A NOVEL DEEP PARALLEL TIME-SERIES RELATION NETWORK FOR FAULT DIAGNOSIS

Introduce a 2022 A paper from the University of Electronic Science and Technology , Deep parallel sequential network

0. Abstract

Considering the application of the context information of time series data, the model can improve the performance of fault diagnosis , Some networks （RNN,LSTM,GRU） It is more effective for the diagnosis results . then , These models are limited by computational complexity , It is difficult to achieve high diagnostic efficiency . Due to large convolution kernel or deep architecture , parallel CNN It is also difficult to achieve high efficiency in dealing with long-term information .BERT The model uses absolute location coding to introduce context information into the model , But it also brings noise . In order to deal with the above problems , Put forward deep parallel time-series relation network(DPTRN). There are three advantages ：1. DPTRN be based on MLP framework , It can effectively improve the computing efficiency .2. Improve absolute position coding , Using decoupled position coding unit, context information can be learned .3. It has obvious advantages in feature interpretability .

1. introduction

PCA,ICA,RVM,DBN, Other diagnostic methods . However, industrial processes change continuously , therefore , It is necessary to model each stage of industrial process by using temporal characteristics .

  The model architecture is as follows ：

1. Establish time series data set

2. Divide the current time node D(T) And historical time nodes D(k),k=1,2,…,T-1.

3. Input each historical time node and current time node into the timing unit .

4. Generate absolute position code , And use location query query Matrix and position key The matrix is mapped to decoupled position coding .

5. Add the outputs of the timing information unit and the decoupling position coding unit to obtain the time relationship weight of each historical node .

6. The feature stitching of historical information vector and current time node （contact） As the input of the classification layer, get the fault diagnosis results .