Reduction –based increase of robustness of a neural-net model of gas-turbine engine monitoring
UDK 621.438, 004.855 BBK З363.3, 32.813, 22.18
The article presents a technique of increasing the robustness of a neural-net model of gas-turbine engine monitoring during stand tests due to decreasing the number of insignificant links in the neural network (neural network reduction). The technique is based on converting the problem of neural network learning into that of multi-criteria optimization which includes the error minimization criterion and the criterion of minimizing the absolute values of the weight links of a neural network. The latter requirement leads to reveal of insignificant links which can be deleted without any loss of accuracy. As a result, the ability of the model to summarize increases greatly, robustness increases as well, the calculation error of monitored parameters decreases.
Key words: model robustness, neural networks, stand tests, monitoring of a gas-turbine engine, reduction of a neural network.
Download article in PDF 219.4 kB