1. Hyperparameter Selection
This project implements an autoencoder neural network dedicated to anomaly detection on the Shuttle dataset. The model learns to reconstruct only normal data in order to identify deviations during the test phase.
Parameter selection is based on a statistical analysis of the data to ensure optimal compression.
Selected Parameters
- Latent Space (k = 6): Based on eigenvalue analysis — 6 significant values are observed
- Learning Rate: 0.0015 — ensures smooth loss stabilization
- Optimizer: Adam — efficient convergence
- Epochs: 80 — convergence plateau reached around epoch 76
Eigenvalue Analysis
The choice of k=6 for the latent space is justified by the following spectral analysis, where 6 eigenvalues clearly stand out:
λ₄ = 6.2564 × 10⁻¹ | λ₅ = 9.8503 × 10⁻¹ | λ₆ = 1.0004
λ₇ = 1.0245 | λ₈ = 1.7436 | λ₉ = 3.6172
Technology Stack
Figure 1: MSE loss evolution (K=6) showing model stabilization.
2. Network Architecture
The network adopts a symmetric "hourglass" structure to extract the essential features from the signals.
Architecture Specifications
- Input/Output dimensions: 1×9 → 1×9
- Encoder: Three progressive linear layers (9 → 8 → 7 → 6)
- Activations: ReLU after each linear layer
- Latent Space: Compressed dimension of 6
- Decoder: Mirror structure of the encoder (6 → 7 → 8 → 9)
- Final output: No activation — faithful reconstruction of real values
ReLU layers after each linear layer introduce the non-linearity needed to capture complex relationships in the data. The absence of output activation allows unconstrained reconstruction of continuous values.
Figure 2: Encoder-Decoder network architecture diagram.
3. Training Strategy
Training is optimized to process normal data only, effectively ignoring anomalies.
Training Parameters
- Batch Size: 256 — maximizes computation speed while avoiding excessive generalization
- Training: Exclusively on normal class data (class 1)
- Loss Function: Mean Squared Error (MSE)
Preprocessing & Normalization
- Preprocessing: Removal of invalid data and label separation to keep only the 9 sensor dimensions
- Normalization: Computed only on valid training data (class 1)
- Test set: Normalized using the training set's mean and standard deviation to prevent anomalies from biasing the scale
4. Optimal Threshold Determination
Final classification relies on computing the reconstruction error (MSE). A threshold is determined to separate "normal" from "anomalous".
Determination Method
- Objective: Find the ideal trade-off minimizing False Negatives (FN) and False Positives (FP)
- Visual analysis: Loss histograms to identify the separation between the two distributions
- Evaluation: Validated by the F-measure (F1-score) curve as a function of the threshold
F-Measure formula used for evaluation.
Reconstruction Loss Histograms
Full distribution of reconstruction errors — normal data vs anomalies.
Figure 4: Zoomed view — separation between normal and anomalous distributions.
F-Measure & Accuracy Curve
Figure 5: F-measure and Accuracy evolution as a function of the selected threshold.
5. Results
Final performance demonstrates the model's high precision on this dataset.
Final anomaly detection results on the Shuttle dataset.
6. Conclusion
Project Summary
This project demonstrates that a lightweight autoencoder with a latent space of dimension 6 is sufficient to effectively detect anomalies in the Shuttle dataset. The reconstruction approach — training only on normal data then thresholding the MSE error — proves to be robust and interpretable.
Key Highlights
- F-score of 0.9912 — near-perfect performance on this dataset
- Accuracy of 98.62% — excellent generalization
- Minimalist symmetric architecture — 9 → 6 → 9 with ReLU
- Stable convergence reached by epoch 76 out of 80
- Optimal threshold determined by F-measure analysis, minimizing FP and FN
Future Improvements
- Variational Autoencoder (VAE): Probabilistic latent space modeling for better generalization
- Automatic threshold search: Threshold optimization via cross-validation
- Other datasets: Test robustness on KDD Cup, MNIST-Anomaly, etc.
- LSTM Autoencoder: Leverage the temporal dimension of sensor data
Resources & Links
View Source Code on GitHubAssociated Documents
- Dataset: UCI Shuttle Dataset — 9 sensor features, normal/anomaly classes
- Source Code: Model, training and evaluation scripts on GitHub