應用雙向長短期記憶神經網路於異常風機部件之研究

Abstract

本研究分別使用前饋式神經網路[1]、捲積神經網路[2]、長短期記憶神經網路[3]及雙向長短期記憶神網路[4]建立健康狀況診斷演算法，比較這些神經網路並討論這四種神經網路的優劣。再以最佳神經網路尋找適當的神經元數與隱藏層數，建立健康狀況診斷演算法，經比較後採用雙向長短期記憶神經網路(Bidirectional Long Short-Term Memory Neural Networks)建立風機部件健康診斷演算法，並以輸出值與實際值之誤差，定義部件健康指標。最後再以整合移動平均自迴歸模型(Autoregressive Integrated Moving Average Model)所預測之信賴區間設定閾值，以建立故障警訊之標準。 以麥寮12號風機、彰濱工業區2號風機與麥寮11號風機為研究對象，並利用其運作資料與故障維修紀錄分別建立發電機、齒輪箱與液壓系統之風機部件的維護預警系統。經分析後可在至少三天前提出預警。

In this study, the feedforward neural networks [1], convolutional neural networks [2], long-short-term memory neural networks [3] and bidirectional long-short-term memory neural networks [4] were used to establish the health diagnosis algorithms of wind turbine. The used neural networks were compared with their advantages. Then the best neural networks was used to find the appropriate number of neurons and hidden layers, and to establish a health diagnosis algorithm. The Bidirectional Long Short-Term Memory Neural Networks was selected to establish the health diagnosis model for wind turbine component after comparisons. The component health indicators were defined based on the difference between the model output value and the actual value. Finally, the threshold value was set by the confidence interval predicted by the Autoregressive Integrated Moving Average Model to establish the standard of fault warning. The validation examples include Mai-Liao No. 12 wind turbine、Chang-Kong No. 2 wind turbine and Mai-Liao No. 11 wind turbine. The SCADA data and the fault maintenance records were used to establish the early warning system for the generators 、 gearboxes and hydraulic system of the wind turbine, respectively. The validation results show that the early warning can be issued at least three days ago.