Autoencoder-Based Anomaly Detection for Turbofan Engine Sensors Data

Turbofan engines generate large amounts of multivariate sensor data whose unlabeled nature complicates direct anomaly identification. This study utilizes the NASA C-MAPSS FD001 turbofan engine degradation dataset, focusing on 21 sensor readings while excluding non-sensor attributes, to develop an unsupervised anomaly detection approach. An autoencoder neural network with a 15-8-15 architecture is trained on 20,631 normalized training records using Mean Squared Error (MSE) as a reconstruction loss metric. Anomalies are defined using the 95th percentile threshold of the MSE distribution computed from the training data. When evaluated on 13,096 test records, the model achieves an average MSE of 0.001823 with a standard deviation of 0.001032 and flags 131 points (about 1%) as anomalous using a threshold of 0.003785. The low reconstruction error for most data indicates that the model captures normal behavior patterns effectively, while high-error instances correspond to anomalous sensor behavior. These results demonstrate that autoencoder-based reconstruction error is an effective method for detecting anomalies in unlabeled turbofan engine sensor data.