it is difficult and significant to research on related problems.The innovation of this dissertation is to propose a fault diagnosis method for underwater vehicle propulsion system based on deep learning,we build a data acquisition platform independently,and use LabVIEW to simulate and collect 37 kinds of faults and normal state data,then simply preprocess data in MATLAB,then use TensorFlow to train and save fault diagnosis model,and use integrated learning to upgrade model.Finally,the model is imported into LabVIEW for on-line fault diagnosis.The experimental results show that,compared with traditional methods,this method can simultaneously realize feature extraction,feature dimensionality reduction and fault diagnosis in one model,greatly improving the efficiency of diagnosis.The average training set accuracy of the model is 97.85%,the average debugging set accuracy is 94.72%,the average test set accuracy is 96.68%,and the integrated learning test set accuracy is 98.52%.The experimental method and process presented in this dissertation can be used for off-line training and on-line fault diagnosis of underwater mechanical and electrical systems and other similar fields.It can achieve better human-computer interaction,higher diagnostic efficiency and accuracy.