透過相關文獻的探討，目前牙齒疾病的盛行率居高不下，主要原因是目前市面上的牙刷產品沒辦法提供使用者判斷刷牙清潔程度的參考指標，而影響刷牙清潔程度的主要因素是刷牙時間，因此，我們可以透過刷牙區域的分類辨識搭配時間計算提供使用者判斷清潔程度的參考標準。本論文著重於刷牙區域的辨識，使用慣性感測元件設計一套智慧牙刷系統，並提出一種基於人工智慧演算法的個人化刷牙區域自動辨識演算法。本論文主要分為硬體、韌體部分以及演算法部分，在硬體、韌體方面，我們使用小尺寸的MPU9250慣性感測元件來設計智慧牙刷系統，收取刷牙時的慣性訊號，在Arduino Nano中透過Madgwick algorithm使用加速度、角速度以及磁力計算牙刷的姿態角度，透過藍芽傳輸到電腦端進行後續的演算法分析；在演算法方面，我們使用Matlab進行開發，其中包含訊號處理、特徵選取、人工智慧分類器建構、交叉驗證，在訊號處理部分，本論文使用移動平均濾波器去除刷牙動作造成的雜訊，並且移除受到濾波延遲影響的資料，保留含有姿態訊息的資料；在特徵選取中，我們選擇三軸加速度、三軸磁力以及姿態角度(Roll, Pitch, and Yaw)作為輸入到分類器的分類特徵；而在人工智慧分類器的部分，本論文總共嘗試4種分類演算法：線性判別分析(Linear Discriminant Analysis, LDA)、二次判別分析(Quadratic Discriminant Analysis, QDA)、集成學習(Ensemble learning)以及長短期記憶(Long Short-Term Memory, LSTM)，前三種為機器學習模型，而LSTM為深度學習模型；交叉驗證的部分，本論文採用k-fold交叉驗證來評估本演算法的穩定性及表現。在實驗的部分，我們總共進行三種實驗：短期測試、長期測試以及深度學習測試，每項實驗都是基於個人化學習的方式進行(使用同一人的資料訓練分類器，並用同一人的另一段資料進行測試)。短期測試使用的分類器包含：LDA、QDA以及Ensemble learning，總共有28位受試者，每人分別收取兩次完整刷牙訊號，一段訊號作為訓練資料，另一段為測試資料，並且交換兩段資料得到兩個實驗結果，最後取平均準確率為最終結果；長期測試中使用的分類器為LDA，我們招募3位受試者，每人連續收取10天的刷牙訊號，用第1天的資料訓練，剩下9天資料做測試，取9個結果的平均準確率，並且同時進行10-fold交叉驗證；最後的深度學習測試，我們使用LSTM來進行分類器建構，受試者有7人，每人分別收取一段大約30分鐘的刷牙資料，利用這比較大段的資料進行5-fold交叉驗證得到最終的分類準確率。結果顯示：短期測試中，LDA的平均辨識準確率最佳，達到85.9%，並且有許多受試者的準確率達到95%以上，Ensemble learning為83.9%，QDA為80.9%；在長期測試中，三位受試者的平均準確率都在90%以上，其中許多天的測試結果都有達到100%，10-fold交叉驗證的結果也都在93%以上；而在深度學習測試的部分，整體平均準確率高達99.3%，每位受試者的結果都比短期測試中的結果進步許多，其中更是有使用者的準確率提高了16.9%之多。與相關的文獻進行比較，Y. J. Lee等人所提出的智慧牙刷演算法準確率為97.1%，但是其所使用的演算法限制較多，需要加入許多人為判斷，例如實驗中的刷牙方式只能是單一方向的刷牙，無法由使用者自行使用習慣的刷牙方式[24]，而本論文提出的演算法則不需要這些限制，因為我們使用深度學習演算法以及個人化學習的設計，可以讓使用者用自己習慣的刷牙方式，並且在最終的刷牙區域辨識達到較高的準確率99.3%。透過機器學習演算法與深度學習演算法的結果比較，我們認為LSTM適合本論文的原因在於慣性訊號的時間連續性，因為刷牙時的慣性訊號是時序訊號，而LSTM會將前一個時間點的運算結果加入下一個時間點的輸入進行運算，這可以加強同一個刷牙區域內慣性訊號的關聯性，達到較好的分類效果。綜上所述，本論文成功開發了一套智慧牙刷系統以及個人化自動刷牙區域分類演算法，透過人工智慧演算法以及個人化學習的設計，減少使用狀況限制，更方便使用，並且達到比相關文獻更高的分類準確率。

According to the literature review, the prevalence of dental diseases is still high, mainly because the current toothbrush products on the market cannot provide users a reference index for judging the degree of effectiveness of the brushing. Therefore, we can provide users a reference standard for judging the degree of cleanliness through the toothbrushing region classification and the calculation of brushing duration. This study focuses on the classification of the brushing regions. We used an inertial measurement unit to design a smart toothbrush system and proposed a personalized toothbrushing region classification algorithm based on artificial intelligence algorithms. This study contained hardware, firmware, and algorithm. In the hardware and firmware part, we used the small-sized MPU9250 IMU to design a smart toothbrush system and collect the inertia signal during toothbrushing. We applied Madgwick algorithm with acceleration, angular velocity, and magnetic force to calculate the orientation of the toothbrush in the Arduino Nano. Then transmitted the data to the computer through Bluetooth for analysis. In the algorithm part, we used Matlab for development, which includes signal processing, feature selection, artificial intelligence classifier construction, and cross-validation. In the signal processing part, we used a moving average filter to remove noise caused by toothbrushing motion, and removed the data affected by filtering delay, and retained data containing orientation information. In the feature selection part, we chose tri-axial acceleration, tri-axial magnetic force, and orientation (Roll, Pitch, and Yaw) as the classification features of the classifiers. In the part of the classifier construction, four different classifiers were utilized: Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Ensemble learning, and Long Short-Term