Thesis Title: A butterfly stroke recording and analyzing system based on computer vision and machine learning

Graduate Student : Ying-Hsien Chen

Advisor: Chih-Hsien Huang

Abstract:

Motion capture technology using wearable devices and cameras is essential for tracking and recording athletes’ movements. However, most wearable devices cannot operate in the water, and manually analyzing the swimming data from the videos is time-consuming. Hence, this study aims to develop an intelligent swimming recording system that could quantify the swimming results of the butterfly stroke and integrate them into a replay video to offer comprehensive and accurate information to swimmers and coaches. First, this work builds a data collection setup that could simultaneously capture the complete scene of a 25 m track by stitching the videos of 8 cameras (four under and four above the water). After that, this study developed a swimmer recognition model by applying transfer learning to 120,031 labeled images while using YOLOv4 as the pre-trained model. Next, this work used the six bounding box coordinates produced by the recognition model to calculate the split time of 15m/25m, the number of strokes, and the underwater dolphin kicking angles. Then, the accelerometer readings were synchronized using the timing and number of butterfly strokes. Finally, the developed system would integrate the analyzed data, three-axis acceleration readings, and the underwater dolphin kick trajectories into a replay video. This study filmed 85 videos of 25-meter butterfly strokes of five participants to examine the efficacy of the proposed system. The swimmer recognition model achieved an Intersection over Union (IoU) of 81.51%, and the mean absolute error (MAE) of the split time was 0.36 seconds. Besides, the F1-score of identifying strokes was 97.85%, and the Pearson correlation of the kicking angle detection was 0.64. The entire processing time from identification to producing a playback video was four and a half minutes for a 30-second video. Hence, the proposed system should be considered a valuable tool for providing prompt and comprehensive quantified and visual feedback to the swimmers and coaches after each practice.

Thesis Title: Design, simulation, and characterization of a novel optical piezoelectric micromechanical ultrasonic transducer (OpMUT)

Graduate Student : Shao-Wei Wu

Advisor: Chih-Hsien Huang

Abstract:

In this study, to overcome the high impedance and electromagnetic interference susceptibility of piezoelectric micromechanical ultrasonic transducers (pMUTs) due to the use of piezoelectric materials, silicon photonic components, which are less susceptible to electromagnetic interference and have high sensitivity, were combined with pMUTs. The silicon photonic component adopted in this study is the optical ring resonator, which is easier to integrate with pMUTs. The combination of the optical ring resonator and pMUT has been successfully used to study the novel optical piezoelectric micromechanical transducer (OpMUT). The working principle of OpMUT is different from that of pMUT. When the pMUT receives ultrasound, causing the thin film to vibrate, the radius of the ring waveguide in the optical ring resonator changes, thereby changing the light output intensity. The light is then converted into voltage through a photodiode and a transimpedance amplifier circuit. This study uses the finite element method and numerical analysis to optimize the optical ring resonator’s width, radius, and placement. After that, we compare the signal-to-noise ratio (SNR) and noise equivalent pressure (NEP) of the proposed OpMUT with state-of-the-art pMUT at the same resonance frequency. When the width and radius of the ring waveguide are 400 nm and 35.45 µm, the SNR of a 150 kHz OpMUT is about 59 dB @ 1Pa, and the NEP is 0.0007 Pa. It is significantly better than the competitor, who has an SNR of 37 dB @ 1Pa and NEP of 0.0014 Pa.

Thesis Title: Investigation of increasing planar imaging size of the acoustic volumetric display by controlling the edge speed of the suspending object

Graduate Student : Pu-Chun Liu

Advisor: Chih-Hsien Huang

Abstract:

Researchers have recently considered using acoustic levitation as a potential technique for implementing volumetric displays. However, current acoustic volumetric displays can only present small and simple shapes due to the speed limitations and the distortion caused by the instability of the suspended object’s trajectory. Previously, some studies have discussed increasing the imaging size in the vertical plane. Hence, this study aims to raise the imaging size of the horizontal plane by applying an edge deceleration method to increase the average speed of the entire route since the suspending objects usually fly out when changing direction. First, the authors built an acoustic levitation system to realize three kinds of edge deceleration distributions and filmed the object with a high-speed camera. After that, this study identified and analyzed the coordinates of the suspended object to analyze the speed distribution and the degree of distortion. Besides, we captured images with longer exposure time to compare the visual performance of the persistence of vision. Compared to the traditional fixed interval distribution, the maximum frame rate and achievable size for square, triangle, and hexagon patterns increase by 45.5%, 25.1%, and 53.1%, respectively. In parallel with it, the imaging distortion of displaying square, triangle, and hexagon shapes decreases by 73.3%, 81.1%, and 80.5%, respectively. Consequently, the proposed edge deceleration method successfully increased the planar imaging size and quality of the acoustic hologram. A three-dimensional acoustic volumetric display capable of producing large patterns should be feasible by using more transducer arrays and combining different levitating position-selecting methods.

