Reinforcement Learning for Predictive Maintenance

Important notes for Thesis and Presentation

• Only raw features. NO derived features like Dr Sameer Sayyad.
• Proof of generalization:
  • Fundamentally different datasets - material, sensors etc.
  • Features are different

Important implementation notes

• Rewards for NUAA and PHM similar range: Keep number of records similar - bout 1000
• RUL value threshold should have some value and not 0

Approach: PhD Thesis work

V.1.0 11-Oct-2024:

• Create two environments: PHM and NUAA

• To demonstrate robustness - Use multiple data sets from each 3 + 3 = 6

• Additional: Add noise and Break-down

• Publish results

• NUAA

Notes:

Environment:

• Only raw features. NO derived features like Dr Sameer Sayyad.

    high = np.array(  [
        1.0,          # Max. force_x
        1.0,          # Max. force_y
        1.0,          # Max. force_z
        1.0,          # Max. vibration_x
        1.0,          # Max. vibration_y
        1.0,          # Max. vibration_z                
    ], dtype=np.float32,) 

PHM	Uniwear
time	time
force_x	axial_force
force_y
force_z	force_z
vibration_x	vibration_x
vibration_y	vibration_y
vibration_z	vibration1
	vibration2
tool_wear	tool_wear

Dataset	Workpiece	Tool
PHM2010	Stainless steel (HRC52)	Tungsten Carbide
NUAA	Titanium (TC4)	Solid Carbide