Memory (LSTM), the first three were machine learning models, and LSTM was a deep learning model; in the cross-validation part, we used k-fold cross-validation to evaluate the stability and performance of the proposed algorithm. For the experiments, we conducted three kinds of experiments: Subjects with Single Trial (SST), Subjects with Multiple Trials (SMT), and Subjects with Multiple Trials/Mass Data (SMTMD). Every experiment was based on a personalized classification setting (trained the classifier with the same person's dataset and used another dataset from the same person for testing). The classifier used in the SST included: LDA, QDA, and Ensemble learning. There were 28 subjects in total. Each person brushed teeth twice for data collection, one dataset was used as training data, the other was testing data, and two datasets were exchanged for the second test. Two experimental results were obtained, and we took the average accuracy as the final result. The classifier used in the SMT was LDA. We recruited three subjects, we collected their brushing data on different ten days, and trained the classifier with the dataset from the first day. The remaining nine datasets were testing data, the average accuracy of 9 results is taken, and we also performed 10-fold cross-validation at the same time; for the SMTMD, we used LSTM classifier, there were seven subjects, each of them brushed teeth for about 30 minutes for data collection, and we used these large datasets to perform 5-fold cross-validation to get the final classification accuracy. The results showed that: in the SST, the average accuracy of LDA was the best, reaching 85.9%, and many results of subjects were more than 95%. The accuracy of ensemble learning was 83.9%, and QDA was 80.9%; in the SMT, the average accuracies of the three subjects were all above 90%, and several test results of different days achieved 100%. The results of the 10-fold cross-validation were also all above 93%; in the part of SMTMD, the average accuracy achieved 99.3%, and the results of each subject were much better than those in the SST. Among them, the accuracy of 1 subject increased by 16.9%. The accuracy of the smart toothbrush algorithm proposed by Y. J. Lee et al. was 97.1%, but their algorithm was more restrictive than ours, and many artificial judgments needed to be added. For example, the subjects could only brush teeth with the rolling stroke method in the experiments, and they could not use their own brushing methods [24]. In contrast, the proposed algorithm in this study did not need these restrictions because we used deep learning algorithms and personalized classification setting, which allowed users to brush teeth with their own habits. Also, the results of our toothbrushing region classification algorithm with accuracy equaled to 99.3% was better than that in Lee’s study. By comparing the results of machine learning algorithms and the deep learning algorithm, we thought that the LSTM classifier was suitable for this study because the inertial signal of the toothbrush when brushing was a sequence of time-series data, and LSTM classifier would take the previous hidden layer output as the input. It can enhance the association between the inertial data in the same region and achieve a better classification result. In summary, this study successfully developed a smart toothbrush system and a personalized toothbrushing region classification algorithm. Through artificial intelligence algorithms and personalized classification setting, we reduced the restrictions on use conditions and made the system more convenient for users with high classification accuracy.

