DSpace community: 光機電工程研究所

飛秒雷射叢發模式於薄石英晶圓微鑽孔之動態燒蝕機制研究;Dynamic Ablation Mechanisms in Femtosecond Laser Burst-Mode Microdrilling of Thin Quartz Wafers

title: 飛秒雷射叢發模式於薄石英晶圓微鑽孔之動態燒蝕機制研究;Dynamic Ablation Mechanisms in Femtosecond Laser Burst-Mode Microdrilling of Thin Quartz Wafers abstract: 自十九世紀末居禮兄弟發現單晶石英之壓電特性以來，石英因其優異之物理特性與化學穩定性，長期被廣泛應用於頻率振盪元件。藉由單晶種子於高壓釜中長晶，可獲得高純度且可用體積大之合成石英（Synthetic Quartz），並採用具備優異頻率/溫度特性之 AT-CUT 切割角度作為頻率振盪基板之核心。隨著高頻化與高速運算時代之來臨，元件持續朝精密微型化發展，石英基板之微型鑽孔製程已成為微米等級佈線之關鍵前置作業。然而，單晶石英之硬脆特性使傳統接觸式機械加工於微孔製程逐漸遭遇崩裂與精度之瓶頸，因而需以更高效率及高精度之非接觸式超快雷射精密微鑽孔技術取而代之。本研究以波長1030 nm之Satsuma X-50飛秒雷射系統，採用MHz 叢發模式（Burst Mode）搭配固定焦點之由下而上（Bottom-up）單點鑽孔策略，於純空氣環境中進行 AT-cut 單晶石英晶圓（厚度80 μm）之微孔加工特性研究。本研究藉由高速微距攝影技術，完整觀測微孔逆向燒蝕之動態過程，揭示燒蝕前緣自底面起始、沿光軸逆向推進、貫穿與壓力釋放，以至孔壁修整與順向排屑之完整時序演化，並與傳統 Top-down 配置進行對比，闡明 Bottom-up 配置在避免電漿屏蔽與維持排屑路徑暢通上之機制優勢。實驗進一步指出，在子脈衝能量密度（1.08 J/cm²）低於單晶石英單脈衝燒蝕閾值之條件下，MHz叢發模式之子脈衝間熱累積效應為驅動逆向燒蝕之關鍵機制。實驗結果顯示，於原始參數（平均功率 46.8 W、叢發脈衝數 180、加工時間 1 ms）下，可於單晶石英中形成準直度優異之微孔，其錐角僅約 0.16°。該微孔之深寬比可達約 8，且加工時間僅需 300 μs 即可完成貫穿，展現本製程於加工效率與孔型品質上之優勢。經氫氧化鉀（35 wt.% KOH）蝕刻後處理後，孔壁之重鑄層與熔融碎屑獲得選擇性移除，微孔錐角進一步降至0.08°，孔壁最大高度粗糙度Rz由3.313 μm降至1.986 μm。透過聚焦離子束（FIB）製備試片並以穿透式電子顯微鏡（TEM）搭配選區電子繞射（SAED）進行檢測，確認蝕刻前孔壁前緣之雷射熱影響區為厚度約700 nm 至1 μm 之非晶態改質層，且晶格損傷並未大範圍擴散至基材深處，經 KOH 蝕刻後處理後則大幅縮減至約 150 nm，使孔壁更接近原始單晶結構。此外，本研究進一步將相同製程應用於厚度 100 μm 之單晶石英，成功以陣列形式完成微孔加工，且深寬比可提升至約10，驗證了該製程於不同晶圓厚度下之適用性。本研究證實，於純空氣環境下，MHz 叢發模式之 Bottom-up 鑽孔搭配適當之蝕刻後處理，即可獲得高品質之單晶石英微孔。;Since the Curie brothers discovered the piezoelectric properties of single-crystal quartz in the late nineteenth century， quartz has been widely used in frequency-control devices owing to its excellent physical properties and chemical stability. High-purity synthetic quartz of large usable volume can be grown from single-crystal seeds in an autoclave， and the AT-cut orientation， which offers superior frequency–temperature characteristics， serves as the core substrate for frequency oscillation. With the advent of higher operating frequencies and high-speed computing， devices continue to advance toward precision miniaturization， making the micro-drilling of quartz substrates a key preliminary step for micron-scale interconnection. However， the hard and brittle nature of single-crystal quartz causes conventional contact-based mechanical machining to encounter limitations of edge chipping and precision in micro-hole fabrication， prompting a shift toward non-contact ultrafast-laser precision micro-drilling， which offers higher efficiency and accuracy. In this study， AT-cut single-crystal quartz wafers (80 μm thick) were micro-drilled in an ambient-air environment using a Satsuma X-50 femtosecond laser system at a wavelength of 1030 nm， employing an MHz burst mode combined with a fixed-focus， bottom-up single-point percussion drilling strategy. Using high-speed macro photography， the complete dynamic process of reverse ablation was observed， revealing the full temporal evolution of the ablation front—from initiation at the bottom surface， upward propagation along the optical axis， breakthrough and pressure release， to wall refinement and forward debris ejection. A comparison with the conventional top-down configuration clarifies the mechanistic advantages of the bottom-up scheme in avoiding