VAE Deviation for Detecting Bearing Anomalies

Yukio Hiranaka, Koichi Tsujino

Abstract

Anomaly of rotating machines are usually inferred from vibration measurements. However, it is not easy to determine the normal range for conventional crest factor or primary component analysis. In this paper, we try to use the Artificial Neural Network technique to make judgments based on the degree of deviation from the learned normal range. Specifically, we evaluated VAE which compresses the measured sensor data into the latent space of smaller number of dimensions with standard normal distributions. We propose an anomaly score which indicates the deviation from the center of the normal distribution using linear VAE calculation and dimensionality compensation. The proposed anomaly score shows good performance with several test data sets and measured real data sets.

Keywords

Variational Auto-Encoder, Bearing Anomaly Detection, Anomaly Score, Latent Space

Download

IMEKO-TC10-2020-024.pdf

DOI

-

IMEKO TC

TC10 - Measurement for Diagnostics, Optimization and Control

Back to the proceedings