← All projects

Autoencoder Anomaly Detection

Unsupervised anomaly detection for time-series data based on reconstruction error.

Problem
Detect abnormal machine behaviour when labelled fault data is rare.
Data
CWRU bearing vibration dataset with normal and faulty bearings, processed on Hugging Face.
Approach
Models are trained on normal data only. Convolutional, WaveNet-based and attention-based autoencoders are compared, with several loss functions in the time or frequency domain.
Tools
Python, PyTorch, Lightning, Librosa, Hugging Face Datasets
Publication
Unsupervised anomaly detection of permanent-magnet offshore wind generators through electrical and electromagnetic measurements. Wind Energy Science, 2024.
Data processing with overlapping training samples in the time and frequency domain, and the convolutional autoencoder architecture trained on reconstruction error
Overview of data processing and the convolutional autoencoder (CAE) architecture implemented in the paper.