plasma shielding and maintaining an unobstructed debris-ejection path. The results further reveal that， under a sub-pulse fluence (1.08 J/cm²) below the single-pulse ablation threshold of single-crystal quartz， the inter-pulse heat-accumulation effect of the MHz burst mode is the key mechanism driving reverse ablation. Experimental results show that， under the baseline parameters (average power 46.8 W， 180 pulses per burst， processing time 1 ms)， micro-holes with excellent straightness were formed in single-crystal quartz， with a taper angle of only approximately 0.16°. The micro-holes reached an aspect ratio of about 8， and through-hole formation was achieved in as little as 300 μs， demonstrating the advantages of this process in both machining efficiency and hole quality. After post-processing by potassium hydroxide (35 wt.% KOH) wet etching， the recast layer and resolidified debris on the hole wall were selectively removed， the taper angle was further reduced to 0.08°， and the maximum-height roughness Rz of the hole wall decreased from 3.313 μm to 1.986 μm. Through focused ion beam (FIB) sample preparation followed by transmission electron microscopy (TEM) with selected-area electron diffraction (SAED)， the laser heat-affected zone at the hole-wall front before etching was confirmed to be an amorphous modified layer approximately 700 nm to 1 μm in thickness， with lattice damage not extending extensively into the bulk substrate; after KOH post-processing， this layer was substantially reduced to about 150 nm， bringing the hole wall closer to its original single-crystal structure. Furthermore， the same process was applied to 100 μm-thick single-crystal quartz， successfully producing micro-hole arrays with an aspect ratio of up to about 10， validating the applicability of the process across different wafer thicknesses. This study demonstrates that， in an ambient-air environment， bottom-up drilling in MHz burst mode combined with appropriate post-processing can produce high-quality single-crystal quartz micro-holes.

Human Trajectory Prediction Model: A Lightweight Latent-Variable Framework For Efficient Multi-Agent Motion Forecasting;Human Trajectory Prediction Model: A Lightweight Latent-Variable Framework For Efficient Multi-Agent Motion Forecasting

title: Human Trajectory Prediction Model: A Lightweight Latent-Variable Framework For Efficient Multi-Agent Motion Forecasting;Human Trajectory Prediction Model: A Lightweight Latent-Variable Framework For Efficient Multi-Agent Motion Forecasting abstract: 摘要準確預測行人軌跡，是自主系統在與人類共享的環境中運作時的一項基本需求。本論文提出人類軌跡預測模型（Human Trajectory Prediction Model， HTPM），這是一種用於多智能體行人軌跡預測的機率式潛在變數架構。HTPM 使用具有時間注意力池化的長短期記憶網路（Long Short-Term Memory， LSTM）局部時間編碼器，對每一位智能體的觀測運動歷史進行編碼。情境投影層將所得的智能體嵌入映射為精簡表示，而輕量化全域聚合器則透過跨智能體的注意力池化形成場景層級描述子。條件式潛在模組參數化一個對角高斯先驗，並在訓練期間參數化一個變分對角高斯後驗。取樣得到的潛在碼由自迴歸門控循環單元（Gated Recurrent Unit， GRU）解碼器進行解碼；該解碼器以涵蓋完整預測時間範圍的多步負對數似然目標進行訓練。本模型在 ETH-UCY 基準上進行評估，使用五個標準行人場景：ETH、Hotel、Univ、Zara-1 與 Zara-2。在留一法（leave-one-out）best-of-20 協定下，HTPM 在五個資料分割上的平均 ADE 為 0.1174 m，平均 FDE 為 0.1969 m；場景層級 ADE 的範圍從 Zara-2 的 0.099 m 到 ETH 的 0.145 m。另於 Stanford Drone Dataset（SDD）上進行額外評估，並採用 PECNet/Y-Net 的像素空間前處理慣例（÷20），得到 ADE = 5.97 ± 0.03 pixels 與 FDE = 9.12 ± 0.08 pixels，結果來自 100 次獨立測試執行。