LSTM VS. Feed-Forward Autoencoders for Unsupervised Fault Detection in Hydraulic Pumps

TL;DR

This paper compares LSTM and feed-forward autoencoders for unsupervised fault detection in hydraulic pumps, demonstrating high reliability without using fault data during training.

Contribution

It introduces and evaluates two autoencoder schemes for early fault detection, highlighting the effectiveness of LSTM and feed-forward models trained solely on healthy data.

Findings

Both models achieve high reliability in fault detection.

LSTM captures temporal dependencies effectively.

Feed-forward model analyzes individual sensor snapshots.

Abstract

Unplanned failures in industrial hydraulic pumps can halt production and incur substantial costs. We explore two unsupervised autoencoder (AE) schemes for early fault detection: a feed-forward model that analyses individual sensor snapshots and a Long Short-Term Memory (LSTM) model that captures short temporal windows. Both networks are trained only on healthy data drawn from a minute-level log of 52 sensor channels; evaluation uses a separate set that contains seven annotated fault intervals. Despite the absence of fault samples during training, the models achieve high reliability.