Industry Project
Monitoring abnormal Bayer-process precipitation and ranking candidate process drivers with XGBoost, SHAP, and historical directional rules.
This project investigates abnormal decomposition efficiency in the precipitation section of the Bayer process. It connects 1,139 historical records with inlet chemistry, tank temperatures, solids, and oxalate measurements to help engineers prioritize investigation.
Measured outlet dissolved alumina (AO) is the diagnostic target. Its relationship to efficiency follows the process chemistry formula; the project measures decomposition efficiency rather than hourly production throughput.
The workflow separates abnormality monitoring from factor explanation. Four statistical process-control rules flag large deviations, sustained deviations, rising trends, and adjacent jumps in measured outlet AO.
Twelve engineered process factors feed an XGBoost classifier for high versus low outlet AO. When the prediction matches the measured class, local SHAP values rank supporting factors, with extra weight for values outside the training interquartile range. When the classes differ, historical directional rules identify supporting factors; unmatched cases remain unknown.
Training-derived thresholds and reference ranges stay fixed during inference. Process inputs exclude outlet measurements and efficiency, and missing measurements are not replaced with zero. Diagnosis requires the corresponding observed outlet AO; explanation rules leave the original prediction unchanged.
In the random test split, supporting factors were available for 95.65% of the lowest-efficiency 10% and 20% cohorts (22/23 and 44/46 records). In the chronological historical test, the corresponding coverage was 100% (15/15 and 29/29 records).
These figures measure explanation coverage: the availability of SHAP or historical-rule support for observed outcomes. They do not measure prediction accuracy, physical loss allocation, or verified causal correctness. Causal confirmation requires additional engineering evidence.
The chronological evaluation used earlier records for training and later records for testing, but had no 48-hour label purge: six training labels reached or crossed the test boundary. Inner model selection used random five-fold cross-validation. These historical results do not guarantee future performance.