DSpace community: 光機電工程研究所

基於時間序列模型預測控片剩餘壽命;Remaining Useful Life Prediction of Control Wafers Based on Time-Series Models

Fri, 06 Mar 2026 11:03:00 GMT

title: 基於時間序列模型預測控片剩餘壽命;Remaining Useful Life Prediction of Control Wafers Based on Time-Series Models abstract: 控片（Control Wafer）於半導體製程量測中扮演關鍵角色，其薄膜狀態可作為製程與量測系統穩定性之重要指標。然而，隨著時間推移，控片表面薄膜會因氧化效應而產生結構與光學性質變化，進而影響量測結果之準確性與一致性。若能有效預測控片之氧化趨勢與剩餘壽命（Remaining Useful Life, RUL），將有助於提升製程監控效率並降低量測風險。本研究以鈦薄膜控片為研究對象，結合物理模型與機器學習方法，建立一套基於時間序列之控片氧化與剩餘壽命預測架構。首先，透過鈦薄膜氧化模型模擬鈦與氧化鈦厚度隨時間之演化行為，並依據已知薄膜結構之光學理論，生成對應之橢圓偏光光譜。再利用橢圓偏光儀反推等效薄膜厚度與擬合度（Goodness of Fit），建立具物理意義之合成時間序列資料，作為模型訓練與測試之基礎。在模型設計方面，本研究以等效薄膜厚度與擬合度之歷史量測序列作為輸入，分別建構多層感知器神經網路與循環式神經網路模型，進行單步預測以評估模型對即時氧化狀態之掌握能力，並透過多步遞迴預測推估控片之剩餘壽命。此外，針對實務應用中過度預測可能帶來之風險，本研究進一步設計具方向性之懲罰型損失函數，以引導模型降低對剩餘壽命之過度高估行為。實驗結果顯示，相較於多層感知器神經網路，循環式神經網路在多數懲罰倍率設定下，於等效薄膜厚度與擬合度預測誤差上皆呈現較穩定之表現，並能有效利用歷史時間資訊描述薄膜老化之連續性行為。在剩餘壽命預測方面，所提出之懲罰型損失函數可有效抑制過度預測現象，使模型預測結果更符合實務風險評估需求。整體而言，本研究所提出之方法可作為控片氧化監測與剩餘壽命預測之可行解決方案，並具備延伸應用至不同材料與量測情境之潛力。;Control wafers are essential in semiconductor process metrology, as the condition of their thin films serves as a key indicator of process stability and measurement system reliability. However, over time, surface thin films on control wafers undergo oxidation-induced structural and optical property variations, which can degrade measurement accuracy and consistency. Accurate prediction of oxidation behavior and remaining useful life (RUL) is therefore crucial for improving process monitoring efficiency and reducing measurement-related risks. In this study, titanium thin-film control wafers are investigated, and a time-series-based prediction framework is developed by integrating physical modeling with machine learning techniques. A titanium thin-film oxidation model is first employed to simulate the temporal evolution of titanium and titanium oxide thicknesses. Based on optical theories for known thin film structures, corresponding ellipsometric spectra are generated. Thin-film thickness and goodness of fit (GOF) are then extracted through ellipsometric inversion, forming physically meaningful synthetic time-series datasets for model training and evaluation. For model development, historical sequences of thin-film thickness and goodness of fit are used as inputs to construct multilayer perceptron (MLP) and recurrent neural network (RNN) models. One-step-ahead predictions are performed to evaluate the models’ capability to capture instantaneous oxidation states, while multi-step recursive predictions are applied to estimate the remaining useful life of control wafers. To address the practical risks associated with overestimation of remaining useful life, a directional penalty-based loss function is further introduced to guide the models toward more conservative and risk-aware predictions. Experimental results indicate that, compared to multilayer perceptron models, recurrent neural networks exhibit more stable prediction performance for both thin-film thickness and goodness of fit under various penalty factor settings, owing to their ability to effectively utilize historical temporal information and model the continuity of thin-film aging behavior. In terms II of remaining useful life prediction, the proposed penalty-based loss function successfully suppresses overestimation, resulting in predictions that better align with practical risk assessment requirements. Overall, the proposed approach provides a feasible solution for control wafer oxidation monitoring and remaining useful life prediction, and demonstrates strong potential for extension to different materials and metrology scenarios.