這些結果顯示，在所述前處理設定下，隨機取樣具有穩定性；然而，更廣泛的跨資料集泛化仍取決於資料集尺度與攝影機幾何條件。本論文亦提出一套運動感知視覺化流程，結合 YOLOv12x 行人偵測、BYTETracker 身分追蹤與 HTPM 軌跡預測。三層級運動分類器會對靜止智能體抑制預測扇形，對慢速移動智能體繪製精簡扇形，並對步行智能體顯示五假設扇形。ETH Hotel 與 Zara-1 上的定性範例顯示，在所選影格中，該方法能減少視覺雜訊，並改善預測扇形與觀測運動方向之間的對齊。HTPM 含有 516 K 個參數，需要 28 M 次乘加運算（multiply-accumulate operations， MACs），並在 NVIDIA RTX 3090 GPU 上達到每位智能體平均 23 µs 的推論延遲，支持即時桌上型 GPU 推論，並為未來與機器人導航系統整合提供動機。關鍵詞：軌跡預測、行人運動、潛在變數模型、變分推論、LSTM、GRU、ETH-UCY、Stanford Drone Dataset、BYTETracker、運動分類、自主系統。;Abstract Accurate prediction of pedestrian trajectories is a fundamental requirement for autonomous systems operating in shared human environments. This thesis presents the Human Trajectory Prediction Model (HTPM)， a probabilistic latent-variable architecture for multi-agent pedestrian trajectory prediction. HTPM encodes each agent’s observed motion history using a Long Short- Term Memory (LSTM) local temporal encoder with temporal attention pooling. A context projection layer maps the resulting agent embedding into a compact representation， and a lightweight global aggregator forms a scene-level descriptor through attentive pooling across agents. A conditional latent module parameterises a diagonal Gaussian prior and， during training， a variational diagonal Gaussian posterior. Sampled latent codes are decoded by an autoregressive Gated Recurrent Unit (GRU) decoder trained with a multi-step negative log-likelihood objective over the full prediction horizon. The model is evaluated on the ETH-UCY benchmark using the five standard pedestrian scenes: ETH， Hotel， Univ， Zara-1， and Zara-2. Under the leave-one-out best-of-20 protocol， HTPM obtains a mean ADE of 0.1174 m and a mean FDE of 0.1969 m across the five splits， with scene- level ADE ranging from 0.099 m on Zara-2 to 0.145 m on ETH. Additional evaluation on the Stanford Drone Dataset (SDD)， using the PECNet/Y-Net pixel-space preprocessing convention (÷20)， gives ADE = 5.97 ± 0.03 pixels and FDE = 9.12 ± 0.08 pixels over 100 independent test runs. These results indicate stable stochastic sampling under the stated preprocessing settings， while broader cross-dataset generalisation remains dependent on dataset scale and camera geometry. The thesis also introduces a motion-aware visualisation pipeline that combines YOLOv12x pedestrian detection， BYTETracker identity tracking， and HTPM trajectory prediction. A three- tier motion classifier suppresses prediction fans for stationary agents， renders compact fans for slow-moving agents， and displays five-hypothesis fans for walking agents. Qualitative examples on ETH Hotel and Zara-1 illustrate reduced visual clutter and improved alignment between predicted fans and observed motion direction in selected frames. HTPM contains 516 K parameters， requires 28 M multiply-accumulate operations (MACs)， and achieves a mean inference latency of 23 µs per agent on an NVIDIA RTX 3090 GPU， supporting real-time desktop-GPU inference and motivating future integration with robotic navigation systems. Keywords: Trajectory Prediction， Pedestrian Motion， Latent Variable Model， Variational Infer- ence， LSTM， GRU， ETH-UCY， Stanford Drone Dataset， BYTETracker， Motion Classification， Autonomous Systems.