Thesis Title: Investigation of predicting carbon brick temperature of a blast furnace using machine learning approaches

Graduate Student : Yu-Wen Huang

Advisor: Chih-Hsien Huang

Abstract:

The blast furnace is crucial for iron production, and its lifespan directly impacts steelmaking costs. Recently, the erudition speed of carbon bricks in one blast furnace was abnormal and will shorten its lifespan. Hence, this study proposed a machine learning model to address this issue by predicting hearth temperature and finding the critical operation parameters. Besides, because the operation parameters of a blast furnace were numerous (1470 in this study), this work applied Explainable AI (XAI) for feature selection to enhance model performance. First, this research optimized the training setups of the Long Short-Term Memory (LSTM) model and reduced the parameters to 150 through the Pearson correlation. After that, this work proposed a progressive training method to reduce the RMSE of the testing set from 34.2 °C to 17.5 °C. Next, the authors trained prediction models for six temperature sensing positions and calculated the SHAP values of each feature. Then, this work ranked the features based on their SHAP values and retained each prediction model with top-ranked features. The RMSE of the retrained models improved from 17.8 °C, 21.9 °C, 19.4 °C, 25.2 °C, 13.6 °C, and 8.1 °C to 11.1 °C, 15.6 °C, 12.1 °C, 17.7 °C, 6.7 °C, and 5.6 °C. Besides, retraining the model with SHAP-selected features successfully reduced prediction delays from more than predicting days to less than it. In parallel, the correlations between rising and falling periods of practical and predicted temperatures increased, respectively. Finally, this work identified several potential critical parameters by intersecting the SHAP-selected parameters of each sensing position’s prediction model that furnace experts had not previously discussed.

Thesis Title: Investigation of intelligent barbell squat coaching system based on computer vision and machine learning

Graduate Student : Yuhao-Lee

Advisor：Chih-Hsien Huang

The global population has been aging rapidly, leading to an increase in chronic diseases and physical deterioration. Research shows that doing strength training at any age could reduce the impact of these issues. Strength training beginners usually require coaches’ intervention to ensure motion quality and reduce the risk of injury. However, it is essential and unavoidable to practice after coaching classes. Therefore, improving the quality and safety of training in the absence of a coach is critical. This study developed an intelligent barbell coaching system with a real-time mode that could immediately diagnose the issue and provide feedback after each squat. Besides, a replay mode allows users to examine their previous squats and check their comments. First, four key features of the barbell squat were collected, including bending angles of the body, dorsiflexion, ratio of knee/hip movements, and the wobble of the barbell. This work collects 8151 squats from 77 subjects and categorizes them as good squats and six issues. Then, we trained the diagnosis models with three machine-learning architectures, including 1D-CNN, LSTM, and Random Forest. In addition, this research applied the SHAP method to improve the accuracy of predicting each issue and reduce the computation time by feature selection. The F1-Score of the six issues reached 86.86%, 69.01%, 77.42%, 90.74%, 95.83% and 100%, respectively. Besides, the diagnosis time of each squat is less than 0.5 s. Finally, this work examined the efficacy of the proposed system with two groups of subjects trained with and without the system. Consequently, the scores of those trained with the system showed significant improvement, either rated by the system or the professional weightlifting coach.

Thesis Title: Reconstructing 2D gas temperature distribution with deep neural networks

Graduate Student : Chun-Yuan Chen

Advisor：Chih-Hsien Huang

Abstract:

Acoustic tomography is considered a promising technique for visualizing gas temperature distributions. Typically, reconstruction algorithms for acoustic tomography utilize the speed of sound along multiple paths within the region of interest to reconstruct the temperature map. To improve the accuracy of reconstruction, increasing the number of iterations is often necessary, but this process can be time-consuming.

Therefore, this study proposes a convolutional neural network (CNN) to train a machine learning model that can reconstruct gas temperature distributions from sound speed measurements. By adjusting the labels and loss functions within the neural network, six different models were successfully trained to reconstruct gas temperature distributions. Two models were trained using ideal temperature distributions, three using temperature distributions derived from reconstruction algorithms, and one using specific temperature data extracted from ideal temperature distributions. The performance of these six models was evaluated based on temperature distributions from 13 different peak temperature locations, and their ability to track hotspots was compared to existing reconstruction methods. Additionally, the 2D visualization performance was assessed using three different temperature distributions: one similar to the training data and two completely different.

Results indicate that models trained with ideal gas temperature distributions could more accurately track hotspots. Models trained with reconstructed gas temperature distributions demonstrated similar performance to the selected reconstruction algorithms. When the input sound speed differed from the training data, models trained with ideal gas temperature distributions showed poorer results compared to models trained with reconstructed temperature distributions. This suggests that the proposed model can effectively learn the relationship between temperature distributions and sound speed from reconstruction algorithms.

Finally, the execution time of the model was 0.109 seconds, representing a 96% reduction compared to the iterative reconstruction algorithms. Therefore, the proposed reconstruction method should be considered a reliable and efficient approach for 2D gas temperature distribution reconstruction.

論文名稱 : 以機器學習探討爐床碳磚異常侵蝕狀況之對策

研究生 : 吳佳熹

口試委員 : 黃致憲(指導教授)、張天豪(口試委員)、李宜真(口試委員) 、陳建志(口試委員)

摘要-

在工業鐵生產中高爐非常重要。他是鋼鐵持續生產的核心。高爐的使用壽命會直接反映在煉鋼的總成本上。它受多種原因影響，如冷卻板和高爐碳磚的厚度。其中碳磚厚度是最關鍵的。近十年來，將能夠根據即將發生的情況改變高爐操作方式，因此，本研究將提出一個機器學習模型來預測爐床溫度。。中鋼4號高爐碳磚的侵蝕速度異常的快。碳磚的厚度是根據爐床溫度經由特殊公式計算得出。由於鋼鐵產量的需求增加，高爐必須長期保持工作狀態，所以高爐的溫度會逐漸升高，當爐床溫度超過歷史高溫時，代表碳磚厚度變薄了，要在不影響鋼鐵產量下調整溫度是不容易的，如果能提前知道溫度趨勢，

在本研究中，主成分分析(PCA)和皮爾森相關係數用於驗證數據集的特徵數目。PCA可以顯示不同特徵數量的複雜性。此外，可以根據皮爾森係數刪除數據集中的高度相關的特徵。

