Feature Reduction using Deep Learning Methods in Time-Series Data

This project explores different methods to reduce features, evaluate predictive power, and interpret model outputs on CNC machine time-series data.

---

⚙️ Configuration

The file config.py controls key experiment settings:

• data_path → Path to the dataset file (you can switch between datasets by editing this).
• split → Train/validation/test split ratio (default: 0.6, 0.2, 0.2).
• window_size → Sequence length for time-series models.

Update these values in config.py to adapt the experiments to different datasets or configurations.

---

⚙️ Installation

Before running the project, make sure to install all dependencies.
You can install them directly from the requirements.txt file:

pip install -r requirements.txt

---

🔹 Feature Reduction

1. Principal Component Analysis (PCA)

• Applies linear dimensionality reduction.
• Retains 98% variance (configurable).
• Produces:
  • pca_latent.csv containing reduced features (PC1..PCk) + target
  • pca_scree.png showing explained variance across components

Run PCA Feature Reduction

Default (uses paths & dataset split ratio from config.py)

python main.py --model pca_model

---

2. Dense Autoencoder

• Non-linear compression using a fully connected autoencoder.
• Encoder bottleneck (latent layer) provides reduced features.
• Produces:
  • dense_latent.csv containing latent features (latent_0..latent_n) + target
  • Saved models: dense_autoencoder.h5, dense_encoder.h5

Run Autoencoder Feature Reduction

python main.py --model dense_ae

---

3. LSTM-CNN Autoencoder

• Temporal autoencoder combining convolution and recurrent layers.
• Learns latent representations of sequential windows.
• Produces:
  • lstm_cnn_latent.csv with latent features aligned to the target

Run LSTM+CNN Autoencoder Feature Reduction

python main.py --model lstm_cnn

---

🔹 Evaluation of Reduced Features

For each reduced representation, an LSTM Regressor is trained to predict the target variable.

1. PCA Latent Evaluation

• Input: pca_latent.csv
• Uses context window (Window_size = 60, as it produced the best accuracy score in our experimentation) to build sequences
• Produces:
  • Metrics (R², MAE, RMSE)
  • pca_true_vs_pred.png (True vs Predicted plot)

Command for Evaluation

Default (uses paths & dataset split ratio from config.py)

python -m evaluation.lstm_regressor --latent_csv output/pca_latent.csv --model_type pca

---

2. DenseAE Latent Evaluation

• Input: dense_latent.csv
• Uses context window (Window_size = 60)
• Produces:
  • Metrics (R², MAE, RMSE)
  • denseae_true_vs_pred.png

Command for Evaluation

python -m evaluation.lstm_regressor --latent_csv output/dense_latent.csv --model_type dense_ae

---

3. LSTM-CNN Latent Evaluation

• Input: lstm_cnn_latent.csv
• Uses shorter context window (Window_size = 10) as it produced best accuracy score.
• Produces:
  • Metrics (R², MAE, RMSE)
  • lstm_cnn_true_vs_pred.png

Command for Evaluation

python -m evaluation.lstm_regressor --latent_csv output/lstm_cnn_latent.csv --model_type lstm_cnn

---

🔹 Interpretability

1. PCA Loadings

• Calculates loading scores to quantify how much each original feature contributes to each principal component.
• Aggregated importance provides a ranking of features.
• Produces:
  • pca_top30_loadings.png — Top 30 most important features
  • pca_bottom30_loadings.png — Bottom 30 - least important features

Command for PCA Loadings Plot

python -m interpretability.loadings --data_path datasets/DMC2_S_CP2.csv

---

2. SHAP for Dense Autoencoder

• Uses SHAP GradientExplainer on the encoder’s latent space.
• Attributes contributions of original features to learned latent dimensions.
• Produces:
  • denseae_top30.png — Top 30 features shaping the latent space
  • denseae_bottom30.png — Bottom 30 - least important features

Command for SHAP Plot

python -m interpretability.shap_ae --model_path output/dense_autoencoder.h5

---

✅ Summary of Outputs

• Latent CSVs

  • pca_latent.csv, dense_latent.csv, lstm_cnn_latent.csv

• Evaluation Plots

  • pca_true_vs_pred.png, denseae_true_vs_pred.png, lstm_cnn_true_vs_pred.png

• Interpretability Plots

  • pca_top30_loadings.png, pca_bottom30_loadings.png
  • denseae_top30.png, denseae_bottom30.png

---

📊 Results

• Reduced original feature set from 52 → ~25 latent features using PCA and Autoencoder Architectures.
• Using the reduced features, Supervised LSTM-CNN achieved R² ≈ 0.8, exceeding the original full-feature performance (≈ 0.7).

✨ Significance

• Identified the most important sensor features, enabling prioritized monitoring and maintenance in CNC manufacturing.
• Highlighted redundant and less useful sensors, allowing data collection and storage costs to be reduced.
• Improved model efficiency: less input data → faster training/inference, lower memory and computational demands.
• Enhanced interpretability and trust in model decisions, which is critical for industrial applications and deployment in sensor systems.

---

👩‍💻 Author

Shweta Bambal
Research Assistant (DS/ML) at OVGU, Magdeburg

📧 Email: shwetabambal18@gmail.com

🔗 LinkedIn | GitHub

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

---