為了確保捷運系統的運營安全和可靠性，轉轍器的維保管理至關重要。目前台灣文獻未有提出基於長短期記憶網路(LSTM)和機器學習的方法，為辨識警戒值作為維修人員提醒與追蹤設備運作狀態。 使用機器學習技術來訓練模型，預測電動式轉轍器未來的運營狀態，辨識警戒值進行大數據資料學習對於警戒值的提醒進一步探索是否排程檢修之最佳維護保養時間。 實驗結果顯示，使用長短期記憶網路(LSTM)實驗中，觀測值數之間沒有時間相依性，也就是說它們是獨立且隨機取樣的，那麼預測模型可以更簡單地應用。在這種情況下，我們可以假設每個觀測值都是獨立且來自相同的分佈，並且過去觀測值對未來觀測值沒有影響。在進行預測時，可以使用各種機器學習和統計方法，例如決策樹、支持向量機等。在應用機器學習找出最佳分類法的實驗結果，以 Orange 將訓練資料集進行 5、10 折交叉驗證，單一分類器是 SVM 最佳，而集成分類器是 Random Forest 表現較為出色。對比於應用機器學習預測資料集數量較多(設備多)，kNN 整體評估指標表現最好，其次是 Random Forest。這是因為 SVM 在資料量少時的優勢，具有較高的預測準確率和較好的工程效率，Random Forest 在資料多時的優勢，具有良好的擬合能力和抗擬合能力。 ;To ensure the operational safety and reliability of the metro system, maintenance management of the electric point machines is crucial. Currently, there is no literature available in Taiwan that proposes a method based on Long Short-Term Memory (LSTM) and machine learning to identify warning thresholds for alerting and tracking the operational status of the equipment for maintenance personnel. Machine learning techniques are utilized to train the model and predict the future operational status of the EPMs, as well as identify warning thresholds for further exploration of scheduling maintenance at the optimal maintenance time based on big data learning of these thresholds. The experimental results indicate that there is no temporal dependency between the observed values in the Long Short-Term Memory (LSTM) experiment, implying that they are independent and randomly sampled. In such cases, the prediction model can be applied more simply. Under this assumption, we can consider each observed value to be independent and drawn from the same distribution, with no influence from past observations on future ones. Various machine learning and statistical methods, such as decision trees and support vector machines, can be employed for prediction in such scenarios. n the experimental results of applying machine learning to find the best classification method, the training data set was subjected to 5 and 10-fold cross-validation with Orange. The single classifier is the best for SVM, and the integrated classifier is the best for Random Forest. Compared with the application of machine learning to predict a large number of data sets (more equipment), the overall evaluation index of kNN is the best, followed by Random Forest. This is because SVM has the advantages of higher prediction accuracy and better engineering efficiency when the amount of data is small. The advantage of Random Forest when there is a lot of data is that it has good fitting ability and anti-fitting ability.