基於深度學習之影像辨識技術對皮膚疾病與癌症的辨別

Fri, 06 Mar 2026 10:29:25 GMT

title: 基於深度學習之影像辨識技術對皮膚疾病與癌症的辨別 abstract: 隨著醫療科技的快速發展，人工智慧（AI）與深度學習技術逐漸被應用於輔助診斷領域。皮膚疾病是臨床上最常見的疾病之一，然而在偏遠地區，皮膚科醫師資源分配不均，造成患者無法及時獲得正確診斷。同時，部分皮膚疾病之間外觀相似，非專業醫師或一般患者難以分辨，導致延誤治療。此外，皮膚癌（特別是黑色素瘤）具有高度惡性與致死率，早期辨識對於提升存活率至關重要。本研究以深度學習為核心，旨在建立一套「皮膚疾病與癌症的辨別模型」，期望透過影像自動化分析協助臨床判別，減輕醫療人員負擔並提升診斷效率。研究中使用公開皮膚疾病資料庫共10064張影像，涵蓋五類主要皮膚疾病（含良性與惡性腫瘤）及正常皮膚影像。模型採用EfficientNetV2深度學習架構，並結合線上資料擴增（online data augmentation）策略，於模型訓練過程中即時生成多樣化影像樣本，以提升模型對於不同拍攝條件與影像變異的適應能力。在資料擴增設計上，本研究除採用幾何變換、色彩變換與雜訊擾動外，亦引入 Cutout遮擋擴增方法，透過隨機遮擋局部影像區域，模擬臨床影像中常見的遮蔽與影像缺陷情境，以增強模型對關鍵病灶特徵的學習能力並提升其泛化表現。實驗結果顯示，所建立之模型於測試資料集上可達約83%的分類準確率，顯示其具備良好的辨識能力與潛在臨床應用價值。本研究成果可作為皮膚疾病早期偵測與輔助診斷之基礎，未來若與臨床影像系統整合，有望協助醫療資源不足地區進行初步篩檢，提升診療效率與病患治療成效。 ;With the rapid advancement of medical technology, artificial intelligence (AI) and deep learning techniques have gradually been applied to the field of computer-aided diagnosis. Skin diseases are among the most common clinical conditions; however, in remote areas, the unequal distribution of dermatology specialists prevents patients from obtaining timely and accurate diagnoses. In addition, some skin diseases exhibit similar visual appearances, making them difficult to distinguish for non-specialists or patients, which may lead to delayed treatment. Furthermore, skin cancers—particularly melanoma—are highly malignant and associated with high mortality rates, and early identification is crucial for improving patient survival. This study is centered on deep learning and aims to establish a skin disease and cancer classification model. Through automated image analysis, the proposed approach is expected to assist clinical decision-making, reduce the workload of medical professionals, and improve diagnostic efficiency. A publicly available skin disease dataset containing 10,064 images was used in this study, covering five major categories of skin diseases (including both benign and malignant tumors) as well as normal skin images. The model is based on the EfficientNetV2 deep learning architecture and incorporates an online data augmentation strategy, in which diverse image samples are generated dynamically during the training process to enhance the model’s adaptability to different imaging conditions and variations. In terms of data augmentation design, in addition to geometric transformations, color variations, and noise perturbations, Cutout augmentation was introduced. By randomly masking local regions of the input images, this method simulates common occlusions and image defects encountered in clinical photography, thereby strengthening the model’s ability to learn discriminative lesion features and improving its generalization performance. Experimental results show that the proposed model achieves approximately 83% classification accuracy on the test dataset, indicating strong recognition capability and promising potential for clinical application. The results of this study provide a foundation for early detection and computer-aided diagnosis of skin diseases. With future integration into clinical imaging systems, the proposed model may assist preliminary screening in regions with limited medical resources, ultimately enhancing diagnostic efficiency and patient treatment outcomes.

