DYNA JOURNAL ENGINEERING

Key words:

gas turbine, gas path, condition monitoring, artifitial neural networks, multilayer perceptron, deep learning, convolutional neural networks, engine fault diagnosis, measurement deviations, simulation software, validation testing.

Article type:

ARTICULO DE INVESTIGACION / RESEARCH ARTICLE

Section:

RESEARCH ARTICLES

Gas turbines are the main technology used in the generation of electricity in Mexico and worldwide. Therefore, the techniques that maintain high reliability of the turbines are in demand, in particular, diagnostic systems. To design the best system, this paper chooses the most accurate gas path diagnostic algorithm by comparing two algorithms that utilize different artificial neural networks. The first algorithm uses the Multilayer Perceptron MLP), and the second employs deep Convolutional Neural Networks (CNN). In order to carry out this comparison, both algorithms use the same input data that were generated by a specially designed engine simulation software. The development of this software was inspired by the operation of
the software ProDiMES. In contrast to ProDiMES simulating aircraft engine measurements at every flight, our software generates data measured at a stationary gas turbine powerplant every minute. This greater measurement frequency is required for CNN because it works with large input formation.
With this frequency, measured gas path variables are generated
for 5 fault classes: 4 different faults and a healthy engine
class. To have diagnostic features sensitive to these faults,
the measured variables are transformed into measurement
deviations. The vector of deviations of all variables forms a
pattern that is recognized by each network. This pattern is a
network input, and the corresponding fault class is an output.
Both networks are learned on the same multiple pairs of input
and output vectors. After learning, the networks are applied
to validation and testing data, and the probabilities of correct
diagnosis are computed for each network. The comparison of
these probabilities shows that CNN slightly yield to perceptron,
in spite of a common opinion of the high performance of
convolutional networks.