RL for Predictive Maintenance

Implementation Notes

Working PHM settings

1. ALGO = 'PPO'
2. EARLY_DETECT_FACTOR = -0.125
3. r1 = 1; r2 = -4; r3 = -0.5
4. SAMPLING_RATE = 25
5. EPISODES = 200 k  # ** For C06 Most repeatable** runs.
6. ADD_NOISE = 5*1e2 = 500
7. MAX_EPISODE_STEPS_FACTOR = 10 
8. BATCH_SIZE = 16

Working NUAA settings

1. EARLY_DETECT_FACTOR = -0.125
2. r1 = 2; r2 = -4; r3 = -0.5
3. Tried for NUAA W1
4. ADD_NOISE = 5*1e2
5. SAMPLING_RATE = 25
6. EPISODES = 200_000

Most repeatable implementation: PHM_C06, with settings above

What we have

1. Trained PdM agent: "Agent_PHM_C01" implies trained on C01
2. THREE training run results and
3. THREE trained agent
4. TensorBoard plots
5. Trainining results as saved images and .csv results

What we can demonstrate

1. Three runs -- so show REPEATABILITY
2. Trained PdM agent -- so show ROBUSTNESS or TRANSFERABILITY by testing on another set

Presentation of Results

1. Trained model agents: C01
2. Show C01 Tool wear data - normal
3. Show with noise - Mention for robustness
4. Evaluate on C04 and C06
5. Refresh untill reasonable REPLACEMENT
6. Attempt an evaluation on NUAA W1
7. Repeat with C04 on rest i.e. C01 and C06 etc.

Visualizations:

9. Tool wear plot normal and with noise
10. TRANSFERABILITY: Tensorboard reward learning multiple curves (C01, C04, C06) - self explainable
11. Tool replacement time reduction - self explainable
12. RUL improvement - will need explaining so show last