Accomplishments
- Data Collection: Completed power-fault data collection
- Started reduced flow data collection loop
- Regression model: Finalized and validated model for predicting CPU power
- Autoencoder anomaly detection model: Implemented AE using RF residuals
- Completed initial tuning: window size, latent dimension, dropout, batch size
- Implementing event based fault detection instead of window based to fix poor separation between normal vs. fault cases
PCA After Tuning 
PCA Before Tuning
Significant Risks
- Normal fault overlap in AE latent space: PCA shows fault and normal windows are not cleanly separable even after tuning
- Window level anomaly detection may be unreliable due to overlap
- Risk mitigation: Switching to event-based detection, which only requires detecting a fault at least once per event
Design Changes
- No major design changes this week
Schedule Changes
- Project remains on track
- ML model tuning in progress