Turbofan Engine Anomaly Detection

Unsupervised anomaly detection on multivariate sensor data from turbofan engines using a dense Autoencoder built in PyTorch. The model is trained exclusively on early-life (healthy) engine data and uses reconstruction error to flag degradation. No labeled anomalies required.

Key result: Average early warning lead time of 131.8 cycles before failure across 100 engines (FD001).

Dataset

NASA CMAPSS Turbofan Engine Degradation Simulation Dataset (FD001–FD004)

Approach

• Dropped low-variance sensors
• Normalized with MinMaxScaler
• Defined normal as first 30 cycles of each engine
• Trained dense Autoencoder on normal data only
• Flagged anomalies where reconstruction error exceeds 95th percentile threshold

Stack

Python, PyTorch, pandas, scikit-learn, matplotlib