In collaboration with our partner, Cylad, we addressed these obstacles by leveraging our combined expertise. Our client, a European base metal refinery, engaged us to apply machine learning algorithms for analyzing the time data series gathered during their base metal purification process. Our joint objective was to develop a data-driven solution that could enhance operational performance by optimizing process parameters, minimizing impurity levels, and achieving overall process optimization.

Time-series data can be challenging for humans to interpret due to its sequential nature, complex patterns, and high dimensionality. Thus we utilized a combination of machine learning models to interpret the time data series and identify key patterns and relationships within the process. Our team employed the following models:

• Long Short-Term Memory (LSTM): LSTMs are powerful deep learning models capable of learning long-term dependencies in sequential data. By training an LSTM model on the time series data, we were able to capture and analyze complex temporal patterns within the base metal purification process.

• ROCKET (Randomized Convolutional Kernel Transform): ROCKET is a novel algorithm that extracts features from time series data using randomized convolutional kernels. This model enabled us to uncover essential characteristics and trends in the time series data, providing valuable insights for process optimization.

• XGBoost: XGBoost is a popular gradient-boosting algorithm known for its high performance and interpretability. By applying XGBoost to the processed time series data, we were able to build a predictive model that gives early warnings about future operational problems in the process flow.

Through our collaborative efforts, we achieved significant improvements in the base metal purification process. By analyzing the time series data, our models identified the key parameters that had the most significant impact on impurity levels. This information allowed us to focus our optimization efforts on those specific parameters, resulting in substantial efficiency gains.

By training our models on the time series data, we were able to accurately predict critical process variables at different stages of the process. This predictive capability enabled proactive decision-making and facilitated early intervention to prevent the escalation of impurities.