Research On Equipment Remaining Useful Life Prediction Algorithm Based On Deep Learning

With the general progress of war theory,commanders are asked to acquire the health status of weapons in real time.At the same time,the traditional equipment maintenance strategy is inefficient and expensive,which can’t meet the needs of modern war.In this case,Prognosis and Health Management of weapon systems has become the key point to war.As the essential segment of PHM,the prediction of Remaining Useful Life has direct impact on the combat deployment and command decisions of weapon maintenance,which is worthy to study.The stable operation of weapon depends on key components such as engine and battery pack,how to improve the RUL prediction accuracy of components have always been the focus of research.And due to the influence of data noise,imperfect algorithm,how to characterize the uncertainty of the results while obtaining the prediction results is also a big challenge.The main works of this thesis are as follows.Firstly,considering the sensor data of weapon have many features,a feature selection algorithm based on clustering and joint mutual information is proposed.According to the characteristics of the data,the algorithm automatically generates a feature subset with high correlation and low redundancy.The feature subset has a good performance in improving the calculation efficiency and prediction accuracy.Secondly,in order to improve the prediction accuracy of the algorithm,a hybrid neural network model based on attention mechanism is proposed by making comprehensive use of the powerful feature extraction ability of Convolutional Neural Network,the characteristics of Long-Short Term Memory Neural Network good at processing long-time series data and the ability of attention mechanism to automatically learn key information.And compared with recent research results on the CMPASS data set,it proved the effectiveness of the hybrid network model.Finally,in order to obtain the uncertainty of the prediction results,the Bayesian Neural Network and the Dropout-based Approximate Bayesian Neural Network technology are studied,the characteristics of this two networks are analyzed experimentally,and the dropout technology is applied on the newly proposed hybrid model to obtain the uncertainty while maintaining high prediction accuracy.