Autoencoder neural network learns "normal" and flags unusual patterns
How This Works
The Autoencoder Approach:
1. Train on normal flights - The network learns to compress and reconstruct typical engine data
2. Feed it new data - It tries to reconstruct what it sees
3. Compare input vs reconstruction - If they're very different, the data doesn't match learned "normal" patterns
4. Flag anomalies - High reconstruction error = "I haven't seen patterns like this before"
No one labels "bad" data. It just learns what "normal" looks like and flags anything that doesn't fit.
① Train the Model (on normal flight data)
0
Training Samples
-
Reconstruction Loss
Untrained
Model Status
Training...
② Test Different Scenarios
③ Current Engine Reading
Cylinder 1
EGT
--
CHT
--
Cylinder 2
EGT
--
CHT
--
Cylinder 3
EGT
--
CHT
--
Cylinder 4
EGT
--
CHT
--
④ Anomaly Detection Result
Reconstruction Error (Anomaly Score)
-
0 (Normal)0.15 (Threshold)0.3+ (Anomaly)
⑤ What the Network "Expected" vs What It Saw
The autoencoder reconstructs what it thinks the data "should" look like based on patterns it learned.
Large differences (highlighted) indicate the input doesn't match normal patterns.
Actual Input
Network's Reconstruction
⑥ Detection Log
Train the model, then test scenarios to see results here