tbigru
The TBiGRU approach uses a feature selection mechanism to mine transferable features and a bearing running state detection to determine the first-time-to-predict. The training is done with an MMD approach.
The feature selection uses a distance measure based on Dynamic Time Warping and the Wasserstein distance. From a set of 30 common vibration features the ones with the smallest distance between source and target domain are selected. These features serve as inputs to the network.
The first-time-to-predict (FTTP) is used to generate the RUL labels for training. FTTP is the time step where the degradation can be detected for the first time. The RUL labels before this time step should be constant. The TBiGRU approach uses the moving average correlation (MAC) of the energy entropies of four levels of maximal overlap discrete wavelet transform (MODWT) decompositions to determine four running states of each bearing. The end of the steady running state marks the FTTP.
TBiGRU was introduced by Cao et al. and evaluated on the FEMTO Bearing dataset.
VibrationFeatureExtractor
This class extracts 30 different features from a raw acceleration signal.
The features are: RMS, kurtosis, peak2peak, standard deviation, skewness, margin factor, impulse factor, energy, median absolute, gini factor, maximum absolute, mean absolute, energies of the 16 bands resulting from wavelet packet decomposition, standard deviation of arccosh and arcsinh. If the input has n features, n*30 features are extracted. Additionally, it features a scaler that can be fit to scale all extracted features between [0, 1].
__call__(features, targets)
Extract the features from the input and optionally scale them.
The features should have the shape [num_windows, window_size,
num_input_features] and the targets [num_windows].
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
features |
ndarray
|
The input features. |
required |
targets |
ndarray
|
The input targets. |
required |
Returns:
| Type | Description |
|---|---|
Tuple[ndarray, ndarray]
|
The extracted features and input targets. |
__init__(num_input_features, feature_idx=None)
Create a new vibration feature extractor with the selected features.
The features are sorted as f1_1, .., f1_j, ..., fi_j, where i is the index of
the computed feature (between 0 and 30) and j is the index of the raw
feature (between 0 and num_input_features).
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
num_input_features |
int
|
The number of input features. |
required |
feature_idx |
Optional[List[int]]
|
The indices of the features to compute. |
None
|
fit(features)
Fit the internal scaler on a list of raw feature time series.
The time series are passed through the feature extractor and then used to fit
the internal min-max scaler. Each time series in the list should have the
shape [num_windows, window_size, num_input_features].
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
features |
List[ndarray]
|
The list of raw feature time series. |
required |
Returns:
| Type | Description |
|---|---|
VibrationFeatureExtractor
|
The feature extractor itself. |
mac(inputs, window_size, wavelet='dmey')
Calculate the moving average correlation (MAC) of the energy entropies of four levels of maximal overlap discrete wavelet transform (MODWT) decompositions.
The wavelet is a wavelet description that can be passed to pywt. The default
wavelet was confirmed by the original authors. For more options call
pywt.wavelist. The input signal should have the shape [num_windows,
window_size, num_features].
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
inputs |
ndarray
|
The input acceleration signal. |
required |
window_size |
int
|
The window size of the sliding window to calculate the average over. |
required |
wavelet |
str
|
The description of the wavelet, e.g. 'sym4'. |
'dmey'
|
Returns:
| Type | Description |
|---|---|
ndarray
|
The MAC of the input signal which is |
modwpt(inputs, wavelet, level)
Apply Maximal Overlap Discrete Wavelet Packet Transformation (MODWT) of level
to the input.
The wavelet should be a string that can be passed to pywt to construct a
wavelet function. For more options call pywt.wavelist. The implementation was
inspired by this repository.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
inputs |
ndarray
|
An input signal of shape |
required |
wavelet |
str
|
The description of the wavelet function, e.g. 'sym4'. |
required |
level |
int
|
The decomposition level. |
required |
Returns:
| Type | Description |
|---|---|
ndarray
|
The 2**level decompositions stacked in the last axis. |
select_features(source, target, num_features)
Select the most transferable features between source and target domain.
30 features are considered: RMS, kurtosis, peak2peak, standard deviation, skewness, margin factor, impulse factor, energy, median absolute, gini factor, maximum absolute, mean absolute, energies of the 16 bands resulting from wavelet packet decomposition, standard deviation of arccosh and arcsinh. If the input has n raw features, n*30 features are extracted.
The dev splits of both domains are used to calculate a distance metric based on
Dynamic Time Warping and the Wasserstein Distance. The indices of the
num_feature features with the lowest distances are returned.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
source |
AbstractReader
|
The reader of the source domain. |
required |
target |
AbstractReader
|
The reader of the target domain. |
required |
num_features |
int
|
The number of transferable features to return. |
required |
Returns:
| Type | Description |
|---|---|
List[int]
|
The indices of features ordered by transferability. |