국문

<aside> 📌 Contents

</aside>

<aside> 📌 Short-cut

1st KSNVE 2024 AI Challenge

</aside>

This document is about the baseline system of the 1st Noise and Vibration AI Challenge by the Korean Society for Noise and Vibration Engineering.

Baseline files

KSNVE_baseline.zip

Description

• The purpose of this Noise and Vibration AI Challenge is to detect abnormal signals using vibration data under variable speed conditions.

<aside> 📌 The baseline system consists of a total of 4 python codes.

• preprocess.py: Data preprocessing code
• train.py: Model training code
• eval.py: Code to extract the performance per defect of the trained model
• test.py: Code to extract the anomaly score for the evaluation data </aside>

Baseline system

• The baseline system utilizes a spectrogram-based AutoEncoder model.
• AutoEncoder learns features through the process of compressing and restoring data, and it is expected that abnormal data will fail to restore data, and anomaly diagnosis is performed.
• The input spectrogram is $X = \{ X_t \}{t=1}^T$, and $X_t \in \mathbb{R}^{F}$, where $T$ and $F$ are the numbers of time frames and frequency bins of the spectrogram, respectively. When the AutoEncoder model is $\psi{\phi}$, the anomaly score for each data is as follows.

$$ \text{Score}(X) = \frac{1}{TF} \sum_{t=1}^T \|X_t - \psi_{\phi}(X_t) \|_2^2 $$

Dataset