為了證明預測爐膛溫度的可能性。使用隨機森林訓練一個溫度分類器，可以對爐床溫度進行分類。分類器在測試集中達到95%的準確率。最後，基於LSTM建立了一個爐床溫度模型，利用高爐運行數據預測爐膛溫度。此外，為了優化溫度預測模型，本研究將提出一個新的訓練方法。此外，找出輸入和輸出長度的最佳組合，以減少模型的損失。該模型的MSE為0.00939，而未經訓練以平均溫度計算的基準線為0.0615。

Thesis Title: Investigation of Countermeasures for Abnormal Erosion of Furnace Hearth Carbon Bricks Using Machine Learning

Graduate Student: Chia-Hsi Wu

Oral Defense Committee:

Chih-Hsien Huang (Advisor), Tien-Hao Chang (Committee Member), Yi-Chen Li (Committee Member), Chien-Chih Chen (Committee Member)

Abstract:

In industrial iron production, blast furnaces are critical, serving as the core of continuous steel manufacturing. The lifespan of a blast furnace directly impacts the overall cost of steel production, and it is influenced by various factors, including the thickness of the cooling plates and furnace carbon bricks. Among these, the carbon brick thickness is the most crucial. In recent years, the ability to adjust furnace operations in response to upcoming conditions has become increasingly important. This study presents a machine learning model designed to predict furnace hearth temperature, addressing the issue of rapid carbon brick erosion observed in China Steel’s No. 4 blast furnace.

The thickness of carbon bricks is calculated from furnace hearth temperature using a specific formula. As steel demand increases, the furnace must remain operational for extended periods, causing its temperature to rise gradually. When the furnace hearth temperature exceeds historical highs, it signals a thinning of the carbon bricks. Adjusting the temperature without affecting steel output is challenging, but if temperature trends can be predicted in advance, more efficient measures can be taken.

In this study, principal component analysis (PCA) and Pearson correlation coefficients were used to verify the number of features in the dataset. PCA revealed the complexity of different feature sets, and highly correlated features were removed based on Pearson coefficients.

To demonstrate the feasibility of predicting furnace hearth temperature, a temperature classifier was trained using the random forest algorithm. The classifier achieved 95% accuracy on the test set. Finally, a long short-term memory (LSTM) model was developed to predict furnace hearth temperature using operational data from the blast furnace. To further optimize the temperature prediction model, a new training method was proposed, and the optimal combination of input and output lengths was identified to minimize model loss. The LSTM model achieved a mean squared error (MSE) of 0.00939, compared to a baseline MSE of 0.0615 using untrained average temperature estimates.

論文名稱 : 聲波二維氣體溫度重建演算法之開發與系統實現

研究生 : 蘇瑪

黃致憲(指導教授)、廖育德(口試委員)、李順裕(口試委員) 、戴政祺(口試委員)、鄺獻榮(口試委員)、鄭國順(口試委員)

摘要-

本論文介紹了聲熱成像或斷層成像 (AT) 作為一種可行的非侵入性方法來監測高爐、鍋爐等工業過程中的氣體溫度分佈的調查和分析。溫度被認為是製造企業中用於保證氣體溫度分佈的關鍵參數。高品質的產品線。例如，在高爐中，鋼的質量與爐內氣體溫度的分佈密切相關。因此，監測高溫氣體並跟踪爐內熱點的位置對於提高生產率至關重要。

大多數現有的基於聲學的重建算法都是在相對均勻的溫度分佈下開發的。他們跟踪給定區域的峰值溫度或熱點的能力很少被討論。因此，在本研究中，提出並演示了一種專用於高度集中氣體溫度分佈的聲溫層析成像重建方法。所提出的方法可以有效地可視化和測量峰值溫度，準確度為 5.89%。它還可以跟踪熱點向新位置的移動。

AT使用平面上多條路徑的聲速來重建二維氣體溫度分佈。因此，對於溫度圖的最終重建，重建方法和正確的聲速估計都是至關重要的。然而，由於缺乏關鍵信息描述，例如估計飛行時間 (TOF) 或聲速的方法，現有文獻中報告的結果難以重現。因此，在本研究中，提出並演示了另一種提高用於測量聲速的聲學平台精度的新技術。使用所提出的方法，在具有臨界溫度分佈的環境中實現了 ±3% 到 ±5% 的聲速精度。隨後，通過採集的聲速重建的二維溫度圖也與熱電偶陣列的結果相匹配。

Thesis Title: Development and System Implementation of a Two-Dimensional Gas Temperature Reconstruction Algorithm Using Acoustic Waves

Graduate Student: Ma Su

Oral Defense Committee : Chih-Hsien Huang (Advisor), Yu-Te Liao (Committee Member), Shun-Yu Li (Committee Member),

Cheng-Chi Dai (Committee Member), Hsiang-Rong Kwang (Committee Member), Kuo-Shun Cheng (Committee Member)

Abstract:

This thesis introduces acoustic tomography (AT) as a viable non-invasive method for monitoring gas temperature distributions in industrial processes such as blast furnaces and boilers. Temperature is a crucial parameter for ensuring the quality of products in manufacturing enterprises. For instance, in blast furnaces, the quality of steel is closely related to the distribution of gas temperature within the furnace. Monitoring high-temperature gases and tracking the location of hotspots is essential for improving productivity.

Most existing acoustic-based reconstruction algorithms are developed under the assumption of relatively uniform temperature distributions, with limited discussion on their capability to track peak temperatures or hotspots. This study proposes and demonstrates a specialized acoustic tomography reconstruction method tailored for highly concentrated gas temperature distributions. The proposed method can effectively visualize and measure peak temperatures with an accuracy of 5.89% and track the movement of hotspots to new locations.