應用強化學習以機器手臂進行軸孔餘隙配合裝配任務;Application of Reinforcement Learning to Robotic Arm Assembly for Shaft–Hole Clearance Fit Tasks

Fri, 06 Mar 2026 10:29:13 GMT

title: 應用強化學習以機器手臂進行軸孔餘隙配合裝配任務;Application of Reinforcement Learning to Robotic Arm Assembly for Shaft–Hole Clearance Fit Tasks abstract: 本研究探討強化學習（Reinforcement Learning, RL）於組裝任務的應用，實驗平台為七軸機械手臂，以軸孔裝配（Peg-in-Hole）任務作為實驗場景。系統整合視覺與力覺感測器，其中視覺模組負責辨識孔位位置，力覺回饋則協助接觸判斷與阻抗控制，使手臂在組裝過程中具備順應性與穩定操作能力。控制策略以笛卡兒阻抗控制（Cartesian Impedance Control）為基礎，並透過強化學習演算法進行策略優化。本研究比較三種常用演算法——近端策略最佳化（Proximal Policy Optimization, PPO）、深度確定性策略梯度（Deep Deterministic Policy Gradient, DDPG）及軟性行為者評論家（Soft Actor-Critic, SAC）——以評估其在連續控制任務中的學習效率、穩定性與泛化能力。所有訓練皆於 NVIDIA Isaac Sim 模擬環境完成，並建立數位雙生（Digital Twin）模型，完整模擬機器手臂、工件及作業環境的動態與交互關係，確保策略學習可反映實際操作特性。訓練後模型再部署至實體手臂進行 Sim2Real 驗證。實驗結果顯示，PPO 與 DDPG 模型均能成功學習組裝策略，並具備一定的 Sim2Real 遷移能力，其中 PPO 於實體驗證中成功率約為 90%，DDPG 約為 75%。相較之下，SAC 模型在模擬階段未能穩定收斂，雖偶爾能完成插入動作，但學習過程不穩定；於實體測試中完全無法執行成功組裝（成功率為 0%）。此外，在不同手臂剛度設定下的測試結果顯示，PPO 與 DDPG 模型均具良好泛化能力，能在不同孔洞位置與插件初始角度條件下維持穩定性能。研究結果證實，結合強化學習、阻抗控制與數位雙生模擬的 Sim2Real 流程，能有效提升七軸機械手臂在軸孔裝配任務中的精度、穩定性與自適應能力，並降低實體試誤學習所帶來的風險與成本，對自動化組裝與智慧製造之實務應用具有重要參考價值。;This study investigates the application of Reinforcement Learning (RL) in assembly tasks, using a 7-DOF robotic arm as the experimental platform and the Peg-in-Hole assembly task as the test scenario. The system integrates visual and force sensing, where the vision module identifies hole positions and force feedback assists contact detection and impedance control, enabling compliant and stable manipulation during assembly. The control strategy is based on Cartesian Impedance Control and optimized through RL algorithms. Three widely used algorithms—Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), and Soft Actor-Critic (SAC)—were compared to evaluate their learning efficiency, task stability, and policy generalization in continuous control tasks. All training was performed in the NVIDIA Isaac Sim environment, where a Digital Twin fully simulates the dynamics and interactions between the robotic arm, the workpiece, and the operational environment. The trained models were then deployed on the physical robot for Sim2Real validation. Experimental results show that both PPO and DDPG successfully learned assembly strategies and demonstrated certain Sim2Real transfer capabilities. PPO achieved approximately 90% success in physical validation, while DDPG reached about 75%. In contrast, the SAC model failed to converge stably in simulation and completely failed to perform successful assembly in real-world tests (0% success). Additionally, the experimental results under different arm stiffness settings demonstrate that both the PPO and DDPG models exhibit strong generalization capabilities, maintaining stable performance across varying hole positions and different initial peg orientation conditions. These results confirm that a Sim2Real framework combining Reinforcement Learning, Impedance Control, and Digital Twin simulation can effectively improve accuracy, stability, and adaptability of a 7-DOF robotic arm in Peg-in-Hole assembly tasks, while reducing risks and costs associated with real-world trial-and-error learning and providing practical guidance for automated assembly and smart manufacturing applications.

