整合壓電能源擷取與自供電扭力感測技術並透過卷積神經網路進行旋轉環境即時監測

Abstract

本論文主旨為整合能源擷取與感測技術於壓電元件中。其可應用於自供電扭力感測器的開發，並與旋轉機械系統進行整合。透過雙分支輸出神經網路的輔助，此自供電扭力感測器可對時規皮帶進行健康狀態診斷。以下將研究成果分為寬頻能源擷取、自供電扭力感測與人工智慧即時診斷三大部分進行摘要說明。 (寬頻能源擷取)本論文始於探討電致阻尼和能源擷取之輸出功率大小與頻寬之間的關係，而這項研究主要應用於壓電升頻轉換器的開發。其中壓電升頻轉換器的特性為可在低速環境中產生高頻共振。與傳統方法相比，本論文所提出一嶄新升頻轉換器，其機制為透過雙點脈衝磁力以激發雙模態共振。相較於傳統升頻轉換器，嶄新壓電升頻轉換器可提升原本非共振區之輸出功率高達25倍，此結果展示了其顯著的寬頻能源擷取效能。 (自供電扭力感測)本論文為了將賦予扭力感測器能源擷取功能，將壓電懸臂與飛輪進行整合。此扭力感測器利用雙點磁力激振機制，量測飛輪因扭力而產生的相位角變化。扭力感測原理為識別壓電電壓頻譜中模態振幅，與相位角之間的對應關係。然而相同的模態振幅，卻會對應到對稱於臨界相位角的兩個不同相位角，產生多值問題此一無法有效進行扭力感測之情境。因此本論文提出利用同步電荷提取電路中壓電電壓之開路狀態特性，反映出模態振幅的即時變化以解決此問題。最後透過將壓電電壓轉換為時頻譜，使二維卷積神經網路在多值問題發生時進行辨識後，將感測區間進行分段，而此方法已透過實驗驗證。 (人工智慧即時診斷)本論文將自供電扭力感測器，應用於旋轉環境中時規皮帶的健康狀態監測。通過壓電交流電壓的波形變化，將時規皮帶區分為正常、暫態與故障狀態。然而若僅通過時域之交流電壓，難以進行準確的健康狀態辨識。因此本論文提出利用雙分支輸出的卷積神經網絡模型，以提升辨識精度。此模型可在有限的數據下透過監督式學習，對壓電電壓進行高精度辨識，實現對時規皮帶健康狀態的準確診斷。

The thesis endeavors to advance piezoelectric energy harvesting and sensing technology utilizing piezoelectric elements. Its primary focus lies in its application for the creation of self-powered torque sensors and the condition monitoring of timing belts, leveraging a dual-branch output convolutional neural network (CNN). The thesis is structured into three main parts: broadband energy harvesting, torque-sensing identification, and real-time artificial intelligence diagnosis. (Broadband energy harvesting) The study commences by exploring electrically induced damping to enhance broadband capacity and power output. This investigation leads to the development of frequency up-conversion (FUC) harvesters capable of eliciting high-frequency resonance from low-speed environments. Contrasting conventional methods, a novel FUC harvester is engineered to simultaneously induce two distinct resonant modes through a two-point magnetic plucking technique. This approach showcases significant broadband energy harvesting capabilities, resulting in a remarkable output power increase of up to 2500% compared to traditional FUC harvesters in off- resonance regions. (Torque-sensing identification) A torque sensor, incorporating piezoelectric cantilever beam and flywheel, is designed to integrate energy harvesting capabilities. This sensor, employing a two-point magnetic plucking setup, measures torque-induced phase angles on the flywheel. Torque-sensing identification involves detecting changes in the piezoelectric voltage spectrum''s modal amplitude corresponding to induced phase angles from varying torque loads. Challenges arise from multi-valued equal voltages occurring when employing various phase shifts away from critical angles, complicating interpretation in operational scenarios. To address this, a novel approach utilizing SECE (synchronized electric charge extraction) piezoelectric voltage characteristics in the open-circuit state is proposed. Visualization of real-time changes in modal amplitude is achieved by converting the piezoelectric voltage into a spectrogram. Additionally, a 2D- CNN aids in segment identification within the sensing interval, with experimental results demonstrating alignment with predicted outcomes. (Real-time condition monitoring) The device''s application extends to monitoring timing belt conditions in rotating environments, distinguishing healthy, transient, and faulty states through AC piezoelectric voltage fluctuations. However, identifying these states solely via AC voltage in the time domain poses challenges. To address this, a proposed solution involves a dual-branch output CNN. This CNN aims to precisely recognize piezoelectric voltage patterns using limited supervised learning data, enabling accurate identification of belt conditions.