Comparison of multilayer perceptron with deep learning neural networks applied to gas turbine diagnosis

Abstract

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 power 
plant every minute. This greater measurement frequency is 
required for CNN because it works with large input information. 
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.