AT reconstructs two-dimensional gas temperature distributions using the speed of sound along multiple paths on a plane. Thus, accurate reconstruction of the temperature map relies on both the reconstruction method and precise speed of sound estimation. However, due to the lack of critical information, such as methods for estimating time-of-flight (TOF) or sound speed, the results reported in existing literature are difficult to reproduce. Therefore, this study proposes and demonstrates a new technique to enhance the accuracy of acoustic platforms used for sound speed measurement. Using the proposed method, sound speed accuracy of ±3% to ±5% was achieved in environments with critical temperature distributions. Subsequently, the two-dimensional temperature maps reconstructed using the measured sound speeds matched the results from thermocouple arrays.

Thesis Title: A Novel ADE System with Tunable Droplet Size and Ejection Height

Graduate Student : Shih-Hung Shen

Advisor：Chih-Hsien Huang

Abstract:

Acoustic droplet ejection (ADE) has been proven to move liquids in droplet form from a container to mid-air. However, the droplet sizes produced using the traditional ADE setups are proportional to their heights if the physical framework is unchanged. This limitation obstructs further employment of ADE technology since having the exact sizes of droplets at a specific location or different sizes at different locations is required for many applications. To overcome this limitation and enable more possible applications of ADE, this study proposed an innovative ADE configuration that could manipulate the size and height of the ADE droplets with only electrical signals. To achieve this, a low-voltage driving period and a pinhole structure were added to create a water mound before ejection. First, simulations were conducted to validate the proposed method and find the parameters of the novel ADE setup. After that, a driving circuit featuring a high-voltage pulser and a field-programmable gate array was built. Besides, a 3D spherical resin model was printed to focus the acoustic wave on the water surface, and a cover with a pinhole was added to create water mound. To observe the behavior of the droplets, a recording system and detection algorithm were developed to capture and identify the dimension/height of the droplets, respectively. Finally, the proposed ADE configuration successfully manipulated the droplet size at the same ejecting height under three driving voltages (65 V, 70 V, and 75 V) and pinhole diameters (3.7 mm, 4 mm, and 4.4 mm).

論文名稱 :適用於超音波觸覺之壓電式微型超音波傳感器的初步探討

研究生 : 蘇芯儀

口試委員 : 黃致憲(指導教授)、李永春(口試委員)、楊慶隆(口試委員) 、薛漢鼎(口試委員)

摘要-

超音波觸覺為一種觸覺反饋技術，利用高強度超音波的聚焦刺激人體皮膚以形成觸覺，其具有高解析度與非接觸式等優點，適合與VR/AR技術或3C產品結合。

目前超音波觸覺技術可透過傳統傳感器組成的模組來實現，然而現有模組因傳感器體積大，不易與可攜式產品結合，因此本研究提出使用壓電式微型超音波傳感器(pMUT)取代傳統傳感器，不僅可以大幅縮小整體尺寸，且pMUT相較電容式微型超音波傳感器(cMUT)最大優勢在於不需要直流偏壓即可驅動，功耗相對較低，適合與可攜式產品結合。

本研究透過圓盤分析模型與有限元素模型進行模擬，以設計響應頻率為40kHz時pMUT之尺寸，結構部分選用鋯鈦酸鉛(PZT)作為pMUT之壓電層，並嘗試射頻濺鍍法、脈衝雷射沉積法及溶膠凝膠法三種製備方法，透過調整退火參數製備出具有良好結晶情形及壓電特性的PZT，再將此PZT薄膜應用於整個pMUT製程中，以絕緣層上覆矽(SOI)作為基板與pMUT結構層，依序沉積下電極、PZT、上電極，最後進行背蝕刻形成空腔。

本研究製備出的pMUT經雷射都卜勒振動儀(LDV)分析其響應頻率，接近目標頻率40kHz，對單顆pMUT輸入脈衝驅動訊號，藉由高頻麥克風量測到0.227Pa的聲壓，透過陣列模擬可以輸出足以產生觸覺之聲壓，證明此pMUT未來有機會應用於超音波觸覺技術中。

Thesis Title: Preliminary Study of a Piezoelectric Micromachined Ultrasonic Transducer (pMUT) for Ultrasonic Haptics

Graduate Student: Hsin-Yi Su

Oral Defense Committee:

Chih-Hsien Huang (Advisor), Yung-Chun Lee (Committee Member), Ching-Long Yang (Committee Member), Han-Ding Shueh (Committee Member)

Abstract:

Ultrasonic haptics is a tactile feedback technology that uses focused high-intensity ultrasound to stimulate human skin, creating a sensation of touch. It offers advantages such as high resolution and non-contact interaction, making it suitable for integration with VR/AR technologies or 3C products.

Current ultrasonic haptic technologies are typically realized through sensor modules composed of traditional transducers. However, these modules are bulky and difficult to integrate with portable devices. This study proposes the use of piezoelectric micromachined ultrasonic transducers (pMUT) to replace traditional transducers, significantly reducing the overall size. A major advantage of pMUTs over capacitive micromachined ultrasonic transducers (cMUTs) is that pMUTs do not require a DC bias for operation, resulting in lower power consumption and better compatibility with portable products.

In this research, both disk analysis models and finite element models were employed to simulate the design of a pMUT with a resonant frequency of 40 kHz. Lead zirconate titanate (PZT) was selected as the piezoelectric layer for the pMUT structure. Three fabrication methods were explored: RF sputtering, pulsed laser deposition, and sol-gel deposition. Through adjustments to the annealing parameters, PZT films with excellent crystallinity and piezoelectric properties were successfully prepared. These films were then applied in the full pMUT fabrication process. The device utilized a silicon-on-insulator (SOI) wafer as the substrate and structure layer, with sequential deposition of the bottom electrode, PZT, and top electrode, followed by back-etching to create a cavity.

