- 4 Datasets (FD001-4) consist of multiple multivariate series.
- Each dataset has 3 operational settings and number of sensor measurements which are contaminated with noise over time.
- Taking a look at example of sensor measurements for FD001 jet, we can see non-linear trends in some, caused by noises in sensors.
- Since the non-linearity holds, neural network approach should be advised.
- We want to build a RNN that can handle a stream of sensor measurements and produce a one-hot-encoded vector in real time.
- Each jet supposedly has different type and number of sensor measurements. Each unit has a different number of time cycles. Therefore, we will make a model for each of the jets.
- We can structure the data such that each unit's data is treated as a separate sequence and train our batches like that.
- Reformat the problem: If we can model the RUL of a certain jet, we should be able to split the measurements to * INFO -> WARNING -> FAULT*.
- Measurements classification splitting (garbage-in, garbage-out). It's a subject matter expertise knowledge to decide when the model should be warned and when does the fault start.
- Our base model was prone to the overfitting, we also got double descent at 130 epochs, probably because of over parametrization of the model, since the dataset is relatively small.
- We solved some of the overfitting problems by:
- implementing early stopping (patience)
- increasing dropout rates in model layers
- keeping model complexity reduced to only a few layers
- adding L2 regularization penalties
- Validation calculation is slow and should be optimized for bigger datasets.
- Classes 0,1 and 2 are not evenly distributed.
| FDOO1 | FD002 | FD003 | FD004 |
|---|---|---|---|
 |
 |
 |
 |
| FDOO1 | FD002 | FD003 | FD004 |
|---|---|---|---|
 |
 |
 |
 |
Residual is calculated as:prediction - target:
- In our case, when we subtract target vector from prediction vector, we get a set of values
{-2, -1, 0, 1, 2}:0is our goal, since it means that the prediction is the same as the target- Since models are lightweight and not perfect, we agree to make our models a bit more sensitive (skewness to the right in the residuals graphs), so that in real life, the prediction of fault would come earlier than the actual fault.