README.md

---
license: mit
---

<center>
  <h1>CANDAS Dataset: A Cooling Fan Sound Dataset with Modeled Disturbances and Controlled Experimental Conditions<br>
    <small>
      <i>Henrique Lima, Cristofer Silva,
      Ricardo Nery, Wilmer Córdoba Camacho, Ricardo Brazileiro, Rodrigo de Paula Monteiro, Andrea Maria Nogueira Cavalcanti Ribeiro, Carmelo José Albanez Bastos-Filho,
      Mariela Cerrada, Diego Pinheiro
      </i>
    </small>
  </h1>
</center>


## 📄 Citation

If you use this dataset in your research, please cite our paper:

LIMA, Henrique et al. **CANDAS Dataset: A Cooling Fan Sound Dataset with Modeled Disturbances and Controlled Experimental Conditions**. In: ENCONTRO NACIONAL DE INTELIGÊNCIA ARTIFICIAL E COMPUTACIONAL (ENIAC), 22., 2025, Fortaleza/CE. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2025. p. 1670-1681. ISSN 2763-9061. DOI: https://doi.org/10.5753/eniac.2025.13864

### BibTeX
```bibtex
@inproceedings{eniac2025candas,
 author = {Henrique Lima and Cristofer Silva and Ricardo Nery and Wilmer Camacho and Ricardo Brazileiro and Rodrigo Monteiro and Andrea Ribeiro and Carmelo Bastos-Filho and Mariela Cerrada and Diego Pinheiro},
 title = {CANDAS Dataset: A Cooling Fan Sound Dataset with Modeled Disturbances and Controlled Experimental Conditions},
 booktitle = {Anais do XXII Encontro Nacional de Inteligência Artificial e Computacional},
 location = {Fortaleza/CE},
 year = {2025},
 pages = {1670--1681},
 publisher = {SBC},
 address = {Porto Alegre, RS, Brasil},
 issn = {2763-9061},
 doi = {10.5753/eniac.2025.13864},
 url = {https://sol.sbc.org.br/index.php/eniac/article/view/38843}
}
```

---

## 📖 Brief Summary

Predictive maintenance reduces economic losses and safety risks in rotating machinery by leveraging vibration and acoustic data, which machine-learning models convert into intelligent fault detectors.
Acoustic signals are particularly effective for early fault detection. Cooling fans, due to their relatively simple rotational dynamics, provide convenient proxies for studying more complex rotor systems.
Despite the growing interest in acoustic-based fault diagnosis, existing cooling fan datasets often lack **modeled disturbance scenarios** and **controlled experimental conditions**.
We present **CANDAS**, a controlled sound dataset containing **28 hours of recordings** from **two cooling fan models** under **five modeled disturbance configurations** and **two voltage levels (9V and 12V)**.

---

## ⚙️ Experimental Setup

The experimental setup was designed to systematically collect acoustic signals from cooling fans under controlled operating conditions. Two commonly used cooling fan models were selected:

<div align="center">

| Fan ID | Model           | Description                          |
|--------|-----------------|--------------------------------------|
| A2     | Delta AUC0912D  | commonly found in electronic devices |
| A3     | Intel AUC0912D  | commonly found in electronic devices |

</div>

For each fan, five weight configurations were applied in order to simulate different levels of mechanical imbalance. Each configuration was evaluated under **two voltage levels**: **9 V** and **12 V**.
This design allows controlled variation in rotational behavior, enabling reproducible disturbance scenarios.

---

## 𖣘 Vibratory Disturbance Model

A model of vibrational disturbances was employed using neodymium magnets as weights, motivated by recent works on the creation of cooling fan datasets with the introduction of weights on the cooling fan blades.
In this model, a magnet is fixed to each blade of the cooling fan using superglue, forming a layer of magnets. This first layer makes it possible for new magnets to be stacked on top of the existing ones, creating new configurations of weight distribution


![Modelo de perturbação](https://cdn-uploads.huggingface.co/production/uploads/689a959b5ec21199bf4cf3fb/VS7bIMXaHfTIIAyBgMiJI.png)
*Figure 1 — The black dots represent first-layer magnets attached to the cooling fan blades, while the blue dots indicate additional magnets stacked on top. It is possible to design both balanced configurations (i.e., Config 1) and severely unbalanced ones (i.e., Config 5).*