Development of a soft sensor for predicting dense medium cyclones performance

This project details the development of a DMC soft sensor for modelling a coal processing circuit consisting of a desliming screen and two DMCs in series. The soft sensor comprises a number of mathematical models for the coal washability, DMC operation, desliming screen efficiency, and pump operation. The mathematical models are continuously tuned to online data from all of the instrumentation around the circuit, enabling the soft sensor to best fit all of the available physical sensor information while simulating useful additional information about the circuit operation.

The washability model plays a key role in the soft sensor. Although laboratory washability is available and can be linked with the coal type being processed, it is known that this is not necessarily the best reflection of the coal being processed at any given time. Variability of the coal within the seam will occur, meaning the laboratory washability results will not accurately reflect the coal being processed. By modelling the washability using the laboratory results and a two-parameter model, it is possible to adjust the modelled washability of the coal being treated. This adjustment in the washability improves the DMC soft sensor prediction capacity.

Wear has shown to be important to consider in soft sensor development, and so an exponential model for spigot wear is developed. Nevertheless, wear data shows inconsistent values where the spigot diameter decreases without an actual spigot replacement. This highlights the importance of improving the wear data measurements to obtain a more accurate wear model. Ideally, a pump wear model would also have been considered, but details on pump wear were unavailable for this project.

Combining the models to simulate the flowsheet over a years' worth of data shows that a static model does not provide good outcomes, mainly due to the fixed washabilities used in this scenario. When the soft sensor approach of adjusting specific model parameters to minimise the overall error between measured and modelled parameters is used, results are shown to be vastly improved.

Further model calibration is undertaken to improve the accuracy of the DMC soft sensor by manually setting key model parameters. While the DMC soft sensor outputs are shown to have a good relationship with the DMC product flow rates and the medium density validation data, predictions of ash content were poorer than desirable. Multiple machine learning models and configurations are tested to determine if post-processing can result in better ash content predictions. While some results looked very promising further analysis suggests that they are likely the effect of overfitting to data, and if applied to new time series data, then the model performance would not be nearly as accurate as the machine learning results suggest.