Deterministic behavior of temperature field in turboprop engine via shallow neural networks

Abstract

A study of machine learning approaches for temperature field (rake) prediction in a turboprop engine is presented. The potential of supervised machine learning and shallow feed-forward neural network (NN) architectures is studied to predict and reconstruct temperature fields based on a single sensor measurement. The reason to study shallow (not deep nor gated) NN first before more complex networks is that shallow NN can provide us with robust nonlinear mapping at a low computational cost. Simultaneously, their mathematical structures can be simple, especially for a limited amount of training data. Thus, revealing a governing law in data via a simpler architecture is desirable. Further, the problem suits feed-forward architectures as only one temperature sensor is considered input in real time. It is investigated which type of shallow NN architectures with which learning algorithm can be most accurate with the best generalization for provided turboprop temperature field data. The important finding is that it is possible to capture the deterministic governing law of temperatures in a turboprop engine. Thus, the temperature sensor locations in rakes can be analyzed to allocate the important positions for a limited number of temperature sensors inside the engine. The machine learning results also confirm the importance of slow heat transfer between internal engine parts and temperature sensors alongside the air propulsion. Thus, the proposed neural network application concept is promising as a funding base for further design of modern turboprop health monitoring system.

Appendix

See Tables 2, 3 and Figs. 14, 15.

Table 2 Details for testing all neural architectures and learning algorithms for dataset I as in Fig. 5 (the best achieved results are in bold)

Table 3 Details for testing all neural architectures and learning algorithms for dataset II as in Fig. 6 (the best achieved results are in bold)

Fig. 14

Dataset I: The example of prediction results comparison showing the mapping of measured temperatures $\theta_{17} = \theta_{17} (\theta_{1} )$ and the trained neural architectures $\tilde{y}_{HONU17} = \tilde{y}_{HONU17} (\theta_{1} )$ and $\tilde{y}_{MLP17} = \tilde{y}_{MLP17} (\theta_{1} )$ (z-scored data)

Fig. 15

Dataset II: The example of prediction results comparison showing the mapping of measured temperatures $\theta_{19} = \theta_{19} (\theta_{13} )$ and the trained neural architectures and $\tilde{y}_{MLP19} = \tilde{y}_{MLP19} (\theta_{13} )$