Research Article
Sequential Network with Residual Neural Network for Rotatory Machine Remaining Useful Life Prediction Using Deep Transfer Learning
Algorithm 1
The RUL algorithm for training.
| RUL algorithm for training is defined as follows: | | Input: Historical run-to-fail sensors data of rotatory machine | | For do: | | j is size of the time window | | Convert time-series sample into grey image | | Duplicate and extend channel into 3 | | | | The whole datasets | | end | | Define corresponding RUL value | | Function TRAIN (X, N, d): | | t = [t − Min(ttraining)]/[Max(ttraining) − Min(ttraining)] Normalization t | | NN means Maximum number of iterative training d represents minimum allowable training error | | Res-Conv2 rain the ResNet-50 model by ImageNet datasets and use the first two blocks of ResNet-50 as pretrained feature extractor. | | , b Initialize parameters of Bi-LSTM module and fully connected layers | | Z = Random. Sample(X,2000) randomly select 2000 samples as training sets Z | | While N > n or L < d: | | For do: c = Res-Conv2(Z) | | h = Bi-LSTM (c) ⟶ h is the vector extracted by Bi-LSTM | | RUL = MLP(h) ⟶ The RUL value is output from fully connected layer. | | | | | | | | Output: the trained proposed model. |
|