The fabricated pMUT was analyzed using a laser Doppler vibrometer (LDV), which confirmed a resonant frequency close to the target of 40 kHz. When driven by a pulse signal, the pMUT generated a sound pressure of 0.227 Pa, as measured by a high-frequency microphone. Simulations of pMUT arrays demonstrated the ability to generate sound pressure sufficient to produce tactile sensations, indicating potential applications of pMUTs in ultrasonic haptic technology.

論文名稱 :適用於二維氣體溫度重建演算法開發之超音波測溫平台

研究生 : 謝淨川

口試委員 : 黃致憲(指導教授)、李永春(口試委員)、楊慶隆(口試委員) 、薛漢鼎(口試委員)

摘要-

聲學溫度計(Acoustic Thermometer)是一種以非侵入性方式量測待測空間內氣體溫度分佈的技術,但目前多以模擬的聲速資料進行演算法的研發。

本研究主要設計一種硬體架構,提供相關研發人員一個實驗平台,研究相關的演算法與硬體設備。本論文敘述了整體實驗之架設,零件與電路設計上的考量,以及測試不同的電路得到較高的訊號品質。透過屏蔽電纜傳遞聲波,可提升 1.8 倍的訊雜比,以及使用本研究所設計的隔離式電源板,可提升 5.3 倍的訊雜比。本實驗環境使用 12 組超音波傳感器進行聲波收發實驗,藉由本研究設計的收發電路使同一組傳感器可發射或接收聲波,聲波接收角度最大可達正負150 度。

最後利用此測溫平台進行測溫實驗,使用本實驗室開發的演算法重建出待測範圍之溫度分佈圖,再藉由外接 37 組熱電偶進行本測溫系統的驗證,根據不同的熱源實驗,其兩者測溫結果皆可匹配。

Thesis Title: Development of an Ultrasonic Temperature Measurement Platform for 2D Gas Temperature Reconstruction Algorithms

Graduate Student: Jing-Chuan Hsieh

Oral Defense Committee:

Chih-Hsien Huang (Advisor), Yung-Chun Lee (Committee Member), Ching-Long Yang (Committee Member), Han-Ding Shueh (Committee Member)

Abstract:

An acoustic thermometer is a non-invasive technique used to measure the gas temperature distribution in a target space. However, most algorithm development in this field relies on simulated sound speed data.

This research focuses on designing a hardware platform that provides researchers with an experimental setup for studying related algorithms and hardware components. The thesis details the experimental setup, considerations for component and circuit design, and testing of different circuits to achieve higher signal quality. By using shielded cables to transmit sound waves, the signal-to-noise ratio (SNR) improved by 1.8 times. Additionally, with the custom-designed isolated power supply board, the SNR increased by 5.3 times. The experimental environment utilizes 12 sets of ultrasonic sensors for sound wave transmission and reception. The designed transmission and reception circuit allows the same sensor to both transmit and receive sound waves, with a maximum receiving angle of ±150 degrees.

Finally, temperature measurement experiments were conducted using this platform, and the temperature distribution within the target area was reconstructed using an algorithm developed by the laboratory. The system’s accuracy was verified with 37 external thermocouples. The results from the thermocouples matched well with the temperature measurements obtained from the system, across various heat source experiments.

論文名稱 :超音波影像辨識於智慧白蝦養殖應用之初探

研究生 : 楊伯威

口試委員 : 黃致憲(指導教授)、鄺獻榮(口試委員)、戴政祺(口試委員) 、黃世杰(口試委員) 、張天豪(口試委員)

摘要-

超音波影像技術在醫學領域上已被廣泛研究,卻鮮少應用在農漁養殖業上,但實際上超音波為藉由介質傳送之機械波,且在水中之衰減速度甚低,在養殖場域中具能穿過懸浮微粒且對養殖物影響極小,對比一般光學影像,超音波的成像機制幾乎不受光線、水色及混濁度的影響,在智慧水產養殖領域中極具發展優勢。

本論文提出一個將人工智慧技術應用於白蝦之超音波影像辨識系統,旨在利用超音波技術提供過去因環境因素無法獲得之影像,並搭配人工智慧自動地判斷出區域內白蝦的數量、位置與長度,以提供給養殖業者促進其改善投餌策略,達到智動化,高報酬率、友善環境、等高科技永續水產養殖的目標。

本研究中使用 YOLOv4 之物件偵測模型來訓練白蝦之超音波影像,並探討了使用不同資料集訓練之模型對白蝦影像辨識之影響。此外利用此模型進行偵測區域影像中白蝦之位置,並藉由 Spectral clustering 分群演算法提出的評分演算法處理模型偵測結果,最終能成功地得到該區域內白蝦的數量、位置與長度。根據實驗結果統計, YOLOv4 模型在測試集上的 mAP 為 98.93%,測試平均 IOU 達到 78.47%。此外該系統在預測白蝦之數量準確度能夠達到 97.3%,且最終預測的白蝦位置與真實情況之平均誤差為 0.97cm ,同時平均長度誤差為 8.99%,成功的展現了此系統在實際應用上之可能性。

成大圖書館連結 : 尚無

Thesis Title: Preliminary Exploration of Ultrasound Image Recognition for Smart White Shrimp Aquaculture Applications

Graduate Student: Bo-Wei Yang

Oral Defense Committee:

Chih-Hsien Huang (Advisor), Hsien-Jung Kuang (Committee Member), Cheng-Chi Tai (Committee Member), Shih-Chieh Huang (Committee Member), Tien-Hao Chang (Committee Member)

Abstract:

Ultrasound imaging technology has been extensively studied in the medical field but is rarely applied in agriculture and aquaculture. In reality, ultrasound is a mechanical wave transmitted through a medium and has low attenuation in water. This allows it to pass through suspended particles with minimal impact on the organisms in aquaculture environments. Compared to optical imaging, ultrasound imaging is nearly unaffected by light, water color, or turbidity, offering significant advantages for development in smart aquaculture.

