Gas turbine engine dynamic model based on variable-memory LSTM architecture

Abstract

The buildup of thermodynamic cycle parameters is the main way to increase gas turbine engine efficiency. However, the growth of engine pressure and temperature ratio leads to the increase in the turbine heat load, which reduces the engine lifetime dramatically. In terms of gas turbine engines, to avoid the engine life loss is a crucial problem especially for small engines, because the limited size of a small gas turbine engine does not allow implementing various measures for nozzle vane cooling. In light of this, the contribution of the turbine heat control is essentially increasing. It places great demands on the accuracy of control over the main engine variables (such as the rotor speed and turbine outlet temperature). The state-of-the-industry gas turbine engines use an on-board engine mathematical model to improve the quality of the control. These models deal with engine processes of short duration and considerable overshooting. For that reason, the model accuracy is the main aspect in the control process. However, the issues of accurate and at the same time resource-saving calculation of rapidly varying processes of changing the rotor speed and the turbine gas temperature remain under-investigated. In the work, neural network methods were used to model the unsteady modes of a small gas turbine engine. Using the data obtained as a result of firing tests of the JetCat P-60 engine, the engine regression neural network model was created. The main issue that arose during the creation of the model was to describe the dynamics of rapidly varying processes with pronounced overshoot. For this purpose, modification of the architecture of the classical LSTM network was carried out, the essence of which was to add a functional dependence of the exit node on the memory tensor. This allowed us to make the memory size independent of the number of model outputs, thereby increasing the modeling accuracy. The developed architecture was proposed a new name - VMLSTM network. As a result of comparison with the traditional Elman network and the classic LSTM network, the developed VMLSTM network showed the least value of the average error with a comparable number of modifiable model parameters. In addition, unlike the existing neural networks, the developed network demonstrated the ability to simulate turbine outlet gas over-temperature at the moments when the engine operating mode changes. The developed neural network architecture increases the reliability of modeling the dynamics of a small gas turbine engine as an object of control, which in the conditions of economical use of computing resources opens up possibilities of its application in on-board microcomputers.

Methodology for Control of Helicopters Aircraft Engines Technical State in Flight Modes Using Neural Networks

“…3). It is also worth noting that a recurrent neural network of the LSTM structure [23] has been successfully applied to solve the problem of identification of an aviation GTE. Therefore, in this work, a recurrent neural network of the LSTM structure is used to solve the problem of monitoring the technical state of aircraft GTEs of helicopters in flight mode.…”

“…According to the research results, the best results were obtained using the error backpropagation algorithm. This is primarily due to the fact that the backpropagation algorithm has been created and optimized for the selected neural network architecture and activation function, which makes it possible to maximize its potential [22,23].…”

“…According to [23], in this work it is advisable to apply an LSTM network based on a dynamic model of a gas turbine engine based on the classical LSTM structure (fig. 5, a) or LSTM structure with variable memory (fig.…”

“…where с t -memory tensor representing a vector of weighted inputs in a step t; f t -tensor at the exit from the forgetting node in the step t, representing the sum of the weighted inputs and outputs in the previous step t 1; c t1 -step memory tensor t 1; i t -tensor at the output of the input node at step t, which is the sum of the weighted inputs at the current step and outputs at the previous step; cc tcandidate tensor to write to memory tensor [23]. As mentioned above, to solve the problem of dynamic monitoring of aircraft gas turbine engines, a special architecture of a recurrent neural network with long-short-term memory (LSTM) was developed, presented in [23]. When using LSTM networks, the problem of the vanishing gradient of the LSTM network arises, which the filter mechanism allows to resist.…”

“…As an example, a test version of a recurrent neural network was developed in the Matlab environment, illustrating the solution to the problem of TV3-117 aircraft GTE technical state control. The neural network was trained according to the following rule [22,23]. First, all 10000 vectors of the training set were sequentially fed to the network input, with the help of which training was performed using the backpropagation algorithm.…”