paper-02-A_multi-sensor_architecture

A multi-sensor architecture for robust identification of incipient short-circuits in wind turbine generators

The source code necessary to run all experiments discussed in the paper.

Repo structure

Clone the full repo and pay attention for the if the data folder follows the structure

.
├── data
│   ├── csv             - csv concataned and chunks
├── models              - models, transformers and pippelines
├── notebooks           - notebooks for data analyses
├── results             - results folder
└── src                 - source code

Data folder

Inside the csv folder there are the sensor settings datasets in a csv format ready to be fed into the model's pipeline. The naming convention refer to parameters, for example, the file v000_SCIG_SC_SENSORC_FOURIER.csv tell us,

• v000: version of the database (concerns only to author).
• SCIG: electrical machine is a Squirrel-Cage Induction Generator (SCIG).
• SC: The inserted fault is short-circuit (SC)
• SENSORC: The sensor used to monitor the machine.
• FOURIER: The feature extracted method used to build the dataset.

All datasets used in the paper are inside this folder. More details about theses dataset are here.

Models folder

It contains all instaces used in the simulations. The objects are serialized in a pickle format using the extension .pkl. Unpackle it using the proper method. The libraty joblib is recommended. This folder is organized as follows

• classifier: contain the trainned classifiers. They follow a straightfoward naming convention. For example, clf_SENSORC_FOURIER_mlp_union.pkl stands for the MLP classifiers, using the feature union process, trained on the dataset built using the feature extraction method Fourier and the raw signal is a three-phases current sensor.
• pipeline: contains the transformer that ought to be used to prepare inputs for the classifiers. The process is decribed in paper's Section II.A. The pipeline is a composition of scalers and selector.
• scaler: The feature scaler used to standandize features.
• selector: The feature selector to be applied in the datasets. The naming SF stands for set of feature and the labeling 1 or 2 refers to the set of features and the extra features, respectivelly. Describe in paper's Section II.B.

How to run

It is very important to maintain this repository structure. Instabilities may occur if files are moved without caution. Open a question if ran into any problem.

Reproductible results

1. Navigate to src and setup the environment:

cd src
virtualenv -p python3.6 .venv
source .venv/bin/activate
pip install -r requirements.txt

You can uso pipenv instead of virtualenv, there is a Pipfile inside the folder.

2. Run the following scripts to reproduce our results

• python test_clf_appended.py and python test_clf_union.py. The results are saved under the folder results.

See the src directory for more details about the scripts.using another scritps. There is a lot more that can be done. And the results folder for more details how the results are organized.

Statistical tests

The statistical tests are made under the notebook dir. To reproducetheses results open the notebook-02-statistical_test.ipynb. The discussion and methods are pretty straightforward.

To run this parte ot is not necessary to run the aforementioned scripts. The results are store in .pklfiles here.

More info:

Navigate to each folder for more information. There should be a readme file explain in details all proccess. Fell free to contact us, open questions, report bug or PR. Please remember to cite us :)