441 Figure 17. Determination of Residence Times in the Phase I Concentrator Plant. (Albarracin, 2021) 3.2. Data Preprocessing Data from ODJIT were used, where alteration variables (Intrusive A, Intrusive B, Skarn Serpentine– Magnetite, Skarn Actinolite, and Hornfels) are available as time series derived from originally georeferenced information. As a quality control step, records were filtered to retain only those with alteration balances between 99% and 101%, ensuring consistency in the representation of the total material. dditionally, negative and inconsistent values were removed. In parallel, operational variables were integrated from the PI System by matching timestamps. Only representative operating conditions were considered, specifically: TPH F1 > 3500 and Warman pump current > 200 A. 3.2.1. Description of Variables Considered in the ODJIT System To capture the short-term temporal dynamics of the SAG milling process, the original 15-minute resolution data were transformed into sliding windows of 1.5 hours (equivalent to six consecutive time steps), the windows were constructed, ensuring strict temporal continuity between observations. Only sequences with uninterrupted timestamps were considered, preventing artificial transitions caused by missing records or operational gaps. This approach guarantees that each input sample represents a physically consistent and continuous segment of the process, preserving the temporal structure required for convolutional feature extraction within the CVAE framework. 3.2.2. Data Normalization Prior to model training, all variables were scaled using the Min-Max normalization method implemented in the scikit-learn library. The transformation rescales each feature to the range [0,1] according to: D !"#$%& = D−D'() D'#* −D'() (1) This normalization ensures numerical stability during neural network training, prevents dominance of variables with larger magnitudes, and facilitates balanced learning across geological, geochemical, and geomechanical features. The scaler parameters were computed exclusively from the training data and subsequently applied to the complete dataset to avoid data leaking.
RkJQdWJsaXNoZXIy MTM0Mzk2