Research On Equipment Key Subsystem Condition Assessment Based On Deep Learning Algorithm

With the development tendency of intelligent manufacturing,the requirement of reliability and intelligent maintenance for modern industrial equipment is further improving.To meet this challenge,assessing equipment health condition through artificial intelligence is the key solution.Large machinery equipments usually have complex structure with many components,and the components of the same subsystem are highly coupled.Therefore,the health condition of key subsystems has an important impact on the quality and reliability of complex equipment.As the bridge between the physics and information,industrial big data contains a wealth of information that can reflect the condition of equipment comprehensively.Hence,there is great significance to apply industrial big data mining by using deep learning algorithms to assess health condition for the key subsystems of equipment.Firstly,this paper summarizes the opportunities and challenges which health condition assessment under mechanical big data.Secondly,the advantages and application feasibility of deep learning algorithm is provided.Then,this paper assessed condition of equipment and its key subsystems by health condition recognition and remaining useful life(RUL)estimation based on the mechanical big data obtained from monitoring.Compared with other methods,the proposed two models have higher intelligentialize degree and better results.This study not only enriches the theory of prognostics and health management but also provides some methods for the key subsystems such as gearbox and engine.The main work of this paper is as follows:(1)The condition recognition of key subsystems based on mechanical big data faced three challenges: inadequate data processing ability of model,low intelligence level and large conflict between evidence.The combination of deep learning algorithm and improved evidence theory can solve these problems.Firstly,this paper constructs an open recognition framework based on deep sparse autoencoder neural network and support vector classification algorithm.Secondly a modified evidence fusion method for subsystem health condition recognition based on information content and conflict measurement between evidences is presented.Further,an open framework evidence fusion model based on information quantity combined with D-S synthesis rules(IE-OFET)is proposed.Finally,case study proves that this method obeys operational law and has higher accuracy.(2)Traditional approaches usually construct health indicators first and then estimate the current condition by a preset threshold.However,human factors have great influence on indicators extraction and the approaches are difficult to process big data.Hence,this paper proposes a novel approach named Deep Self-normalizing Convolutional Neural Network(DSCNN),which predicts RUL directly from sensors data without any requirement of expert knowledge and can learn characteristic automatically.In addition,scaled exponential linear units(Selu)are applied to construct the network.This procedure improves noise immunity of the model and reduces the training difficulty.In case study,results comparison with popular state-of-the-art data-driven methods shows the DSCNN is more accurate and has competitive performance.