This thesis proposes an artificial intelligence-based ultrasound image recognition system for white shrimp. The goal is to leverage ultrasound technology to provide images that were previously unattainable due to environmental factors, and to automatically determine the number, location, and length of white shrimp within the area. This system aims to help aquaculture farmers optimize feeding strategies, achieving intelligent, high-efficiency, environmentally friendly, and sustainable aquaculture practices.

The YOLOv4 object detection model was employed to train on ultrasound images of white shrimp, and the study examined the impact of different training datasets on shrimp image recognition. The model was used to detect the location of shrimp in the target area, and a scoring algorithm based on Spectral Clustering was applied to process the detection results. Ultimately, the system was able to successfully determine the number, location, and length of white shrimp in the area. According to experimental results, the YOLOv4 model achieved a mean average precision (mAP) of 98.93% on the test set, with an average Intersection over Union (IOU) of 78.47%. Additionally, the system was able to predict shrimp quantities with an accuracy of 97.3%. The predicted positions of the shrimp had an average error of 0.97 cm, and the average length error was 8.99%, demonstrating the potential of this system for practical applications in aquaculture.

論文名稱 :應用於穿戴式光聲影像裝置之可撓式透明聲波感測元件

研究生 : 陳立祥

口試委員 : 黃致憲(指導教授)、李永春(口試委員)、楊慶隆(口試委員) 、薛漢鼎(口試委員)

摘要-

光聲影像(Photoacoustic imaging)是一種基於光聲效應來產生影像的方式,利用特定的波長之雷射脈衝光,瞬間加熱生物組織,使其發生收縮及膨脹,從而產生出超音波,再利用超音波傳感器將接收到的超音波訊號轉換成電訊號,最後依據光束位置或傳感器陣列分布,以及聲波之聲速與傳遞時間,將所接收之訊號重建成 2D 或 3D 影像。

當前的光聲影像系統,由於傳感器採用的壓電薄膜及金屬電極大多為不透光的材料,造成光源的路徑必須避開超音波傳感器,導致整體的系統不易微縮,若要發展成穿戴式相關的系統上,需克服較多的困難。因此,本研究針對超音波傳感器的部分進行改良,使用聚偏二氟乙烯(PVDF)為壓電材料開發透明可撓式超音波傳感器,電極部分則是採用銦錫氧化物(ITO)。本研究產出之傳感器具有 86.51%的 3dB 頻寬,以及6.831MHz 之中心頻率,縱向解析度和系統橫向解析度分別為 265μm 以及 400μm,並且將傳感器維持長時間的彎曲狀態,結果顯示於合理的彎曲範圍內,傳感器可進行長時間的彎曲工作。因兩種材料都具有可透光特性,故雷射光可直接穿透傳感器照射待測物體,使的雷射光源、傳感器及欲成像之組織可設置於同一路徑上,大幅提升將光聲影像系統整合至穿戴式裝置的可能性。

本研究採用波長 532nm 的 Nd:YAG 雷射光進行傳感器之成像效能測試,掃描方法採用單點雷射光掃描,訊號接收的方式採用單一傳感器進行接收,將接收到訊號利用放大器放大後,傳送至數位類比訊號轉換模組中,並儲存於電腦,最後,將儲存到的數據重建成二維或三維影像,再將成像圖與自行設計之模型相互比較,於水中成像的深度模型其縱向及橫向位置誤差分別為 0.94%和 1.66%,而可達的成像深度於水中和環氧樹脂模型裡,分別為 36mm 以及 15mm。實驗證明本研究所開發之透明可撓式超音波傳感器,能夠以高靈活度貼於人體組

織表面進行長時間的生理訊號偵測,且能用於淺層靜脈血管成像,具有可開發成穿戴式裝置的潛力。

Thesis Title: Flexible Transparent Acoustic Sensor for Wearable Photoacoustic Imaging Devices

Graduate Student: Li-Hsiang Chen

Oral Defense Committee : Chih-Hsien Huang (Advisor), Yung-Chun Lee (Committee Member), Ching-Lung Yang (Committee Member), Han-Ting Hsueh (Committee Member)

Abstract:

Photoacoustic imaging (PAI) is a technique that generates images based on the photoacoustic effect. It utilizes laser pulses of specific wavelengths to rapidly heat biological tissues, causing them to expand and contract, which in turn generates ultrasound waves. These ultrasound waves are then converted into electrical signals by sensors, and the signals are used to reconstruct 2D or 3D images based on the position of the laser beam, the distribution of the sensor array, and the speed and travel time of the sound waves.

Current PAI systems face limitations because the piezoelectric films and metal electrodes used in traditional ultrasound sensors are often opaque, necessitating the laser source to avoid the sensor, complicating miniaturization efforts. This poses significant challenges when developing wearable systems. Therefore, this research focuses on improving the ultrasound sensor by developing a flexible, transparent sensor using polyvinylidene fluoride (PVDF) as the piezoelectric material and indium tin oxide (ITO) for the electrodes. The developed sensor achieved an 86.51% 3dB bandwidth and a center frequency of 6.831 MHz, with a longitudinal resolution of 265μm and a lateral resolution of 400μm. The sensor was tested under prolonged bending conditions, showing stable performance within reasonable bending limits. Due to the transparency of both PVDF and ITO, the laser can directly pass through the sensor, allowing the laser source, sensor, and target tissue to be aligned along the same path. This significantly enhances the feasibility of integrating PAI systems into wearable devices.

