Remaining Useful Life Prediction from 3D Scan Data with Genetically Optimized Convolutional Neural Networks.

Area of Science:

• Industrial Engineering
• Machine Learning
• Materials Science

Background:

• Optimized asset management is crucial in modern industry.
• Prognostics and Health Management (PHM) enhances maintenance decisions through continuous monitoring and Remaining Useful Life (RUL) forecasting.

Purpose of the Study:

• To investigate convolutional neural network (CNN)-based deep learning for RUL prediction of punch tools.
• To analyze surface deformations during machining using 3D scanning.

Main Methods:

• 3D scanning sensors captured point clouds of punch tool surface deformations with micrometric accuracy.
• 3D point clouds were converted into 2D depth and normal vector maps for CNN feature extraction.
• 15 genetically optimized CNN architectures were compared against 19 pretrained transfer learning models and support vector regression (SVR).

Main Results:

• Genetically optimized CNN architectures achieved superior RUL prediction performance (MAPE = 0.058).
• Transfer learning models showed moderate performance (MAPE = 0.416).
• Support vector regression served as a benchmark with lower performance (MAPE = 0.857).

Conclusions:

• Genetically optimized CNNs are highly effective for punch tool RUL prediction.
• This PHM strategy enables accurate, non-invasive condition monitoring and predictive maintenance.