基於無監督遷移學習之鋰離子電池老化預測;Lithium-Ion Battery Aging Prediction Based on Unsupervised Transfer Learning

title: 基於無監督遷移學習之鋰離子電池老化預測;Lithium-Ion Battery Aging Prediction Based on Unsupervised Transfer Learning abstract: 鋰離子電池之健康狀態(SOH)在預測電動車或工業機台的安全性與可靠性時十分重要。然而實務上，要停機獲取充放電數據是不實際的，但在極端溫度與跨設備的情況下又會使得傳統的數據驅動模型之預測因領域偏移而造成預測準確率大幅下降。此外在實務場域中難以取得足夠的標籤數據，要如何運用少量的早期數據來預測電池壽命後期之SOH變化也無可避免地成了實際運用上的難題。為了克服上述困難，本研究使用一套基於CNN+LSTM與AdaBN之無監督遷移學習架構。本論文分為四部分：(1)實驗數據與特徵前處理:利用本實驗室自行收集之不同溫度(25°C、5°C與45°C)下電池充放電數據與NASA開源電池數據集，並使用降採樣與正規化技術來對齊原始的高頻特徵；(2)基準預訓練模型建立:建立CNN+LSTM之基準預訓練模型，以便模型能同時學習單次循環內的充放電特徵和不同循環之間的容量衰退趨勢；(3)AdaBN無監督遷移學習:在完全無須目標域標籤且凍結核心網路權重的情況下，導入AdaBN演算法僅更新批次常態化層(BN層)中的平均值與變異數，將極端溫度或跨數據集之特徵分佈重新對齊至基準預訓練模型熟悉的特徵範圍之中；(4)預測結果與量化驗證:使用遷移指標(TI)等量化評估標準，對比從頭訓練之模型、直接預測模型與經AdaBN遷移後之預訓練模型之間的表現差異。實驗結果證實，25°C基準模型直接預測5°C #8(Ch4)與NASA(B0006)數據時，因嚴重領域偏移導致均方根誤差(RMSE)高達15.64%與11.68%。作為本研究之主要貢獻，所導入之跨溫度無監督AdaBN架構在「完全無需極端溫度標籤」且「免反向傳播」的低運算條件下，將上述RMSE收斂至5.58%與5.01%，並敏銳追蹤極端容量變化與高頻震盪。在多數高變異測試條件下，遷移模型TI值皆顯著小於1（5°C與45°C平均TI為0.577與0.679，NASA B0006更低至0.205）。本研究充分驗證了所使用之AdaBN架構，在面臨跨溫度及跨數據集時具備從頭訓練之模型無可比擬的預測能力，為邊緣運算設備上之即時自適應預估提供了一個高效的解決方案。;Accurate estimation of the State of Health (SOH) of lithium-ion batteries is crucial for ensuring the safety and reliability of electric vehicles and industrial machinery. However， in practical applications， halting operations to acquire charge-discharge data is unfeasible， and extreme temperatures along with cross-device operations often induce severe domain shifts， leading to a significant degradation in the prediction accuracy of traditional data-driven models. Additionally， acquiring sufficient labeled data in real-world environments is notoriously difficult. Effectively utilizing a limited amount of early-stage data to predict SOH variations in the later stages of battery life remains an inevitable practical challenge. To overcome these challenges， this study adopts an unsupervised transfer learning framework based on a Convolutional Neural Network combined with a Long Short-Term Memory network (CNN-LSTM) and Adaptive Batch Normalization (AdaBN). This thesis is structured into four main components: (1) Experimental Data and Feature Preprocessing: Raw high-frequency features from battery charge-discharge data， collected in-house at varying temperatures (25°C， 5°C， and 45°C) and from the open-source NASA battery dataset， are aligned using downsampling and normalization techniques. (2) Baseline Pre-trained Model Establishment: A CNN-LSTM baseline model is pre-trained to simultaneously learn intra-cycle charge-discharge characteristics and inter-cycle capacity degradation trends. (3) AdaBN Unsupervised Transfer Learning: Without requiring target domain labels and while freezing the core network weights， the AdaBN algorithm is introduced to exclusively update the mean and variance within the Batch Normalization (BN) layers， realigning feature distributions skewed by extreme temperatures or cross-dataset variations back into the baseline model′s established feature space. (4) Prediction Results and Quantitative Validation: The performance disparities among models trained from scratch， models using direct prediction， and the pre-trained model enhanced by AdaBN transfer are compared using quantitative evaluation metrics such as the Transfer Index (TI). Experimental results verify that direct prediction of 5°C Ch4 (battery #8) and NASA (B0006) data using the 25°C baseline model yields high Root Mean Square Errors (RMSE) of 15.64 % and 11.68%， respectively， due to severe domain shifts. As the primary contribution of this study， the introduced cross-temperature unsupervised AdaBN framework operating under low-computational conditions requiring "absolutely no extreme-temperature labels" and "no backpropagation” successfully converges the aforementioned RMSEs to 5.58% and 5.01%， while acutely tracking extreme capacity variations and high-frequency fluctuations. Under most high-variance testing conditions， the TI values of the transferred model are significantly less than 1 (averaging 0.577 and 0.679 for 5°C and 45°C， respectively， and reaching as low as 0.205 for NASA B0006). This study comprehensively validates that the adopted AdaBN framework demonstrates predictive capabilities far superior to models trained from scratch when confronted with cross-temperature and cross-dataset scenarios， offering a highly efficient solution for real-time adaptive estimation on edge computing devices.

