Fault detection for the flotation process

Sammanfattning

This thesis project investigates the possibility of identifying faulty equipment in a flotation process within the mining industry at Boliden AB, at their refinement center connected to the Aitik mine. The objective is to rank the flotation cells in terms of repair need, achieved by detecting abnormal control behavior without explicitly identifying the type of faults present. Two approaches are considered: a model-based method using a linear state-space model with a Kalman filter, and a data-driven method using an LSTM-based autoencoder trained on process data. Both methods aim to detect deviations from expected behavior and produce a ranking of the cells. The results show that both methods are capable of producing ordered cell rankings. The Kalman filter provides a physically motivated framework and demonstrated more consistent results, where elevated innovation scores correspond fairly well with abnormal behavior in the measured signals. However, strong coupling between neighboring cells in the flotation process, abnormalities propagate and give elevated scores for both the faulty cell and its neighboring cell, making it difficult to determine whether anomalies originate from a specific cell or from upstream effects. The inclusion of pump dynamics was investigated but excluded due to scaling issues and modeling inaccuracies. The autoencoder offers a flexible alternative but depends strongly on feature selection. Relevant fault behavior may not be captured, while normal variations in the process can be incorrectly flagged as anomalies. Additionally, the use of unlabeled process data introduces uncertainty, as the models may learn representations that include faulty behavior. The study highlights the challenges of fault detection in industrial processes. While both approaches provide useful insights, the Kalman filter showed greater robustness in this setting. Further work is required to improve model accuracy, handle process coupling, and evaluate the methods using verified fault data.