Introduction

This repository contains all the materials used in the experiments described in the paper "Explainable fault and severity classification for rolling element bearings using Kolmogorov-Arnold networks".

In the study, we augmented the CWRU and MaFaulDa datasets and subsequently extracted feature libraries from them, in order to apply our Kolmogorov-Arnold network (KAN)-based framework for automatic feature and model selection. The experiments focused on three key tasks: fault detection, fault classification, and severity classification. While centered on bearing faults, our approach also extended to other types of machinery failures via the MaFaulDa dataset.

Files

The extracted feature libraries for each dataset can be found in the data folder. The pykan folder contains some modified files from the pykan library (see below for details). The utils.py file contains some basic utilities used throughout all experiments. The main files are CWRU.ipynb and MaFaulDa.ipynb, containing the structure to carry out the experiments presented in the paper.

Installation

We advise creating a virtual environment, for instance via venv as

python3 -m venv env

Then, activate the environment via env\Scripts\activate for Windows or source env/bin/activate for Linux and install dependencies via

pip3 install -r requirements.txt

Note that packages like Jupyter Lab are not included in the requirements, although the experiments are in .ipynb notebooks, so you will have to install one to run them, e.g. via pip3 install jupyterlab. Additionally, torch is not included in the dependencies because we used the GPU version for acceleration. Depending on your CUDA setup, you are advised to install the corresponding version via their website.

As mentioned above, our code runs on a modified version of the pykan library. Providing an extended changelog would be less practical, so instead we include our versions of the modified files in the pykan folder. You may either install pykan via pip (which is also not included in the dependencies) and then replace the corresponding files, or directly download their codebase locally and replace the corresponding files.

During development, our version of torch was 2.5.0+cu118 and our version of pykan was v0.2.7.

Attribution

If the code and/or extracted feature libraries presented here helped you for your own work, feel free to cite our GitHub repo and/or paper as:

@misc{rigas2024explainablefaultseverityclassification,
      title={Explainable fault and severity classification for rolling element bearings using Kolmogorov-Arnold networks}, 
      author={Spyros Rigas and Michalis Papachristou and Ioannis Sotiropoulos and Georgios Alexandridis},
      year={2024},
      eprint={2412.01322},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2412.01322}, 
}