Selected Projects

In this project, the supervised and unsupervised methods of classification are used for fault detection and identification of the Gas Turbine (GT) system. These methods are signal based approach for fault detection and identification. The performance signals of the gas turbine are acquired in normal and fault occurred conditions. Data acquired from Simani gas turbine simulated model and the Siemens V94.2 gas turbine. This Simani model is capable of simulating four common faults of gas turbine. It has been validated with a single-shaft industrial gas turbine so its performance is matching the real gas turbines performance. Thus this simulated model could be a useful instrumental for academic studies. The V94.2 gas turbine is the most useful gas turbine in the power plants in Iran. Its nominal power is 162 MW. The under taken gas turbine is stored in the 3rd plant of the Parrehsar power plant, Iran. In this research the Fuzzy CMeans, Gustafson-Kessel and Gath-Geva clustering algorithms, as unsupervised classification methods, are used to detect and identify fault. The ANFIS, MLP and RBF neural networks are used as supervised classification methods. Feature selection methods have been applied on the data set. Principal component analysis (PCA) and linear discriminant analysis (LDA) used as feature selection methods. The results show that combination of the supervised and unsupervised classification methods, able to detect and identify the normal and fault behavior of the gas turbine. Results indicate that using feature selection methods improves the performance of the fuzzy clustering algorithms. Combination of the LDA feature selection and the GK fuzzy clustering algorithm performs a considerable performance.

Gas turbine, as the prime mover in fossil and combined cycle power plants, is one of the core elements of energy production. Turbine malfunctioning not only reduce the efficiency of the plant, but also will cause its damage and failure. Furthermore, it makes heavy costs on industry owners. Therefore, gas turbine condition monitoring is a logical and preventive operation to avoid damages caused by improper operation of the turbine. In this project, a new condition monitoring system to identify common faults in a gas turbine is designed. It makes use of temperature sensors data. Soft sensors and invented hybrid adaptive threshold are used for making a decision. First, a residual signal is produced using soft sensors designed by a feedforward neural network. Second, a Mamdani Fuzzy system is utilized for an online determination of parameters to generate an adaptive threshold. Usually investigations of the gas turbine are carried out on the simulator of a laboratory set up. The simulator is capable for creating various faults in three different operating points. The superiority of the designed system is specified by comparing obtained results with other research. In addition to simplicity and increasing the confidence coefficient, the system is properly sensitive to faults.

Neural networks (NNs) are an important data driven tool for estimating and classifying petroleum reservoir properties. In this project, seismic attributes have been used to estimate well logs in a fractured reservoir case study from Iran. Three methods have been evaluated: the linear regression, the multilayer perceptron (MLP) and the radial basis function (RBF). For linear case, the selection of appropriate attributes was determined by forward selection and for nonlinear one; the selection was based on the genetic algorithm result. 10-fold cross-validation method was used to train and evaluate the predictor models to overcome the over-fitted problem occurred in the small-data-set case study. Then by adding a decision function to the designed models, the reservoir properties were classified, meanwhile the well logs were estimated. By considering correlation coefficient criterion and confidence matrix, MLP with two hidden layers had the best performance; so reservoir characterization was done by use of this network. At last, thirteen desirable areas to develop the oil field and one area with the risk of troublesome gas production were identified throughout the reservoir.

Neural networks (NNs) are an important data driven tool for estimating and classifying petroleum reservoir properties. In this project, seismic attributes have been used to estimate well logs in a fractured reservoir case study from Iran. Three methods have been evaluated: the linear regression, the multilayer perceptron (MLP) and the radial basis function (RBF). For linear case, the selection of appropriate attributes was determined by forward selection and for nonlinear one; the selection was based on the genetic algorithm result. 10-fold cross-validation method was used to train and evaluate the predictor models to overcome the over-fitted problem occurred in the small-data-set case study. Then by adding a decision function to the designed models, the reservoir properties were classified, meanwhile the well logs were estimated. By considering correlation coefficient criterion and confidence matrix, MLP with two hidden layers had the best performance; so reservoir characterization was done by use of this network. At last, thirteen desirable areas to develop the oil field and one area with the risk of troublesome gas production were identified throughout the reservoir.

In a project, a member of APAC worked on it during his research visit to the University of Alberta, the actual effect of caustic addition on bitumen recovery rate was evaluated based on analyzing the routine operational data. It is theoretically and experimentally shown in the literature that caustic addition helps bitumen recovery when ore grade is low. However, this effect has not been clearly observed from the operational data. Therefore, this project aimed to discovering if the data shows a relation between caustic addition and recovery rate and proposing how to improve caustic control guideline for better recovery rate if possible. The fact that caustic may only have an impact in some situations and not in others, in addition to lack of necessary important measurements, large noise amplitude and complexity of the process, made it challenging to evaluate the actual effect of caustic on recovery rate based on historical data. A combination of careful data analysis with the process knowledge showed that caustic in general has improved the recovery rate in the understudy plant operation, but the improvement might not be observed by the operators due to the less improvement than expected, time delay, slow rate measurement of recovery, and some other relevant factors.