Week 7 Summary: Time-series and process monitoring

Cross-Book Summary

1. Sequential Data & Memory

• Temporal Dimension: Process logs are sequential; order matters.
• RNNs: Possess hidden states that act as memory.
• Vanishing Gradients: Basic RNNs struggle with long-term dependencies.
• LSTMs/GRUs: Use gates to selectively remember/forget information.

2. Materials Process Monitoring

• Preprocessing: Smoothing and “Triggering” to extract cycles.
• Anomaly Detection: Large prediction deviations signal defects.
• Surrogate Models: Fast RNN/LSTM replacements for slow simulations.

90-Minute Lecture Strategy

Part 1: Processing as Sequence

• Time-dependency of microstructure.
• Sensor types (1D vs. logs).
• Event vs. Continuous.

Part 2: Preprocessing

• Denoising filters.
• Triggering cycles.
• Temporal Feature Engineering.

Part 3: RNNs

• Unrolled RNN structure.
• Vanishing Gradient problem.
• LSTM and GRU mechanics.

Part 4: Case Studies

• Predictive Maintenance.
• AM melt pool stability.
• Dilatometry phase predictions.

Part 5: Challenges

• Non-stationarity and machine drift.
• Transformers for 1D data.
• Closed-loop control outlook.

Quarto Website Update (Summary)

Summary for ML-PC Week 7:
- Applies ML to Time-Series Data for process monitoring. - Introduces RNNs and LSTMs for sequential dependencies. - Details essential preprocessing like smoothing and triggering. - Covers anomaly detection and process outcome prediction.