Research On Engine Remaining Life Prediction Based On Data-driven Methods

The complexity of machinery and equipment is increasing as the industrial Internet and industrial manufacturing systems continue to evolve.The traditional maintenance strategy is prone to issues such as "under-maintenance" or "overmaintenance," making meeting actual maintenance needs difficult.In the context of rapid intelligence development,it is critical to make full use of the large amount of data generated by the equipment for Remaining life Prediction(RUL)in order to achieve safe and reliable operation of equipment systems.The author suggests two real-world approaches for equipment application that are data-driven.The following are the primary research findings and contents.(1)An attention-based multichannel hybrid neural network is proposed to address the issue that devices do not adequately consider different degradation patterns in feature extraction,resulting in poor RUL accuracy under complex operating conditions.This method can fully consider the characteristics of short-term fluctuation randomness and long-term change stability hidden in time series.The 1D Convolutional Network(1DCNN)is used to extract the short-term characteristic features in the time series,and the Bi-directional Long Short-Term Memory(Bi LSTM)network is used to extract the long-term characteristic features.Furthermore,the channel attention mechanism improved by Global Stochastic Pooling(GSP)and the spatial attention mechanism improved by Global Covariance Pooling(GCP)are used to apply the improved attention to the two channels respectively,to enhance the extraction ability of the model to extract degraded features and prevent overfitting.When compared to the conventional network,four sub-data in the C-MAPSS dataset are used for validation,which has better accuracy performance in complex working conditions.(2)For the current problem of high model complexity of RUL prediction and a large number of parameters that cannot respond to the device state in a timely manner,a light-form parametric self-gating prediction network is proposed.The method is based on the Light Gated Recurrent Unit(Li-GRU),and it begins by replacing the original Rectified Linear Unit(Relu)with the Parametric Rectifier Linear Unit(PRelu)activation function to ensure efficient operation while avoiding the problem of neuron death.Second,introducing batch normalization in the feedforward connection can improve model convergence speed and avoid the problem of covariate shift caused by distribution changes during data propagation between layers.The experimental results show that the FD001 and FD003 sub-datasets have the best performance.A single sample calculation time is 0.121 ms and 0.137 ms,and the model size is only 20.7%and 26.4% of the model proposed in Chapter 3.The results show that the proposed model has the advantages of being lightweight and having a quick prediction time.