Use of data from anode current distribution for state and parameter estimation and fault detection in an aluminium prebake electrolysis cell

Abstract

The work undertaken in this thesis addresses the use of individual anode current measurements for stat and parameter estimation and fault detection in aluminium electrolysis cells. The work has focused on different methods where measurements of anode current distribution could be used to infer major changes in alumina concentration, and possibly detect abnormal process conditions like faulty feeding  equipment and oncoming  anode effects. Two main approaches have been pursued, a physical model based approach and a pure data driven approach.

 

The main part of the work has focused on the model based approach. Based on physical principles and empirical correlations from the literature, several models have been developed that use measurements of current through the individual anodes in a cell as input and estimate cell voltage and distribution of alumina concentration in the electrolyte. The models are implemented in Extended Kalman Filters allowing states and a number of process parameters to be updated based on measured cell voltage.

 

Experimental data from real aluminium electrolysis cells are used to validate the models and estimate the unknown model parameters. The Extended Kalman Filter estimates are compared to measured outputs, and turning and robustness of the estimator for different operating conditions are discussed. In general the estimates of alumina concentration and cell voltage fit the measurements reasonably well for the available measurement data. The performance and robustness of the Kalman Filter depend on how well the Kalman Filter gain matrix is tuned to blen model and measurement information, as well as the identifiable and observability of the parameters and the states given the available input and measurement data. All states and parameters can rarely be estimated simultaneously, an in practice we must determine what states and / or parameters to estimate based on the available data.

 

The experimental data typically involve some manipulation of feed rate to excite the concentration of the raw material alumina outside of the normal operating fange, and provoke an abnormal condition known as anode effect. The data have been used to study the performance of the Kalman Filter in situations where the actual feed rate of alumina in the cell is different from what is assumed by the process control system. When the inout feed rate to the model is faulty, updating the model states or a few chosen model parameters from the measured cell voltage (or cell resistance) can significantly improve the estimates of alumina concentration. This can be used to detect faulty input feed rate and abnormal changes in alumina concentration distribution in the cell. The results suggest that using the Kalman Filter to generate faults, anode effects in the experimental data could be detected 10-20 minutes ahead.

 

Detection of anode effects and abnormal distribution of alumina concentration has also been studied using the data based fault detection method Principle Component Analysis. This method also shows some potential for improved detection of anode effect, but in general, detection is closer to the oncoming anode effect than for the model based approach. A weakness of data based methods is that they in general cannot explain variation in the data not represented by the training data, and therefore the frequency of false alarms can be high. An advantage with the model based approach is that this method can relate the abnormal situation directly to a specific parameter or state (s), thereby giving a better indication of the cause of the fault. Only faults related to abnormal feed rates are evaluated here, and further work are necessary to assess the robustness and performance for the two methods with other common operating conditions and faults in the data.