刷牙是預防各種口腔疾病的主要方法，但刷牙全面且時間足夠才能夠真正降低牙齒疾病發生率。現有的智慧牙刷相關研究，雖然能夠以刷牙時的姿態角辨識刷牙區域，但使用者的身高及牙刷擺放位置等因素無法確定，因此若使用固定的模型辨識，將導致姿態辨識精確度及穩定性不足，還有無法監測刷牙的正確性和完整度的缺點。本論文因此提出一個遞迴機率神經網路模型DRPNN，應用於智慧牙刷姿態辨識。DRPNN由系統中已存在的PNN模型抽取出適當的個人刷牙特徵建立，模型包含記憶神經單元，具有自適應能力，利用PSO演算法迭代調整參數至模型最佳狀態，實驗結果發現本論文所提出之DRPNN辨識模型，刷牙姿態辨識率可達到98.64%，透過增加遞迴記憶單元平均準確率可達到99.08%，平均辨識率比使用CNN模型辨識高16.2%，也比使用LSTM模型辨識高21.21%。模型大小遠小於CNN與LSTM神經網路模型，能夠於低成本嵌入式系統中進行即時刷牙姿態辨識，改善現有智慧牙刷成本過高、辨識精度低、和智慧化不足等缺失。;Tooth brushing is the main method to prevent various oral diseases, only if thorough and long enough tooth brushing can reduce the incidence of tooth disease. In the existing studies about smart toothbrush, the tooth brushing area can be recognized by the attitude angle during tooth brushing, but the user′s body height and toothbrush location factors are uncertain. Therefore, if a fixed model is used for recognition, the posture recognition accuracy and stability will be insufficient, and the tooth brushing correctness and integrity cannot be monitored. This paper proposes a Dynamic Recurrent Probability Neural Network (DRPNN) for smart toothbrush posture recognition. The DRPNN uses the existent Probability Neural Network model in system to extract appropriate personal tooth brushing feature establishment. The model has memory cell and adaptive capability. The parameters are tuned iteratively by using Particle Swarm Optimization algorithm to the optimum condition of model. The experimental results show that the tooth brushing posture recognition rate of the recognition model proposed by this study is 98.64%. The average accuracy rate is 99.08% after the recurrent unit is used. The average recognition rate is higher than the Convolutional Neural Networks (CNN) model by 16.2%, and higher than the long short-term memory (LSTM) model by 21.21%. The model size is much smaller than the CNN and LSTM neural network models. The real-time tooth brushing posture recognition can be implemented in low cost embedded system. The deficiencies in the existing smart toothbrush can be remedied, such as high cost, low recognition accuracy and insufficient intelligence.
