alignment

These losses are used to create a latent space that is conductive to RUL estimation. They are mainly used by the LatentAlignmentApproach.

DegradationDirectionAlignmentLoss

Bases: Metric

This loss is used to align the direction of the degradation data in relation to the healthy state data in the latent space.

It computes the mean of the cosine of the pairwise-angle of the vectors from the healthy state cluster to each degradation data point. The healthy state cluster location is assumed to be the mean of the healthy state data in the latent space. The loss is negated in order to maximize the cosine by minimizing the loss.

The loss is implemented as a torchmetrics.Metric. See their documentation for more information.

Examples:

>>> from rul_adapt.loss.alignment import DegradationDirectionAlignmentLoss
>>> degradation_align = DegradationDirectionAlignmentLoss()
>>> degradation_align(torch.zeros(10, 5), torch.ones(10, 5))
tensor(-1.0000)

DegradationLevelRegularizationLoss

Bases: Metric

This loss is used to regularize the degradation level of the data in the latent space in relation to the healthy state data.

It computes the mean of the difference between the normalized distance of the degradation data points from the healthy state cluster and the normalized degradation steps. The healthy state cluster location is assumed to be the mean of the healthy state data in the latent space.

The loss is implemented as a torchmetrics.Metric. See their documentation for more information.

Examples:

>>> from rul_adapt.loss.alignment import DegradationLevelRegularizationLoss
>>> degradation_align = DegradationLevelRegularizationLoss()
>>> degradation_align(torch.zeros(10, 5),
...                   torch.ones(10, 5),
...                   torch.ones(10),
...                   torch.ones(10, 5),
...                   torch.ones(10))
tensor(0.)

HealthyStateAlignmentLoss

Bases: Metric

This loss is used to align the healthy state of the data in the latent space.

It computes the mean of the variance of each latent feature which is supposed to be minimized. This way a single compact cluster of healthy state data should be formed.

The loss is implemented as a torchmetrics.Metric. See their documentation for more information.

Examples:

>>> from rul_adapt.loss.alignment import HealthyStateAlignmentLoss
>>> healthy_align = HealthyStateAlignmentLoss()
>>> healthy_align(torch.zeros(10, 5))
tensor(0.)