[1] 國民健康署,“吸菸、嚼檳榔會影響牙周病治療的效果嗎,”國民健康署, Apr. 9, 2018. [Online]. Available: https://reurl.cc/VanmjA. [Accessed Nov. 24, 2019].
[2] T. Attin and E. Hornecker, “Tooth brushing and oral health: how frequently and when should tooth brushing be performed?,” Oral health & preventive dentistry, vol. 3, no. 3, 2005.
[3] 康健雜誌,“小心，牙周病已經悄悄找上你,”康健雜誌, Mar. 12, 2010. [Online]. Available: https://reurl.cc/drOqyV. [Accessed Nov. 24, 2019].
[4] 風傳媒,“健保醫療支出10大疾病排行榜出爐,” 風傳媒, Sep. 2, 2019. [Online]. Available: https://reurl.cc/A1XbM3. [Accessed Nov. 24, 2019].
[5] D. Slot et al., “The efficacy of manual toothbrushes following a brushing exercise: a systematic review,” International Journal of Dental Hygiene, vol. 10, no. 3, pp. 187-197, 2012.
[6] C. Ganss, N. Schlueter, S. Preiss, and J. Klimek, “Tooth brushing habits in uninstructed adults—frequency, technique, duration and force,” Clinical Oral Investigations, vol. 13, no. 2, p. 203, 2008.
[7] G. A. van der Weijden, M. F. Timmerman, E. Reijerse, C. M. Snoek, and U. van der Velden, “Toothbrushing force in relation to plaque removal,” Journal of Clinical Periodontology, vol. 23, no. 8, pp. 724-729, 1996.
[8] S. Turesky, N. D. Gilmore, and I. Glickman, "Reduced Plaque Formation by the Chloromethyl Analogue of Victamine C," Journal of Periodontology, vol. 41, no. 1, pp. 41-43, 1970.
[9] H. Löe, "Oral hygiene in the prevention of caries and periodontal disease," International Dental Journal, vol. 50, no. 3, pp. 129-139, 2000.
[10] A. Choo, D. M. Delac, and L. B. Messer, "Oral Hygiene Measures and Promotion: Review and Considerations," Australian Dental Journal, vol. 46, no. 3, pp. 166-173, 2001.
[11] G. A. Van der Weijden, M. F. Timmerman, M. M. Danser, and U. Van der Velden, “Relationship between the plaque removal efficacy of a manual toothbrush and brushing force,” Journal of Clinical Periodontology, vol. 25, no. 5, pp. 413-416, 1998.
[12] G. A. Van Der Weijden, M. F. Timmerman, M. M. Danser, M. Piscaer, Y. IJzerman, and U. Van Der Velden, “Approximal brush head used on a powered toothbrush,” Journal of Clinical Periodontology, vol. 32, no. 3, pp. 317-322, 2005.