基於智慧型手機之排尿音量測分析以研製等效尿流速裝置研究

Fri, 06 Mar 2026 10:28:50 GMT

title: 基於智慧型手機之排尿音量測分析以研製等效尿流速裝置研究 abstract: 男性在五十歲後，下泌尿道症狀(Lower Urinary Tract Symptoms, LUTS)盛行率約為50～75%，且隨年齡增加而上升。下泌尿道症狀常見成因包含逼尿肌過動、膀胱功能低下與膀胱出口阻塞等。臨床上，尿路動力學為評估排尿功能的重要工具，可量測排尿量、最大與平均排尿流速以及排尿時間等參數，協助醫師判斷排尿狀態。然而診斷方式時常需於醫療院所等封閉式環境進行，或以侵入式方法進行診斷，患者在非自然的環境中接受長時間測試，易產生不適與心理壓力，且檢測結果可能因環境與情緒因素與日常排尿狀況有所差異，導致無法完全反映患者真實情況。為克服上述限制，研究以智慧型手機為診斷平台，建立一套遠端、非侵入式之等效尿流速量測系統。利用手機麥克風錄製尿液撞擊水面聲音，以標準尿流速計作為參考，針對平坦型、滴尿型、斷續型與中斷型四種排尿型態建立等效流速估測算法。量測流程以高通濾波去除低頻環境干擾，並計算聲音訊號之根均方值(Root Mean Square, RMS)曲線；利用頻譜分析選定截止頻率之低通濾波與移動平均平滑方法，使RMS曲線在訊號波形起伏趨勢貼近標準尿流速曲線。透過面積比例法，根據已知排尿總量將聲音訊號轉換為尿流速(ml/s)，完成由音量訊號估測尿流速曲線的等效轉換。考量遠端居家診斷環境，研究建立寬頻衝擊噪音、窄頻衝擊噪音與風扇穩態噪音三種模擬嘈雜環境，探討噪音對尿流音量測之干擾及對診斷指標的影響。利用變分模態分解(Variational Mode Decomposition, VMD)將混合訊號分解為多個窄頻本質模態函數(Intrinsic Mode Function, IMF)，結合分段皮爾森相關係數與能量占比構成之自適應評分函數，自適應挑選最具尿流生理特徵之目標 IMF，以實現單通道盲源分離。本研究於理想環境中，針對四種排尿型態各蒐集十五筆模擬排尿資料，新舊手機皆可重建與標準尿流速計高度相關流速曲線。針對臨床指標，排尿時間估測之決定係數達0.999(新手機)與0.995(舊手機)；最大尿流速估測之決定係數介於0.79～0.85，所有樣本之決定係數約為0.82，顯示本研究所建立之演算程序在排尿起止判定及峰值流速追蹤上皆具良好解釋力。針對曲線相似度評估，以對稱平均絕對百分比誤差(Symmetric Mean Absolute Percentage Error, SMAPE)作為指標，顯示平坦型態因流速變化平順，誤差最低；滴尿與斷續型在高頻起伏誤差較高，但整體波形趨勢仍與標準曲線相符。結果證實本研究之流速估測算法具有跨裝置穩健性，可減少不同手機規格對估測結果的影響。於噪音環境中，未經處理之含噪音訊號往往導致排尿時間被高估或低估，且最大流速峰值被衝擊聲或風扇背景噪音掩蓋。藉由VMD及自適應IMF判定，寬頻、窄頻衝擊噪音與風扇噪音均可被有效去除，模態判定的平均準確率達94.2%，標準差約2.25。多數訊號於降噪後，其流速曲線SMAPE誤差均顯著下降，最大流速與排尿時間之估測可回復至接近理想環境下的準確度，僅少數風扇噪音訊號因模態判定錯誤而未改善，顯示在不同噪音類型與不同排尿型態下，自適應判定方法仍能保持穩定且高準確度的判斷表現。綜上所述，本研究建立一套以聲音為基礎之遠端等效尿流速量測系統，不僅在理想環境下提供與標準尿流速計相當的流速曲線與排尿指標，亦能在模擬居家噪音環境中維持高準確度之流速估測與噪音抑制能力，能夠作為傳統尿流速計之等效替代方案。 ;The prevalence of lower urinary tract symptoms (LUTS) in men over the age of 50 is approximately 50–75% and increases with age. Common causes of LUTS include detrusor overactivity, impaired bladder function, and bladder outlet obstruction. Clinically, urodynamic studies are an important tool for evaluating urinary function. These tests can measure parameters such as voided volume, maximum and average urinary flow rate, and voiding time, helping physicians assess a patient′s urination condition. However, diagnosis often requires testing in controlled clinical environments or through invasive procedures. These settings can be unnatural and lengthy, causing discomfort and psychological stress to the patient. Additionally, results obtained under such conditions may differ from daily urination patterns due to environmental and emotional factors, potentially failing to fully reflect the patient’s actual condition. To overcome these limitations, this study proposes the use of a smartphone-based platform to establish a remote, non-invasive urinary flow rate estimation system. The system records the sound of urine striking the water surface using the smartphone microphone. Using a standard uroflowmetry as a reference, the study developed estimation algorithms for four types of urination patterns: plateau-type, dribbling-type, intermittent-type, and interrupted-type. The measurement process includes high-pass filtering to eliminate low-frequency ambient noise and calculating the root mean square (RMS) of the audio signal. Spectrum analysis is used to determine the cutoff frequency for low-pass filtering and to apply a moving average smoothing method so that the