nasa / prog_algs Goto Github PK

The Prognostic Algorithm Package is a python framework for model-based prognostics (computation of remaining useful life) of engineering systems, and provides a set of algorithms for state estimation and prediction, including uncertainty propagation. The algorithms take as inputs prognostic models (from NASA's Prognostics Model Package), and perform estimation and prediction functions. The library allows the rapid development of prognostics solutions for given models of components and systems. Different algorithms can be easily swapped to do comparative studies and evaluations of different algorithms to select the best for the application at hand.

integrated-system-health-management prognostics prognostics-health-management

prog_algs's Introduction

ProgPy Packages

The prog_algs package has been combined with the prog_models package to create the new progpy python package, here: https://github.com/nasa/progpy.

prog_algs's People

Contributors

Stargazers

Watchers

prog_algs's Issues

RNN based model tuning

State Scatter Plots

Add support for python 3.10

Particle Filter- Add ability to set samples directly

Dynamic Step Size in the MC

Online System Id/Parameter Estimation

Approach for online system id by using a "training period" with a separate model. Assumes health is 1, and trains for model parameters

Finish MC Predictor Test

Add description at top of each example file

A couple lines describing the purpose of each example would be helpful.

Definition of Done:

Description of example purpose and overview at top of each file in triple quotes (""" [Description] """) below copyright line

Fix: MC predict returns times[0]

UKF-UTP Connection

MonteCarlo EOL Return

Event Time should return UncertainData() - look at multiple events.

Needed to support hist

MC - Predict any event

Accessing output eqn in pool

Output eqn access in pool is broken in unreleased version of prog models

Make ground truth optional in metrics

Add Prognostics Horizon Metric

Add metric for prognostics horizon i.e., the prediction time at with the prediction first comes with a specified accuracy (alpha-beta)

D.o.D.

Implement prognostics horizon metric for prediction profile
Implement tests
Add to examples
Add to release notes

Prediction return type

Explore different return types for prediction. I'm thinking some sort of class that behaves the same as the current lists for backwards compatibility, but also provides other functions. Could also integrate metrics into return type

Update particle filter to use measurement_noise from model

Measurement Example- extension

Add an example using the measurement_eqn to handle a case where units are different, or what is measured is some combination of states

Explore methods for EOL estimation in UTPredictor which doesn't rely on the gaussian assumption

The current approach includes an assumption that the sigma points that represent the distribution of states are valid sigma points on the time axis (i.e., the time when the state sigma points cross the threshold form the sigma points of the new EOL distribution). This is based on the strong assumption that the EOL distribution is similar to a gaussian distribution and is consistent to the state distribution.

This may be true for simple state transition or for a simple threshold shape, but may not be true for complex arbitrary threshold functions and state transition equations

@matteocorbetta had an idea for addressing this. Since we know the state to be gaussian, can we at each tilmestep, estimate the portion of the state distribution that is beyond the threshold, recording those numbers instead of just the sigma points.

@teubert suggested that we use @kjjarvis sample gain/shed approach to strategically add extra samples around the sigma points closest to the threshold in order to better define the shape of the distribution as it passes the threshold. The samples would be shed as they pass the threshold, being replaced by others further away.

Add Median to UncertainData

Under consideration: add median as another metric

Fix benchmarking example

Form UnweightedSamples from dict

PF: Initial uncertainty as covar

Make Tests independently executable

add support for covar matrix in particle filter x0_uncertainty

Prediction Plot

Add plot that shows entire prediction with uncertainty. Something like this:

Make Examples independently runnable

Unscented Kalman Predictor: specify event(s) to predict to

Specify subset of events to predict to. When every specified event has been met, simulation will end.

Figure showing examples of new figures

Add to release notes and documentation

Uncertain Data Files

All of Uncertain Data is in init.py

DoD:

Split Uncertain Data Classes into separate files
Setup init.py to automatically import them

Bug

Previous implementation was fundamentally broken. multithreading requires every argument to be fundamentally pickle-able. However, the prediction function is by its nature not pickle able.

Considered Approaches

Set the function attributes to get around this, but this is done globally and breaks down in instances where there are multiple MC's running at once (e.g., PaaS). - this was my previous approach
Pass argument of marshaled code of function (see https://stackoverflow.com/questions/1253528/is-there-an-easy-way-to-pickle-a-python-function-or-otherwise-serialize-its-code). This had promise but didn't work with x being an optional argument. I would have to remove that feature for this to work.
Dill package extends pickle. I might be able to pass in a dill pickled (see extension dill)
Cannibalize simulate_to_threshold - parallelize all but the simulate_to_threshold - this has significant duplicated code