Intelligent engine condition monitoring and early fault detection is becoming a necessity for modern gas turbines to achieve availability and reliability requirements. Degradation or failure of critical control components negatively affects reliability and safety as well as the ability to perform condition monitoring of the engine. Model-based approaches for analytics and condition monitoring show great promise with advances in remote connectivity and available computational power. Development of diagnostics for analytics requires focus on the target machine as well as the integrity of measurement and actuation systems to correctly identify and classify degradation indicators that discriminate between actuation and measurement faults from deterioration in machine performance. In this paper, a method is proposed to use data obtained in closed-loop operation to identify system models that separate the dynamic responses and non-linear characteristics for use in analytics. This is demonstrated on high-fidelity simulation data from a Taurus™ 60 gas turbine generator. The system is modeled with a feedback connection of a known controller in series with a block Hammerstein system.

Many practical applications, such as the fuel control of a gas turbine engine, can be modeled by a feedback connection of a linear controller in series with a Hammerstein system, where the nonlinearity provides a representation of the control element or actuator. An iterative gradient-based method is proposed to simultaneously identify the nonlinear fuel valve characteristic and a low-order linear plant model in gas turbine applications that leverages a priori knowledge of both the nonlinearity and engine dynamics. The identification is a nonlinear prediction error minimization method in a closedloop Hammerstein model framework. It is applied to data from a high-fidelity simulation of a 5 megawatt Taurus T M 60 industrial gas turbine.

No abstract

The intersection of machine learning methods and gas turbine sensor data has expanded rapidly in the last decade to include numerous applications of regression, clustering, and even neural network algorithms. Learning algorithms have pushed traditional engine health management into the realm of prognostic health management. This paper starts with a review of several common computational methods used to monitor the condition of gas turbines currently employed by both industry and academia. Sources of application of machine learning algorithms from outside the gas turbine industry are also brought in. Focus is generally placed on industrial gas turbines with an industry standard monitoring system. The authors explore beyond gas path analysis with a novel use of machine learning algorithms to engine component classification. The paper concludes with a case study of applying learning algorithms to machine data to identify different fuel valves.

Many of the components on a gas turbine are subject to fouling and degradation over time due to debris buildup. For example, axial compressors are susceptible to degradation as a result of debris buildup on compressor blades. Similarly, air-cooled lube oil heat exchangers incur degradation as a result of debris buildup in the cooling air passageways. In this paper, we develop a method for estimating the degradation rate of a given gas turbine component that experiences recoverable degradation due to normal operation over time. We then establish an economic maintenance scheduling model, which utilizes the derived rate and user input economic factors to provide a locally optimal maintenance schedule with minimized operator costs. The rate estimation method makes use of statistical methods combined with historical data to give an algorithm with which a performance loss rate can be extracted from noisy data measurements. The economic maintenance schedule is then derived by minimizing the cost model in user specified intervals and the final schedule results as a combination of the locally optimized schedules. The goal of the combination of algorithms is to maximize component output and efficiency, while minimizing maintenance costs. The rate estimation method is validated by simulation where the underlying noisy data measurements come from a known probability distribution. Then, an example schedule optimization is provided to validate the economic optimization model and show the efficacy of the combined methods.