A 532nm Nd

laser was used to test the imaging capabilities of the sensor. A single-point laser scanning method was employed, with a single sensor receiving the signals. These signals were amplified and sent to a digital-to-analog converter before being stored on a computer for 2D or 3D image reconstruction. The reconstructed images were compared with a custom-designed model, showing longitudinal and lateral positioning errors of 0.94% and 1.66%, respectively, for imaging in water. The imaging depth reached 36mm in water and 15mm in epoxy resin models. The experimental results demonstrate that the developed flexible, transparent ultrasound sensor can adhere to human tissue surfaces with high flexibility for long-term physiological signal monitoring and can be applied for shallow vein imaging, showing potential for wearable device development.

論文名稱 :深度學習於即時超音波手勢辨識系統之應用

研究生 : 許祐齊

口試委員 : 黃致憲(指導教授)、鄺獻榮(口試委員)、戴政祺(口試委員) 、黃世杰(口試委員)、張天豪(口試委員)

摘要-

手勢辨識(Gesture Recognition)是一種可以透過手部動作來代替傳統輸入裝置來進行人機互動的技術。而現今多數非接觸式手勢辨識系統是使用視覺性手勢辨識,而這項技術會因為照明變化而受到嚴重影響,也因為需要使用圖像擷取設備而被大眾擔心隱私權會受到侵害。而超音波手勢辨識不會受到上述的限制,但是現有之超音波手勢辨識多基於都卜勒效應,故多使用於動態手勢辨別。因此本研究將會提出以時頻分析生成時頻圖進行超音波靜態手勢辨識之方法。

此研究使用四顆市售的 40kHz 超音波傳感器來進行訊號發送與資料接收,之後利用數位類比轉換器將超音波訊號轉為數位訊號,再經由短時距傅立葉轉換成時頻訊號強度圖,最後建立兩個深度模型進行訓練並選擇較適合之模型。而文中也透過更改模型的各項參數來進行最佳化,最終讓辨識手勢的準確率提升了約 5%,達到了98.07%。而本研究也透過整合多個軟體來搭建即時手勢辨識系統,並將最佳化模型套用於即時手勢辨識系統也能成功運作,系統上的準確率也達到了 96%,證明了超音波應用於靜態手勢辨識之可行性。

Thesis Title: Application of Deep Learning in Real-Time Ultrasonic Gesture Recognition Systems

Graduate Student: Yu-Chi Hsu

Oral Defense Committee : Chih-Hsien Huang (Advisor), Hsien-Jung Kuang (Committee Member), Cheng-Chi Tai (Committee Member),

Shih-Chieh Huang (Committee Member), Tien-Hao Chang (Committee Member)

Abstract:

Gesture recognition is a technology that enables human-computer interaction using hand movements as an alternative to traditional input devices. Most non-contact gesture recognition systems today rely on vision-based techniques, which can be significantly affected by changes in lighting conditions and raise privacy concerns due to the use of image capture devices. Ultrasonic gesture recognition, on the other hand, is not constrained by these issues. However, existing ultrasonic gesture recognition systems often rely on the Doppler effect and are primarily used for recognizing dynamic gestures. This research proposes a method for recognizing static gestures using time-frequency analysis to generate spectrograms from ultrasonic signals.

In this study, four commercially available 40kHz ultrasonic sensors were used for signal transmission and data reception. The received ultrasonic signals were converted to digital signals using an analog-to-digital converter, and then transformed into time-frequency intensity images using short-time Fourier transform (STFT). Two deep learning models were developed and trained, and the most suitable model was selected. By optimizing various parameters of the model, the gesture recognition accuracy was improved by approximately 5%, reaching 98.07%. The research also integrated multiple software components to build a real-time gesture recognition system, applying the optimized model to the system. The real-time system achieved an accuracy of 96%, demonstrating the feasibility of applying ultrasonic technology for static gesture recognition.

研究生 : 李承祐

口試委員 : 黃致憲(指導教授)、鄺獻榮(口試委員)、戴政祺(口試委員) 、黃世杰(口試委員)、張天豪(口試委員)

Summary-

超音波聲速藉由一端傳感器發射超音波再由另一端傳感器接收聲波,計算從發射聲波的時間點到接收聲波的時間點得知飛行時間(Time-of-Flight,以下簡稱 TOF),並利用已知的距離與 TOF 推估可得到聲波在該介質中的傳遞速率。超音波需要透過介質來傳遞,然而在空氣中,聲速可能被環境因素如溫度、濕度、大氣壓力所影響,導致推算平均聲速結果誤差大。現今已經有的聲速判斷方法如脈衝最大值以及訊號調變包絡方波法,經 TOF 分析過後得知聲速,但並未考量溫度不均勻的情況下,聲波會因為聲阻抗之不同產生的反射和折射使波形偏移,讓 TOF 產生誤差,導致聲速推算錯誤。為了提高聲速估計的空間分辨率,如何在不受真實環境影響下提高聲速計算的準確性正是需要克服的難題。

在真實環境中,聲波傳遞過程中可能因為衰減、反射以及回波干擾等因素,導致聲速判斷不準確,因此在本研究透過演算法以及實驗架構的改善增加聲速判斷準確性。

在實驗架構中,利用架設 12 個超音波傳感器形成一個圓形區域,對於圓中心而言,每間隔 30°就放置一個傳感器,彼此間距離相等且互相皆可當作發射器與接收器,並且在四周圍都放上吸音棉用以降低回波反射影響主訊號。在量測路徑不同時,可以藉由演算法得到訊號到達的接收器的時間點來進行 TOF 的計算,然後推算各路徑上的聲速。

本研究透過超音波傳感器發射與接收的訊號進行分析,找出在非均勻氣體溫度分佈下的能夠正確判斷聲速的方法。根據系統得到之聲波訊號,不受聲波傳遞經過溫度分佈不同的區域產生的反射與折射所造成之干涉的影響,正確計算出該路徑之聲速。當接收器收到聲波訊號,此時訊號就會產生脈衝波形,利用雜訊強度的大小以及Schmitt-trigger 的概念並加入判斷式,使訊號分析在各路徑上的接收訊號中得到主訊號接收時間點,使聲速計算更加準確且適用於溫度分佈不均勻的氣體中量測。

