System Fault Prediction And Maintenance Decision Based On Deep Learning

With the development of new technologies,the products are attached with more functions and better performance.At the same time,as the physical structures and operation mechanisms become more and more complicated,traditional failure mechanism based fault prediction methods are not applicable to new-tech products.Based on the fact of a large amount of monitoring data acquired by sensors and the further development of health management,it is desirable to establish a bridge between big data and intelligent health management,from off-line detection to real-time monitoring and from single mode to intelligent fault prediction,which has become an urgent research topic.Based on the massive monitoring data,this paper researches on the fault prediction and maintenance decision-making problems.Combined with the development of sensor technology and deep learning technology,the research framework of system fault prediction and maintenance decision-making based on deep learning is first presented.Then,guided by the framework,this dissertation sets up relationship between the system degradation characteristics and the remaining life,according to whether the fault is single or whether the system is a single component,respectively.Recurrent neural network(RNN)and convolutional neural network(CNN)are proposed to learn the degradation characteristics of the system,and then to predict its remaining life when making maintenance strategy formulation,this dissertation considers both the component-level maintenance strategy and the transition from component-level system-level maintenance strategy.This main research and innovations of the thesis are summarized as follows:(1)A fault prediction model for single failure mode is proposedFor the system fault prediction and maintenance decision problem with single failure mode,in the fault prediction phase,the mapping relationship between the characteristics and the remaining life is established.The correlation between time dimensions of the monitoring data is used to mine the dynamic degradation of system.Taking into account that the different characteristics of different monitoring indicators from different aspects of system degradation,a bidirectional LSTM model with attention mechanism is proposed.(2)A fault prediction model for multi-failure mode is proposedFor the system fault prediction and maintenance decision problem in multi-fault mode,in the fault prediction stage,based on the mapping between the characteristic variables and the remaining life,high quality degradation information is learned from a large number of multivariate monitoring data.The degradation information has the advantage the characteristics of CNN graph feature extraction and introduces the attention mechanism.A CNN prediction model with channel-space attention mechanism is proposed.In the maintenance decision phase,the relationship between fault prediction,fault diagnosis and maintenance optimization processes are set up,which were combined with utility theories to guide the development of maintenance decisions.(3)A machine learning framework for fault prediction of multi-component system is proposedFor the fault prediction and maintenance decision-making of multi-component systems,considering the complex coupling between component degradation behaviors,it is difficult to model the a priori physical knowledge.The data-driven prediction model becomes an effective means to solve this problem and can help develop effective statebased maintenance decisions.Based on the data generated by the numerical simulator of the naval ship propulsion system,the regression model of the components GT and GTC is established,and the prediction effects under different models are analyzed.The analysis of feature importance is to find highly predictive feature variables that characterize component degradation,thereby helping to develop early warning mechanisms for system monitoring and to guide condition-based maintenance decisions.