Multi-mode Fault Diagnosis And Residual Useful Life Prediction Based Deep Learning

The structure of automation equipment is becoming more and more complex,making the equipment fault problem inevitable.Using equipment predictive maintenance technology can achieve the transition from planned maintenance to condition-based maintenance,it is not only ensuring the safe and efficient operation of equipment but also greatly reducing maintenance costs.Fault diagnosis and residual life prediction are two important steps for predictive maintenance of equipment.However,changes in raw materials,load changes or other factors will cause the system in multi-mode operating conditions,thus affecting the accuracy of fault diagnosis and residual life prediction.Without the precise mechanism model of large-scale equipment,we adopt the feature extraction method of deep learning and study data-driven key technologies for equipment predictive maintenance.Focusing on solving the problem,which traditional deep learning methods cannot predict multi-mode fault diagnosis and residual life accurately.The main innovations of the paper are as follows:(1)The idea of building a Hierarchical Deep Neural Network(HDNN)is proposed.Firstly,we aim to achieve an accurate mode division for multi-mode data.Then,some different DNN diagnosis models are established for each mode on the second level of the HDNN,which achieves accurate diagnosis of multi-mode faults.On the third level of the HDNN,different DNN models are constructed according to the different modes and different faults,realizing to distinguish the severity of different fault types and providing decision-making reference for predictive maintenance.(2)A differential LSTM method is proposed to solve the problem of inaccurate extraction of autocorrelation features of slowly varying faults.The purpose of early multi-mode fault diagnosis can be achieved without mode division.Then,the fault dynamic evolution features are extracted by differential LSTM,which is taken as the input.And the actual residual life of the training samples,which is seen as the output.Constructing the model of residual life prediction on the basis of neural network,and we realize real-time diagnosis of early failure and online prediction of remaining life.