Work in Progress

This package is currently work in progress. The project is projected to be completed by May 2021. A version of this work was published at the AIAA SciTech 2021 Forum

Precursor Analysis Overview

Identification and analysis of precursors of time-series using the Intelligent Methodology for the Discovery of Precursor of adverse Events (IM-DoPE).

Data processing

cd ../PrecursorAnalysis
python -m imdope.buildDataModel --nominal-directory"Data/toy_data/nominal_events" --adverse-directory "Data/toy_data/adverse_events" --verbose 1 --correlation-thres 0.9 --target-feature "Label"

Expected output of data processing using IM-DoPE:

Flight Length set to 100
Keeping only the following columns: Index(['Param1', 'Param2', 'Param3', 'Param4', 'Param5', 'Param6', 'Param7',
       'Param8', 'Param9', 'Param10', 'filename', 'flight_id', 'Anomaly'],
      dtype='object')
Now Concatenating all flights into a dataframe
Data Model Saved! (path: ../PrecursorAnalysis\Data\MyDataContainer.pkl)
List of features used:
Index(['Param1', 'Param2', 'Param3', 'Param4', 'Param5', 'Param6', 'Param7',
       'Param8', 'Param9', 'Param10', 'filename', 'flight_id', 'Anomaly'],
      dtype='object')
Data Model Saved! (path: ../PrecursorAnalysis\Data\MyDataContainer.pkl)
Data Model Saved! (path: ../PrecursorAnalysis\Data\MyDataContainer_MIL.pkl)
training size: (280, 100, 12)
validation size: (24, 100, 12)
test size: (96, 100, 12)
Data Model Saved! (path: ../PrecursorAnalysis\Data\MyDataContainer_MIL.pkl)

Training a model

The model can be trained directly from the root of the repo using:

cd ../PrecursorAnalysis
python -m imdope.train --model-type "imdope_binary" --lr 0.001 --l2 0.01 --ks 8 5 3 --out-channels 10 15 20 --use-str
atisfy True --model-name "test_model.pt" --epochs 100 --data-container "./Data/MyDataContainer_MIL.pkl" --use-cuda True --mini-batch-percent 0.10

Using pre-trained model

A model that was previously trained can be used for inference using:

cd ../PrecursorAnalysis
python -m imdope.train --load-model True --model-name "test_model.pt" --data-container "./Data/MyDataContainer_MIL.pkl" --use-cuda True

The model is evaluated on the test set, whether it was just trained or a pre-trained model was used, and the results will be printed as follow:

	Precision	Recall	f1-score	support
0	1	1	1	48
1	1	1	1	48
accuracy			1	96
macro avg	1	1	1	96
weighted avg	1	1	1	96

Precursor Discovery

The precursor of a time series with id 5 of interest can be identified:

python -m imdope.precursorIdentification --filename "./Data/MyDataContainer_MIL.pkl" --flight-id 19 --use-cuda True

The precursor will be extracted directly from the model architecture as described in the paper List of outputs:

• Feature ranking
• CNN feature map
• Feature means and standard deviations