航空發動機人工神經網路故障診斷法

Fault Diagnosis of Aeroengine Using Neural Networks

陸鵬舉(Peng-Jeu Lu) ； 徐自珍(Tzi-Cheng Hsu)

《中國航空太空學會學刊》 32卷4期 (2000/12) Pp. 301-308

https://doi.org/10.6124/TAASRC.200012_32(4).04

摘要

本研究旨在探討人工神經網路在航空發動機故障診斷領域的應用，並透過程式驗證工作以確保網路程式的正確性。本文採用兩種不同的人工神經網路並進行比較；除了常用的倒傳式神經網路(Back Propagation, BP)外，並針對BP網路學習較慢及網路架構選取過於繁複等缺點而引用串接關聯神經網路(Cascade Correlation, CC)。本文使用了普惠(Pratt & Whitney)公司PW4000型發動機之20組故障案例資料，10組作爲網路學習，另外10紐作爲測試之用。結果顯示，無論是BP網路或是CC網路，對10組供作測試的案例資料在診斷效果上均能有100%的正確率，而CC網路也的確能改進BP網路在學習速度架構決定上的問題。同時，對於一個訓練完成的網路而言，BP網路與CC網路均能在某些參數（如△Wf、△N1、△N2、△PB/P4.9等）無法提供的情況下進行模糊診斷，且達到90%的正確率。

關鍵字

故障診斷 ； 倒傳式神經網路 ； 串接關聯神經網路

並列摘要

The objective of the present research is to study whether the technique of artificial neural networks (ANN) is applicable in the field of aeroengine fault diagnostics. The developed ANN was first validated to ensure the correctness of the networks. Two network architectures were considered with comparison made as well. The first is the modified Back Propagation (BP) network and the second is the Cascade Correlation (CC) network. The latter enjoys the advantage of fast convergence and requires no a priori information in determining the network architecture. In this work, there are 20 PW4000 engine fault cases that are selected from the P&W case history book are used for the network training and testing. Results show that both trained BP and CC networks can achieve a 100% success rate in the fault diagnostics. In the meanwhile, the argument that CC can improve the shortcomings of BP is also demonstrated. For a trained BP or CC network, it is found that a 90% success rate of fuzzy diagnosis can be achieved when some of the parameters (such as △Wf、△NI、△N2、△PB/P4.9, etc) are not available.

並列關鍵字

Fault Diagnostics ； Back Propagation Network ； Cascade Correlation Network