DBSCAN for improved oscillations diagnosis in mineral processing plants, E.C. Nienaber, B.S. Lindner, L. Auret, and J.W.D. Groenewald

Mineral processing plants have complicated control strategies, with different control layers, multiple interacting loops, and feedforward disturbance rejection. The complicated control strategies mean that oscillations can arise and propagate to multiple loops and degrade the control performance. Oscillation detection techniques, such as the spectral envelope method, can be applied to detect oscillations within an observation window. However, the start- and end-points of the oscillations within the observation window cannot be identified with current oscillation detection techniques. The start- and end-points would be useful to label the oscillatory intervals, and then to use these labels to identify process conditions associated with the oscillation. In this paper, a pattern recognition technique, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), was used to label oscillation intervals. An oscillation diagnosis workflow combining spectral envelope and DBSCAN was demonstrated on an industrial case study of oscillations in a flotation circuit concentrate sump. The workflow successfully detected the oscillation, labelled the oscillatory intervals, and enabled accurate root cause analysis. The impact of time series features, namely the signal power ratio (𝑃𝑅) and the oscillation frequency, on the performance and optimal parameter selection of DBSCAN was also investigated. Low oscillation frequencies and low 𝑃𝑅s negatively affected the performance of DBSCAN. Additionally, the results indicated that higher oscillation frequency was related to higher optimal values for the lag parameter. Keywords: Oscillation diagnosis, process monitoring, data mining, clustering, flotation control