RMS curve closely follows the trend of the standard uroflow curve. Through an area ratio method and known total urine volume, the audio signal is converted into urinary flow rate (ml/s), completing the equivalent transformation of volume signals into a flow rate curve. Considering remote home diagnostic environments, the study simulated three types of noisy conditions: broadband impact noise, narrowband impact noise, and fan-like steady-state noise, to investigate their effects on audio-based flow rate estimation and diagnostic indicators. Variational Mode Decomposition (VMD) was employed to decompose the mixed signals into several narrowband intrinsic mode functions (IMFs). An adaptive scoring function, combining segmented Pearson correlation coefficients and energy ratios, was used to select the target IMF with the most relevant urinary flow features, achieving single-channel blind source separation. In an ideal environment, fifteen simulated urination samples for each of the four urination types were collected. The results showed that both new and old smartphones could reconstruct flow rate curves highly correlated with those from standard uroflowmetry. Regarding clinical indicators, the coefficient of determination (R²) for estimating voiding time reached 0.998 (new phones) and 0.995 (old phones). For maximum flow rate, R² ranged from 0.79 to 0.85, with an overall average of around 0.82, demonstrating the algorithm’s strong performance in identifying the start and end of voiding as well as tracking peak flow rates. For curve similarity evaluation, the symmetric mean absolute percentage error (SMAPE) was used. Plateau-type flows, with smoother rate changes, had the lowest errors. Dribbling and intermittent types had higher errors due to high-frequency fluctuations, but the overall waveform trends remained consistent with the standard curves. These results confirm that the developed estimation algorithm is robust across devices, reducing the impact of varying smartphone hardware on the results. In noisy environments, unprocessed audio signals often led to over- or underestimation of voiding time, and maximum flow rate peaks were masked by impact or fan noise. Using VMD and adaptive IMF selection, all types of simulated noise were effectively suppressed. The average accuracy of IMF mode identification reached 94.2%, with a standard deviation of approximately 2.25. After noise reduction, most samples showed significantly decreased SMAPE errors in their flow rate curves. The estimations for maximum flow rate and voiding time returned to near-ideal accuracy, with only a few fan noise samples remaining uncorrected due to incorrect mode identification. This demonstrates that the adaptive method maintained stable and high-accuracy performance across different noise types and urination patterns. In conclusion, this study developed a remote acoustic-based equivalent urinary flow measurement system. It not only provides flow rate curves and urination indicators comparable to standard uroflowmetry in ideal settings, but also maintains high accuracy in simulated home noise environments. This system has the potential to serve as an effective alternative to conventional uroflowmetry.