Aero-engines Anomaly Detection using an Unsupervised Fisher Autoencoder
Saba Sanami, Amir G. Aghdam
TL;DR
This paper addresses unsupervised anomaly detection in aero-engines where labeled faults are scarce. It introduces a Fisher Autoencoder (FAE) with a learnable Gaussian mixture prior for the latent space and trains it by minimizing the Fisher divergence between the true and modeled joint distributions $D_F[ q_{true,\phi}(X,z) | p_{\eta,\theta}(X,z) ]$. The learning objective decomposes into a Fisher-divergence term, a reconstruction term, and a stability regularization term, implemented via the reparameterization trick $z^{(l)} = \mu(X) + \sigma(X) \odot \epsilon^{(l)}$. Experiments on the CMAPSS dataset demonstrate timely anomaly detection with reduced false alarms and a structured latent space that supports diagnostics. The results suggest robustness to imbalanced data and better generalization across engine models, with future work on uncertainty quantification for risk assessment.
Abstract
Reliable aero-engine anomaly detection is crucial for ensuring aircraft safety and operational efficiency. This research explores the application of the Fisher autoencoder as an unsupervised deep learning method for detecting anomalies in aero-engine multivariate sensor data, using a Gaussian mixture as the prior distribution of the latent space. The proposed method aims to minimize the Fisher divergence between the true and the modeled data distribution in order to train an autoencoder that can capture the normal patterns of aero-engine behavior. The Fisher divergence is robust to model uncertainty, meaning it can handle noisy or incomplete data. The Fisher autoencoder also has well-defined latent space regions, which makes it more generalizable and regularized for various types of aero-engines as well as facilitates diagnostic purposes. The proposed approach improves the accuracy of anomaly detection and reduces false alarms. Simulations using the CMAPSS dataset demonstrate the model's efficacy in achieving timely anomaly detection, even in the case of an unbalanced dataset.