No abstract

Gas turbine health monitoring includes the common stages of problem detection, fault identification, and prognostics. To extract useful diagnostic information from raw recorded data, these stages require a preliminary operation of computing differences between measurements and an engine baseline, which is a function of engine operating conditions. These deviations of measured values from the baseline data can be good indicators of engine health. However, their quality and the success of all diagnostic stages strongly depend on the adequacy of the baseline model employed and, in particular, on the mathematical techniques applied to create it.To create a baseline model, we have applied polynomials and the least squares method for computing the coefficients over a long period of time. Methods were proposed to enhance such a polynomial-based model. The resulting accuracy was sufficient for reliable monitoring of gas turbine deterioration effects. The polynomials previously investigated thus far are used in the present study as a standard for evaluating artificial neural networks, a very popular technique in gas turbine diagnostics. The focus of this comparative study is to verify whether the use of networks results in a better description of the engine baseline. Extensive field data for two different industrial gas turbines were used to compare these two techniques under various conditions. The deviations were computed for all available data, and the quality of the resulting deviation plots was compared visually. The mean error of the baseline model was used as an additional criterion for comparing the techniques. To find the best network configurations, many network variations were realised and compared with the polynomials. Although the neural networks studied were found to be close to the polynomials in accuracy, they did not exceed the polynomials in any variation. In this way, it seems that polynomials can be successfully used for engine monitoring, at least for the gas turbines analysed herein.

Gas turbine diagnostic algorithms widely use fault simulation schemes, in which measurement errors are usually given by theoretical random number distributions, like the Gaussian probability density function. The scatter of simulated noise is determined on the basis of known information on maximum errors for every sensor type. Such simulation differs from real diagnosis because instead of measurements themselves the diagnostic algorithms work with their deviations from an engine baseline. In addition to simulated measurement inaccuracy, the deviations computed for real data have other error components. In this way, simulated and real deviation errors differ by amplitude and distribution. As a result, simulation-based investigations might result in too optimistic conclusions on gas turbine diagnosis reliability. To understand error features, deviations of real measurements are analyzed in the present paper. To make error presentation more realistic, it is proposed to extract an error component from real deviations and to integrate it in fault description. Finally, the effect of the new noise representation mode on diagnostic reliability is estimated. It is shown that the reliability change due to inexact error simulation can be significant.

Efficiency of gas turbine condition monitoring systems depends on quality of diagnostic analysis at all its stages such as feature extraction (from raw input data), fault detection, fault identification, and prognosis. Fault identification algorithms based on the gas path analysis may be considered as an important and sophisticated component of these systems. These algorithms widely use pattern recognition techniques, mostly different artificial neural networks. In order to choose the best technique, the present paper compares two network types: a multilayer perceptron and a radial basis network. The first network is being commonly applied to recognize gas turbine faults. However, some studies note high recognition capabilities of the second network. For the purpose of the comparison, both networks were included into a special testing procedure that computes for each network the true positive rate that is the probability of a correct diagnosis. Networks were first tuned and then compared using this criterion. Same procedure input data were fed to both networks during the comparison. However, to draw firm conclusions on the networks’ applicability, comparative calculations were repeated with different variations of these data. In particular, two engines that differ in an application and gas path structure were chosen as a test case. By way of summing up comparison results, the conclusion is that the radial basis network is a little more accurate than the perceptron, however the former needs much more available computer memory and computation time.

Gas turbine diagnostic techniques are often based on the recognition methods using the deviations between actual and expected thermodynamic performances. The problem is that the deviations generally depend on current operational conditions. However, our studies show that such a dependency can be low. In this paper, we propose a generalized fault classification that is independent of the operational conditions. To prove this idea, the probabilities of true diagnosis were computed and compared for two cases: the proposed classification and the conventional one based on a fixed operating point. The probabilities were calculated through a stochastic modeling of the diagnostic process. In this process, a thermodynamic model generates deviations that are induced by the faults, and an artificial neural network recognizes these faults. The proposed classification principle has been implemented for both steady state and transient operation of the analyzed gas turbine. The results show that the adoption of the generalized classification hardly affects diagnosis trustworthiness and the classification can be proposed for practical realization.