From Dataset to Deployment: A Full-Stack Autoencoder-Based Anomaly Detection Pipeline for Gas Turbine BLDC Fuel Pumps

Reliable health monitoring of fuel pumps is essential for maintaining the stability and safety of gas turbine engines. Brushless DC (BLDC) fuel pumps operate under tightly coupled electrical–mechanical–hydraulic dynamics, where fault progression is often subtle and real fault data are scarce. This article presents a full-stack anomaly detection pipeline spanning dataset construction, feature engineering, model development, evaluation, and embedded deployment, built on an unsupervised autoencoder trained exclusively on normal operating data. A real-world dataset was collected from BLDC fuel pumps integrated into gas turbine testbeds over multiple engines and years; however, the available corpus remains modest, comprising 18 log files and nearly three hours of cumulative operation. A 48-dimensional handcrafted feature set was designed to capture statistical, energetic, derivative, and autocorrelation characteristics of pump behavior. The autoencoder learns a compact representation of healthy operation, while anomalies are identified through reconstruction error using a fixed threshold and a persistence-based decision rule. In offline analysis, the proposed autoencoder achieves competitive discrimination relative to strong unsupervised baselines, while its primary practical advantage lies in supporting a deterministic end-to-end TinyML deployment pipeline rather than maximizing offline discriminative performance alone. Real-time on-device experiments on an STM32 platform demonstrate successful detection for the evaluated dry-run and aeration scenarios, while no alarms are produced under normal steady and transient tests after persistence filtering. Under the current feature set and calibration, the evaluated partial-blockage and gradual-restriction scenarios did not produce persistence-triggered alarms, indicating incomplete fault-mode coverage and motivating future work on broader fault characterization and adaptive calibration. These results support the practical feasibility of real-time, on-device anomaly detection for gas-turbine fuel-pump monitoring in a resource-constrained TinyML setting.