Time-series Defect Detection & Explainability

---

Project overview ✅

This project demonstrates a complete pipeline for multi-label defect detection on synthetic multivariate time-series data and provides per-defect explainability. It uses an LSTM-based classifier for detection and an LSTM autoencoder to identify and localize anomalous segments responsible for defects.

Key outcomes:

• High classification performance (binarized accuracy reported ~99.832%)
• Per-anomaly explainability that maps individual anomaly clusters to defect labels

---

Files

• ProjectData.ipynb — Main notebook with data generation, model training/evaluation, and visualizations.
• RNN_model.pt — Saved classifier state (used by notebook when retrain=False).
• lstm_autoencoder.pt — Saved autoencoder state (used by notebook when retrain_autoencoder=False).

---

Dataset & generation 💡

• Data is synthetic and generated in the notebook via the createRow function.
• Each sample is a multivariate time series (3 features), with 5 potential defect types injected probabilistically.
• Dataset size used for experiments: n = 50000 (adjustable in notebook).

---

Models & approach 🔧

• Classifier: RNNClassifier — two LSTM layers (64 → 32) + fully connected output with sigmoid nodes for multi-label classification. Optimized with Adam and BCELoss.
• Explainability: LSTMAutoencoder trained only on no-defect samples. Reconstruction error (MSE) is used to detect anomalies; thresholding and clustering isolate anomaly segments. The classifier is then used on series with single isolated anomaly clusters to determine which defect they represent.

---

How to run (quick start) ▶️

1. Open ProjectData.ipynb in Jupyter or VS Code.
2. Install dependencies (example):

python -m pip install torch numpy pandas matplotlib scikit-learn pillow

3. Run cells from top to bottom. If you want to retrain models:

• Set retrain = True to train the classifier.
• Set retrain_autoencoder = True to train the autoencoder (autoencoder uses many epochs to overfit no-defect data).

Notes:

• If you prefer to use the pretrained models, keep retrain=False and ensure RNN_model.pt and lstm_autoencoder.pt are in the notebook directory.
• The notebook automatically handles padding variable-length sequences using pad_sequence.

---

Important notebook functions & variables

• createRow(...) — synthetic generator of a single time series
• RNNClassifier — classifier model
• LSTMAutoencoder — explainability autoencoder
• visualize_candy_defects(k, n) — visualize examples of defect k
• visualize_many(idx) — visualize series with mapped defect labels
• threshold — reconstruction error threshold (default used: 0.19)

---

Results & evaluation 📊

• Binarize predictions with threshold 0.5 to compute classification metrics.
• Example reported: Binarized Accuracy ≈ 0.99832. Per-class precision & recall printed in the notebook.

---

Reproducibility & tips

• Set seeds (np.random.seed(42) is used) for deterministic generation in examples.
• If you train on GPU make sure torch.cuda.is_available() and set device accordingly (handled in notebook).
• Adjust threshold if you change the autoencoder architecture or dataset.

---

Possible extensions

• Replace synthetic generator with real sensor data and retrain.
• Use attention mechanisms for more detailed temporal explainability.
• Automate threshold selection (ROC/validation-based) for anomaly detection.