Ensemble Methods for APS In-Flight Particle Temperature and Velocity Prediction Considering Torch Electrodes Ageing

Area of Science:

• Materials Science
• Manufacturing Engineering
• Data Science

Background:

• The atmospheric plasma spray (APS) process involves complex nonlinear relationships between input parameters and in-flight particle characteristics, crucial for coating quality.
• Torch electrode aging significantly impacts these relationships, adding another layer of complexity to process control and prediction.
• Machine learning (ML) offers a powerful approach to model and manage these intricate, nonlinear interactions in industrial processes.

Purpose of the Study:

• To apply ensemble ML methods for predicting in-flight particle temperature and velocity in APS, considering the effect of torch electrode aging.
• To compare the effectiveness of Random Forest (RF) and Gradient Boosting (GB) for feature selection and predictive modeling in this context.
• To investigate the utility of time series analysis techniques for improving the prediction accuracy of APS in-flight particle characteristics.

Main Methods:

• Collected experimental data on APS input parameters, in-flight particle characteristics, and electrode usage time.
• Employed Random Forest (RF) and Gradient Boosting (GB) for feature ranking, selection, and predictive model development.
• Explored two time series embedding strategies: direct attribute/target embedding and embedding after differencing for stationarity.

Main Results:

• RF demonstrated superior performance over GB, capable of simultaneously predicting both particle velocity and temperature while capturing inter-target interactions.
• Both RF and GB models showed improved performance when APS data was treated as a time series, highlighting the importance of temporal dependencies.
• Pre-processing time series data using differencing to achieve stationarity further enhanced the predictive accuracy of the RF models.

Conclusions:

• Ensemble ML methods, particularly Random Forest, are effective for predicting in-flight particle characteristics in APS, even with electrode aging.
• Treating APS data as a time series, especially after applying differencing for stationarity, significantly improves prediction accuracy.
• The findings suggest that RF is a more suitable algorithm than GB for simultaneously predicting multiple, interacting in-flight particle characteristics in APS.