Główne pojęcia
本文提出了一種利用數位孿生技術,從系統級狀態監測數據中診斷部件級故障的方法,並探討了該方法在實際應用中的潛力和挑戰。
Statystyki
訓練數據集包含 3600 個數據集,涵蓋 9 種故障模式,每種模式 400 個軌跡。
測試數據集包含 90 個軌跡,涵蓋 9 種故障模式,每種模式 10 個軌跡。
模型訓練使用 Adam 優化器,學習率為 0.0001,批次大小為 32。
模型訓練持續 5000 個時期,以平衡準確性和過擬合風險。
訓練後的模型在訓練數據集上的平均準確率為 98.12%,在驗證數據集上的平均準確率為 92.44%,在測試數據集上的平均準確率為 61.56%。
Cytaty
"The deep learning-based fault diagnosis models, despite of their wide applications and great success, face notable limitations. First, they often require extensive amounts of labeled training data that are difficult to obtain in practice [1]. Second, the majority of deep learning-based models rely on detailed, component-level monitoring data to accurately detect and localize component-level failure [2]."
"In this paper, we attempt to address these two issues by leveraging the high-fidelity simulation and real-time updating capability of digital twins [6]."
"When applied on the real robot, however, the fault diagnosis model trained by the digital twin still faces difficulty in identifying all the failure modes accurately. One of the main reasons is that the digital twin model inevitably has some simulation errors."