Evaluation of ML potential for modelling and process control of a reduction annealing process

Sammanfattning

While many industrial processes have been thoroughly exposed to machine learning models, the realm of powder metallurgy is still underexplored. This thesis aims to investigate the potential for a machine learning model to predict the characteristics of metal powders after a reduction annealing process using time series process data from the furnace. A supervised learning model was developed to predict six target variables: three particle size distributions, two chemical composition properties, and a density. The pipeline involved filtering and preprocessing data, model training, and performance evaluation. Multiple families of machine learning methods were explored and implemented, including linear methods, tree-based ensemble methods, and sequence models. Hyperparameter values were systematically tuned from a range of values to find the best configuration; each model’s performance was then assessed using cross- validated error metrics. Using SHAP analysis, it was possible to identify the most prevalent and important features for predicting each output for each model. Results show that performance varies for each target variable, and while no model was completely dominant, the most successful predictions were made by LSTM, XGboost, and the Elastic Net, with all models presenting some overlapping features with the highest SHAP scores. The findings show that all three types of models tested were applicable and offer promising potential for the further rollout of machine learning in this field.