Monitoring of a semi-autogenous grinding circuit on a gold plant with topology preserving projection methods and neural networks

In this paper, the use of topology preserving projections (Sammon maps) and neural networks are considered as a basis for the construction of multivariate statistical process control charts for a semi-autogenous grinding circuit on a gold plant. Comminution circuits are often operated in a quasi-steady state mode, where the mills tend to move rapidly from one state to another. This makes them difficult to monitor by conventional means designed for continuous steady state processes. As a first step, process data reflecting normal operating conditions are mapped with Sammon's algorithm. Only one feature is extracted at a time and each mapping is captured by a neural network model to allow online application. Reverse neural network models are used to reconstruct the data prior to the extraction of subsequent features from the data residuals. The ensemble of neural network models can then be used for both fault detection and identification. The approach compared favourable with both linear and non-linear principal component models. It was able to detect two simulated fault conditions both with a reliability of 100 per cent, as opposed to 57.8 per cent and 95.2 per cent detected by an approach based on the use of linear principal component analysis and 84.1 per cent and 98.0 per cent detected by an approach based on non-linear principal components extracted with auto-associative neural networks.