Condition Monitoring and Fault Diagnosis for Electric Power Generators

Abstract

The chapter analyzes condition monitoring and fault diagnosis for electric power generators with the following approaches: (i) fault diagnosis in the time domain for distributed interconnected synchronous generators and with the use of the local statistical approach to fault diagnosis, (ii) fault diagnosis in the time domain for synchronous generators and with the use of nonlinear filtering methods, (iii) fault diagnosis for asynchronous generators, making use of neural networks having activation functions that remain invariant to the Fourier transform. First the chapter analyzes a Kalman Filtering-based approach for fault diagnosis in distributed and interconnected power generators. Due to the development of distributed interconnected power units, e.g. inland and offshore wind farms, the need for systematic methods for fault diagnosis in such multi-machine power systems has emerged. In this section a new approach to parametric change detection and failure diagnosis for interconnected power units is proposed. The method is based on a nonlinear filtering scheme under the name Derivative-free nonlinear Kalman Filter and on statistical processing of the obtained state estimates, according to the properties of the χ 2 distribution. To apply this fault diagnosis method, first it is shown that the dynamic model of the distributed interconnected power generators is a differentially flat one. Next, by exploiting differential flatness properties, a change of variables (diffeomorphism) is applied to the power system, which enables also to solve the associated state estimation (filtering) problem. The new filtering technique consists of (i) a change of variables (diffeomorphism) which results into a linearized equivalent model for the power system, (ii) application of the Kalman Filter recursion, and (iii) an inverse transformation based again on differential flatness theory which permits to obtain state estimates for the initial nonlinear model. Next, statistical processing is performed for the obtained residuals, that is for the differences between the state vector of the monitored power system and the state vector provided by the aforementioned filter when the latter makes use of a fault-free model. It is shown, that the suitably weighted square of the residuals' vector follows the χ 2 statistical