Comparing Different Schemes in a Combined Technique of Kalman Filter, Artificial Neural Network and Fuzzy Logic for Gas Turbines Online Diagnostics

Abstract

The paper presents research on the online performance-based diagnostics by implementing a novel methodology, which is based on the combination of Kalman Filter, Artificial Neural Network, Neuro-Fuzzy Logic and Fuzzy Logic. These methods are proposed to improve the success rate, increase the flexibility, and allow the detection of single and multiple failures. The methodology is applied to a 2-shaft industrial gas turbine engine for the automated early detection of single and multiple failures with the presence of measurement noise. The methodology offers performance prediction and the possibility of utilizing multiple schemes for the online diagnostics. The architecture leads to three possible schemes. The first scheme includes the base methodology and enables Kalman Filter for data filtering, Artificial Neural Network for the component efficiency prediction, the Neuro-Fuzzy logic for the failure quantification and the Fuzzy Logic for the failure classification. For this scheme, a performance simulation tool (Turbomatch) is used to calculate the thermodynamic baseline. The second scheme replaces Turbomatch with the Artificial Neural Network, that is used to calculate the deteriorated efficiencies and the reference efficiencies. The third scheme is identical to the first one but excludes the shaft power measurements, which are not available in aero engines or might not be usable for some power plant configurations. The paper compares the performance of the three methodologies, with the presence of measurement noise (0.4% reference noise and 2.0% reference noise), and 24 types of random single and multiple failures, with variable magnitude. The first methodology has been already presented by Togni et al. [10], whereas the other two methodologies and results are part of the PhD thesis presented by Togni [18] and they extend the applicability of the method. The success rate, targeting the correct detection of the of the failure magnitude ranges between 92% and 100% without measurement noise and ranges between 66% and 83% with measurement noise. Instead, the success rate of the classification, targeting the correct detection of the type of failure ranges between 93% and 100% without measurement noise and between 85% and 100% with measurement noise.