Thesis Title: Development and Verification of an Acoustic Velocity Algorithm for Non-Uniform Gas Temperature Distributions

Graduate Student: Cheng-Yu Lee

Oral Defense Committee : Chih-Hsien Huang (Advisor), Hsien-Jung Kuang (Committee Member), Cheng-Chi Tai (Committee Member),

Shih-Chieh Huang (Committee Member), Tien-Hao Chang (Committee Member)

Abstract:

Ultrasonic velocity is determined by transmitting sound waves from one sensor and receiving them at another, calculating the time-of-flight (TOF) from the transmission to reception. By knowing the distance and TOF, the propagation speed of sound in the medium can be estimated. Sound velocity in air can be influenced by environmental factors such as temperature, humidity, and atmospheric pressure, leading to inaccuracies in the average speed calculation. Current methods like pulse peak and signal modulation envelope approaches calculate sound speed using TOF, but they don’t account for temperature non-uniformity, which causes reflections and refractions due to differences in acoustic impedance, shifting waveforms and introducing TOF errors.

This research aims to improve spatial resolution for velocity estimation by addressing the challenge of calculating sound speed accurately without being affected by real-world environmental variations. In real scenarios, sound waves may experience attenuation, reflection, and echo interference, which hinder accurate sound velocity determination. Therefore, this study improves accuracy through algorithmic adjustments and an enhanced experimental setup.

The experimental setup includes 12 ultrasonic sensors arranged in a circular area, with each sensor spaced 30° apart from the center. Each sensor can function as both a transmitter and receiver. Acoustic foam was placed around the sensors to reduce echo reflections. By measuring the time at which signals reach the receivers, the TOF for each path was calculated, allowing for the computation of sound velocity along those paths.

The study successfully identified methods for accurately determining sound velocity in gas with non-uniform temperature distribution. The algorithm minimized interference from reflections and refractions due to temperature differences, enabling the correct calculation of sound speed for each path. By analyzing the signal using noise intensity and applying a Schmitt-trigger-based decision criterion, the method effectively identified the main signal arrival time, improving the accuracy of sound velocity measurements, particularly in environments with uneven temperature distribution.

聲學斷層掃描（Acoustic Tomography）是一種以非侵入性方式獲取目標區域的訊息的技術。在監控系統中目標區域的溫度訊息常常是監控中不可或缺的部分。而使用聲學斷層掃描的方式可以在不影響目標區域的狀況下重建出目標區域的二維溫度分佈圖。其中關鍵的挑戰是如何提高相對空間分佈和絕對溫度值的匹配準確性。

利用超音波測溫的方式，可以藉由聲波傳遞速度與介質溫度之關係，計算目標區域內的介質溫度。使用超音波陣列在目標區域內產生數十條超音波傳遞路徑，並根據所測得之各路徑的聲波速度，推算出二維平面上各點氣體的溫度。

本研究通過分析在二維平面上不同的網格形狀和超音波路徑的影響，提出了一種根據多個聲波傳導路徑重建溫度分佈的演算法並在MATLAB中實現。此演算法主要使用了空間之移動平均技術並使用六角網格進行溫度分佈之重建。對網格形狀與大小之選擇，本研究進行了數個模擬測試，共使用三種網格形狀（六角形分佈，正方形分佈和極座標分佈）分析重建溫度分佈的準確性。根據研究結果，發現六角形網格可以最準確地追蹤熱點的峰值溫度以及移動。在理想模型上可達所有聲速路徑上之平均溫度誤差小於1%，並可在熱點於正中心往外移動之情況下，感測到相對區域直徑3%之熱點移動，同時最高溫誤差小於10%。

Paper Title: Development of 2D Ultrasonic Temperature Distribution Reconstruction Algorithm

Postgraduate: Lin Fusong

Oral examination committee members: Huang Zhixian (advising professor), Lin Zhilong (oral examination committee member), Dai Zhengqi (oral examination committee member), Lu Zhenyuan (oral examination committee member)

Summary-

Acoustic Tomography (Acoustic Tomography) is a technique that acquires information about a target area in a non-invasive manner. In the monitoring system, the temperature information of the target area is often an indispensable part of the monitoring. The use of acoustic tomography can reconstruct a two-dimensional temperature distribution map of the target area without affecting the target area. The key challenge is how to improve the matching accuracy of the relative spatial distribution and the absolute temperature value.

Using the ultrasonic temperature measurement method, the temperature of the medium in the target area can be calculated by the relationship between the transmission speed of the sound wave and the temperature of the medium. The ultrasonic array is used to generate dozens of ultrasonic transmission paths in the target area, and the temperature of the gas at each point on the two-dimensional plane is calculated based on the measured sound wave velocity of each path.

This study proposes an algorithm for reconstructing the temperature distribution based on multiple ultrasonic propagation paths, which is implemented in MATLAB. The algorithm primarily employs a spatial moving average technique and uses a hexagonal grid for the temperature distribution reconstruction. The study explores the impact of different grid shapes and sizes through several simulation tests. Three grid configurations (hexagonal, square, and polar) were analyzed for the accuracy of temperature distribution reconstruction. According to the results, the hexagonal grid most accurately tracks the peak temperature and movement of hotspots. In the ideal model, the average temperature error across all sound velocity paths is less than 1%, and when the hotspot moves outward from the center, the grid can detect the hotspot movement with a relative region diameter of 3%, with the maximum temperature error being less than 10%.

成大圖書館連結(請點我)