融合一維與二維電流特徵之輕量化深度學習框架於馬達故障診斷;融合一維與二維電流特徵之輕量化深度學習框架於馬達故障診斷

title: 融合一維與二維電流特徵之輕量化深度學習框架於馬達故障診斷;融合一維與二維電流特徵之輕量化深度學習框架於馬達故障診斷 abstract: 本論文提出一種基於電機電流特徵分析（Motor Current Signature Analysis， MCSA）結合時頻分析與深度學習技術的馬達故障診斷方法。不同於以往僅使用單一輸入形式的研究，本論文系統性地探討三種輸入表示方式：一維（1D）原始電流時序信號、二維（2D）連續小波轉換（Continuous Wavelet Transform， CWT）時頻圖，以及兩者融合的雙輸入（1D+2D）架構，並比較各種輸入方式在不同深度學習模型下的診斷效能。在時頻分析方面，本研究採用以 Morlet 母小波為基礎的 CWT 將電流信號轉換為時頻圖，並採用頻率軸(Y軸)與強度軸(X軸)均以對數尺度表示的顯示模式，以提升時頻特徵的可辨別性。在模型架構方面，本研究比較了多種深度學習模型，包含 VGG-16、ResNet-18、ResNet-50、ViT-Base、ViT-Small，以及本研究提出之輕量化改良模型 Modified VGG16 與結合卷積神經網路（CNN）和視覺轉換器（Vision Transformer， ViT）的混合架構 HybridViT。為驗證方法的可行性與泛化能力，本研究使用兩個資料集進行實驗：公開的 KAIST 永磁同步馬達資料集，以及本研究自行搭建實驗平台所蒐集的 CVR 三相感應馬達資料集，涵蓋正常、內環故障、外環故障、轉子不平衡及軸不對準五種狀態。實驗結果顯示，本研究提出之輕量化模型在維持高診斷準確度的同時，大幅降低模型參數量與記憶體佔用，並於 NVIDIA Jetson Orin Nano 邊緣裝置上完成 ONNX 推論效能評估，確認其具備於資源受限環境下實際部署的潛力。;This thesis proposes a motor fault diagnosis method based on motor current signature analysis (MCSA) combined with time-frequency analysis and deep learning techniques. Unlike previous studies that rely on a single input representation， this work systematically investigates three input strategies: one-dimensional (1D) raw current time-series signals， two-dimensional (2D) scalograms generated via the continuous wavelet transform (CWT)， and a fused dual-input (1D+2D) architecture combining both. The diagnostic performance of multiple deep learning models under each input strategy is comprehensively evaluated. For time-frequency analysis， the Morlet-based CWT is employed to convert current signals into scalograms， with both the frequency axis and magnitude represented on logarithmic scales to enhance the discriminability of fault-related features. In terms of model architecture， this study compares VGG-16， ResNet-18， ResNet-50， ViT-Base， ViT-Small， a proposed lightweight modified model (Modified VGG16)， and a hybrid architecture (HybridViT) that integrates convolutional neural networks (CNN) with vision transformers (ViT). To validate the generalizability of the proposed framework， experiments are conducted on two datasets: the publicly available KAIST three-phase induction motor dataset， and the self-collected CVR three-phase induction motor dataset constructed on a custom-built experimental platform， covering five fault categories: normal， inner race fault， outer race fault， rotor imbalance， and shaft misalignment. Experimental results demonstrate that the proposed lightweight model achieves high diagnostic accuracy with significantly reduced parameter count and memory footprint. Inference performance is further evaluated on an NVIDIA Jetson Orin Nano edge device using ONNX Runtime， confirming the framework′s suitability for deployment in resource-constrained industrial environments.