Risk-Aware Calibration Scheduling

Risk-aware calibration scheduling on CMAPSS-adapted turbofan data, using a Transformer with an integrated quantile regression head to produce safety-aware time-to-drift predictions and calibration priorities.

Overview

• Adapt CMAPSS into a calibration setting with virtual thresholds, splice/stitch resets, and time-to-drift labels.
• Train sequence models (Transformer with quantile head, LSTM, CNN, TCN) and baselines (trees/boosting).
• Use quantile-triggered, risk-aware scheduling policies to balance violations and calibration cost.
• Generate plots, tables, and summaries for each CMAPSS subset (FD001–FD004).

Key Components

• Calibration adaptation: Virtual thresholds per drift sensor, synthetic calibration resets, sawtooth TTD labels.
• Models: Quantile Transformer (pinball loss), LSTM with quantile head, CNN/TCN, tree/boosting baselines.
• Scheduling: Predictive policies trigger when lower-quantile TTD indicates violation risk; cost model supports calibration vs. violation trade-offs.
• Outputs: Metrics, policy costs, and plots per dataset.

Running the pipeline

python3 calibration_scheduler.py

By default, this will process FD001–FD004 in sequence. Adjust configuration in Config (dataset selection, margins, costs, quantiles).

Citation

If you use this codebase, please cite the accompanying paper or reference this repository:

https://github.com/adithyap/risk-aware-calibration-scheduling