Prediction Of Machinery Remaining Useful Life Based On Recurrent Convolutional Neural Network

In order to meet the needs of industrial development and production,the mechanical structure of modern industrial system is becoming more and more complex,so more attention and investment are needed to ensure its safety.The remaining service life(RUL)of machinery is defined as the length from the current time of machinery to the end of its service life.Accurate RUL estimation of machinery plays a key role in ensuring production safety.The traditional RUL prediction method based on physical model needs specific and accurate physical system,which is not flexible and efficient;The data-driven method of machine learning uses sensor data and uses machine learning method to estimate RUL,but the structure is relatively shallow,and it is difficult to effectively mine high-dimensional features.The design of feature mining and prediction separately is difficult to synchronize and optimize.The performance of traditional algorithms is limited,and it is difficult to obtain satisfactory diagnosis results;The neural network is constructed based on the data-driven method of deep learning.CNN and RNN have the advantages of extracting high-dimensional features and modeling time series data respectively,and can achieve end-to-end efficient prediction of RUL.However,there are two limitations in the current work of applying CNN to RUL prediction:(1)The time dependence of different states is not considered;(2)The uncertainty of the result cannot be predicted by RUL.Aiming at the above two problems,this paper proposes a mechanical RUL prediction method based on cyclic convolution neural network.First of all,a Reccurent Convolution layer is proposed to build a Reccurent Convolution Network.The construction of the network conclude three structures: the CNN-RCNN connected by the convolution layer and a single Reccurent Convolution layer,the CNN-RCNN(Encoder-Decoder)connected by the convolution layer and two Reccurent Convolution layers,and the RCNN overlapped by multiple Reccurent Convolution layers and the Pooling layer stack.In addition,three Reccurent structures are used to generate three Reccurent Convolution layers: Conv LSTM,Conv JANET,Conv GRU.Secondly,in order to quantify the uncertainty of RUL prediction and obtain its probability distribution,this paper proposes to set a Dropout with a probability of π for each Reccurent Convolution layer and full connection layer,and the attenuation coefficient is λ L2 regularization,and then carry out V times random forward learning to quantify the uncertainty of the model through the variance of multiple prediction results.Then the two CNN-RCNN proposed in this paper are compared and studied on the C-MAPSS dataset.The Reccurent Convolution layer uses Conv LSTM and Conv JANET.The comparison idea,evaluation index and loss function of the experiment are designed,and the RUL of the turbofan engine is predicted on the test dataset.The conclusions are as follows:(1)The convergence of all models is good,and the convergence stability of the Encoder-Decoder structure is higher than that of the single-layer structure;(2)Compared with the real RUL,the predicted RUL of the model has large dispersion and certain random error.The error preference is not obvious and the variance is not constant.(3)JANET outperforms LSTM in terms of various indicators,regardless of single-layer structure or codec structure.(4)In addition to FD003,the indicators(RMSE,Score)of the Encoder-Decoder to predict RUL are better than the single-layer,indicating that increasing the processing of the Decoder can improve the prediction effect more than increasing the number of convolution cores.Finally,the performance of the stacked RCNN proposed in this paper is studied on the FEMOT-ST dataset,and the influence of network depth on the prediction effect and the influence of Dropout probability on uncertainty quantification are explored.In order to reflect the prediction advantages of RCNN,the two basic models and the three more advanced models are compared with it.The comparison results show that the RCNN proposed in this paper has the best effect in predicting RUL,which verifies the effectiveness of the RUL prediction method proposed in this paper.