Abstract:
Backfill is used as roof support in deep level gold mines in South Africa. The quality of backfill, in terms of the fraction of material finer than 10 microns, the solids concentration and the fraction of solids recovered from the feed, is critical to its hydraulic transport and support properties. There is a need to detect faults in backfill hydrocyclones that cause deterioration in backfill quality before these faults become so severe as to cause failure of the system. A survey of backfill installations indicated that common faults in backfill plants are wear in the spigot and vortex finders and changes in the feed size distribution. These process variables are not regularly measured. All backfill plants monitor the three indicators of backfill quality and are able to measure the operating pressure. In general, the feed solids concentration is monitored and controlled. To date, no diagnostic system has been developed for hydrocyclones. A quantitative fault diagnosis procedure was developed that uses steady-state models of the process and model parameter estimation to infer the values of the non-measured or non-measurable process variables. The process variables at fault can then be identified by analysing the inferred values. A review of hydrocyclone modelling indicated that mechanistic models based on the physics of fluid flow are becoming suitable for the prediction of hydrocyclone performance. They are, however, too computationally intensive for a parameter estimation approach. Semi-empirical hydrocyclone models describe the performance of hydrocyclones in terms of the partition curve rather than the measured process responses. The development of purely empirical hydrocyclone models that relate measured process responses to the process inputs and non-measured process variables was investigated. Artificial neural networks have been used for the modelling of ill-defined metallurgical processes and their application was investigated. Quadratic models, artificial neural network models and a semi-empirical hydrocyclone model were statistically compared. It was shown that quadratic regression models are suitable for modelling hydrocyclone responses. Such models can be developed using centrally composite rotatable designed experiments to reduce the experimentation required. The fault diagnosis system was implemented and tested using simulation. By introducing changes into the process variables in the simulator, the effectiveness of the fault detection and diagnosis procedure was evaluated. Alarm analysis was performed using qualitative and non-parametric and parametric quantitative methods. It was found that qualitative methods are able to indicate deteriorating variables from their inferred values, but that unique fault cause identification deteriorated at high measurement noise levels. A combination of parametric and non-parametric alarm analysis was found to yield a system that rapidly indicated changes in process variables, yet indicated few false alarms. It was shown that quantitative detection and diagnosis of common faults in backfill hydrocyclones is possible using the process measurements that are taken to characterise backfill quality. The steady-state modelling and parameter estimation strategy has proven to yield a suitable methodology for quantitative fault diagnosis in backfill hydrocyclones.