[13] A. Gallagher et al., “The Effect of Brushing Time and Dentifrice on Dental Plaque Removal in vivo,” American Dental Hygienists Association, vol. 83, no. 3, pp. 111-116, 2009.
[14] T. Winterfeld et al., “Toothbrushing and flossing behaviour in young adults—a video observation,” Clinical Oral Investigations, vol. 19, no. 4, pp. 851-858, May 01 2015.
[15] F. A. V. d. Weijden and D. E. Slot, “Efficacy of homecare regimens for mechanical plaque removal in managing gingivitis a meta review,” Journal of Clinical Periodontology, vol. 42, no. S16, pp. S77-S91, 2015.
[16] E. Bernmark and C. Wiktorin, “A triaxial accelerometer for measuring arm movements,” Applied Ergonomics, vol. 33, no. 6, pp. 541-547, 2002.
[17] H. J. Luinge and P. H. Veltink, “Measuring orientation of human body segments using miniature gyroscopes and accelerometers,” Medical and Biological Engineering and Computing, vol. 43, no. 2, pp. 273-282, 2005.
[18] H. J. Luinge, P. H. Veltink, and C. T. M. Baten, “Ambulatory measurement of arm orientation,” Journal of Biomechanics, vol. 40, no. 1, pp. 78-85, 2007.
[19] S. Madgwick, “An efficient orientation filter for inertial and inertial/magnetic sensor arrays,” Report x-io and University of Bristol (UK), vol. 25, pp. 113-118, 2010.
[20] J. Lee, K. Lee, K. Kim, D. Kim, and K. Kim, “Development of Smart Toothbrush Monitoring System for Ubiquitous Healthcare,” in 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, 2006, pp. 6422-6425.
[21] K. Lee et al., “Tooth brushing Pattern Classification using Three-Axis Accelerometer and Magnetic Sensor for Smart Toothbrush,” in 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2007, pp. 4211-4214.
[22] Y. C. Chang et al., “Playful toothbrush: ubicomp technology for teaching tooth brushing to kindergarten children,” presented at the Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, Florence, Italy, 2008.
[23] Y. Lee et al., “Quantitative assessment of toothbrushing education efficacy using smart toothbrush,” in 2011 4th International Conference on Biomedical Engineering and Informatics (BMEI), 2011, vol. 2, pp. 1160-1164.
[24] Y. J. Lee et al., “Toothbrushing Region Detection Using Three-Axis Accelerometer and Magnetic Sensor,” IEEE Transactions on Biomedical Engineering, vol. 59, no. 3, pp. 872-881, 2012.
[25] Oral-B, “Oral-B GENIUS 9000,” Oral-B.com. [Online]. Available: https://reurl.cc/247agX. [Accessed Nov. 24, 2019].
[26] Freddie, “FlexCare Platinum Connected,” philips.com. [Online]. Available: https://reurl.cc/mdzpnA. [Accessed Nov. 24, 2019].
[27] Prophix, “An innovative way to brush,” getprophix.com. [Online]. Available: https://reurl.cc/YlGKXD. [Accessed Nov. 24, 2019].
[28] Grush, “Grush Smart Toothbrush,” grushgamer.com. [Online]. Available: https://reurl.cc/EKW5Rk. [Accessed Nov. 24, 2019].
[29] Oclean, “Oclean One,” oclean.com. [Online]. Available: https://reurl.cc/yy0558. [Accessed Dec. 18, 2019].
[30] CMUH, “Health Education,” cmuh.cmu.edu. [Online]. Available: https://reurl.cc/M7x7K3. [Accessed Dec. 26, 2019].
[31] O. Rudovic, J. Lee, M. Dai, B. Schuller, and R. W. Picard, "Personalized machine learning for robot perception of affect and engagement in autism therapy," Science Robotics, vol. 3, p. 19, 2018.
[32] X. Miao, "Personalized machine learning models," ed: Google Patents, 2015.
