應用倒傳遞類神經網路預測飛機零組件之故障時間──以T700發動機電子控制單元為例

Using Back-Propagation Neural Network to Predict Time-to-Failure of Aircraft Components – A Case Study on Electrical Control Unit of T700 Engine

攻擊直升機擔負灘岸攻擊與反登陸作戰任務，扮演極重要角色。飛行中，若飛機上重要零組件發生非預期性故障，勢必放棄作戰任務返場檢修；因此，若能對零組件建置故障時間預測系統，於故障前採取適當措施，可有效降低非預期性故障發生，提高任務成功率。本研究以我國某攻擊直升機使用T700發動機之電子控制單元為例，採用修正式德菲法製作第一次專家問卷進行調查，蒐集影響電子控制單元故障時間的關鍵因素，再以李克特量表製作第二次專家問卷進行評分，評選出七項重要關鍵因素。蒐集2011年至2013年的電子控制單元檢測數據為樣本，載入倒傳遞類神經網路軟體Alyuda NeuroIntelligence測試輸入與輸出之間的關係，藉以有效預測出電子控制單元之故障時間。研究結果顯示，經Alyuda NeuroIntelligence學習訓練，關聯性（Correlation）與模式配適度（R-squared）分別達到0.999及0.997，預測準確度達92.45%，其高預測準確度有實務上的應用價值，也印證倒傳遞類神經網路可做為飛機零組件之故障時間預測能力的標準。

The attack helicopters play an extremely important role in all anti-landing and littoral assaults operation, when unexpected malfunctions of critical components during operations occurred, to aboard mission and return for further repairs will become inevitable, therefore, to establish the time-to-failure predicting system against components, also apply appropriate actions prior to failures, operational readiness rate will be enhanced and unexpected malfunctions will be effectively reduced. This study was focused on the Electrical Control Unit which installed on the T700 engine of a random attack helicopter in Taiwan. The Modified Delphi Method was applied when conducting initial questionnaires to collect critical influencing factors of Electrical Control Unit time-to-failure, secondary questionnaires were sent out along with establishing Likert Scale and total 7 critical factors were identified. Testing data of Electronic Control Unit between 2011 to 2013 were collected, then loaded with Back-Propagation Neural Network software Alyuda NeuroIntelligence to test the interactions between input and output, to effectively predict the time-to-failure of Electronic Control Unit. The result of the study revealed that the Correlation and R-squared reached 0.999 and 0.997 respectively with Alyuda NeuroIntelligence educational training applied, the accuracy of prediction was 92.45% and the high precision accuracy proves its value of implementation, the study result also shown that Back-Propagation Neural Network can be an standardization of predicting time-to-failure against aircraft critical components.