[33] J. Escudero et al., "Machine learning-based method for personalized and cost-effective detection of Alzheimer's disease," IEEE transactions on biomedical engineering, vol. 60, no. 1, pp. 164-168, 2012.
[34] S. A. Ludwig and K. D. Burnham, "Comparison of Euler Estimate using Extended Kalman Filter, Madgwick and Mahony on Quadcopter Flight Data," in 2018 International Conference on Unmanned Aircraft Systems (ICUAS), 2018.
[35] F. Destelle et al., "Low-cost accurate skeleton tracking based on fusion of kinect and wearable inertial sensors," in 2014 22nd European Signal Processing Conference (EUSIPCO), 2014, pp. 371-375.
[36] S. O. Madgwick, A. J. Harrison, and R. Vaidyanathan, “Estimation of IMU and MARG orientation using a gradient descent algorithm,” in 2011 IEEE international conference on rehabilitation robotics, 2011, pp. 1-7.
[37] CSDN, “Mathematics of Madgwick algorithm,” CSDN.net, May 28, 2019. [Online]. Available: https://reurl.cc/720nyy. [Accessed Oct. 30, 2019].
[38] CSDN, “Madgwick algorithm,” CSDN.net, Feb. 17, 2016. [Online]. Available: https://reurl.cc/mdM7vl. [Accessed Oct. 30, 2019].
[39] Mathuranathan, “Moving Average Filter,” GaussianWaves.com, Nov. 23, 2010. [Online]. Available: https://reurl.cc/vnao1e. [Accessed Oct. 30, 2019].
[40] InvenSense, “MPU-9250 Nine-Axis MEMS Motion Tracking™ Device,” MPU-9250 datasheet, Jun. 2016.
[41] MathWorks, “Statistics and Machine Learning Toolbox,” mathworks.com. [Online]. Available: https://reurl.cc/pD3VD8. [Accessed Oct. 30, 2019].
[42] MathWorks, “Deep Learning Toolbox,” mathworks.com. [Online]. Available: https://reurl.cc/31XZD8. [Accessed Oct. 30, 2019].
[43] S. Mika et al., "Fisher discriminant analysis with kernels," in Neural networks for signal processing IX: Proceedings of the 1999 IEEE signal processing society workshop (cat. no. 98th8468), 1999: Ieee, pp. 41-48.
[44] MathWorks, “Discriminant Analysis Classification,” mathworks.com. [Online]. Available: https://reurl.cc/Rdq3eD. [Accessed Oct. 30, 2019].
[45] Scikit-learn, “Linear and Quadratic Discriminant Analysis,” Scikit-learn.org. [Online]. Available: https://reurl.cc/pD3lob. [Accessed Oct. 31, 2019].
[46] Towards Data Science, “Simple guide for ensemble learning methods,” towardsdatascience.com. [Online]. Available: https://reurl.cc/XXG1zM. [Accessed Oct. 31, 2019].
[47] Wikipedia, “Ensemble learning,” en.wikipedia.org. [Online]. Available: https://reurl.cc/0zvA5M. [Accessed Oct. 31, 2019].
[48] MathWorks, “fitcensemble,” mathworks.com. [Online]. Available: https://reurl.cc/M7Oa1K. [Accessed Oct. 31, 2019].
[49] Wikipedia, “Long short-term memory,” en.wikipedia.org. [Online]. Available: https://reurl.cc/e5zelK. [Accessed Oct. 31, 2019].
[50] Wikipedia, “Cross-validation,” en.wikipedia.org. [Online]. Available: https://reurl.cc/rl9mGx. [Accessed Oct. 31, 2019].
[51] Towards Data Science, “Deep Learning vs Classical Machine Learning,” towardsdatascience.com, Apr. 4, 2018. [ Online]. Available: https://reurl.cc/Nabzgk. [Accessed Oct. 31, 2019].