| 摘要: | 現今社會在買房交屋前會先進行驗屋，檢測室內地磚是否敲擊有空心的聲音並且是否符合標準。傳統驗屋方式為使用打診棒敲擊地磚五個點(左上、左下、右上、右下、中點)，如果超過三點（含三點）空心就要更換重貼，現今驗屋方式更加嚴格謹慎，將地磚分成九宮格進行敲擊，一樣超過三點（含三點）空心就要換掉重貼。地磚檢測過度依賴有經驗的技術人員判斷，並且會產生檢測標準過度主觀且無法紀錄的問題。因此本論文以機器學習領域中的深度學習做為解決問題的核心，將地磚空心辨識功能嵌入在Arduino Nano 33 BLE Sense微控制器上，以便於攜帶，使民眾或技術人員能以更加簡單且更有效率的方式進行地磚檢測，達到民眾可自行檢測且可以以數值的方式記錄，讓檢測結果更加有可信度，也能及早發現地磚空心問題並得以立即進行補強或更換。本論文先是設計並利用3D列印技術製作出空心磚模型，使用打診棒敲擊地磚並以Arduino Nano 33 BLE Sense內建的麥克風採集音訊資料，將所有音訊資料分成數個類別並透過Edge Impulse進行音訊前置處理和採用頻譜圖(Spectrogram)、梅爾濾波器組能量(Mel-filter bank energy,MFE)和梅爾頻率倒普係數(Mel-Frequency Cepstral Coefficients,MFCC)三種方法提取特徵，使用卷積神經網路(Convolutional Neural Networks ,CNN)訓練模型，之後藉由驗證和測試，比較Spectrogram、MFE和MFCC三種方法之衡量指標，最後發現Spectrogram模型且辨識門檻為0.75準確率最高，驗證集準確率為97.7%，測試集準確率為92.48%，實際敲擊地磚準確率最高可達81.25%，因此將此模型部署至Arduino Nano 33 BLE Sense上並安裝在自行設計打診棒上以便直接利用。一般民眾或檢測人員可自行在家或驗屋時利用智慧打診棒檢測地磚，透過簡易的敲擊便能即時辨識出地磚是否空心和是否需要更換或補強，以降低過度依賴主觀判斷和無法記錄的問題。
;Before buying a house, an inspection is carried out to check whether the indoor floor tiles have a hollow sound when knocked and whether they meet the standards. The traditional home inspection method uses a medical stick to hit five points on the floor tiles (upper left, lower left, upper right, lower right, and center). Floor tile inspection relies too much on the judgment of experienced technicians, and the inspection standard is too subjective and cannot be recorded. Therefore, this paper takes deep learning in the field of machine learning as the core of solving the problem and embeds the hollow recognition function of floor tiles on the Arduino Nano 33 BLE Sense microcontroller so that it is easy to carry so that people or technicians can use it more simply and conveniently. The efficient way to test the floor tiles enables the public to test by themselves and record them numerically so that the test results are more credible. The hollow problem of floor tiles can be found early and can be reinforced or replaced immediately. This thesis first designs and uses 3D printing technology to make a hollow brick model, taps the floor brick with a medical stick and collects audio data with the built-in microphone of Arduino Nano 33 BLE Sense, divides all audio data into several categories, and conducts audio through Edge Impulse Preprocessing and extracting features using three methods: Spectrogram, Mel-filter bank energy (MFE) and MelFrequency Cepstral Coefficients (MFCC), using convolution Neural network (Convolutional Neural Networks, CNN) training model, then through verification and testing, compare the measurement indicators of the three methods of Spectrogram, MFE, and MFCC, and finally found that the Spectrogram model and the identification six thresholds of 0.75 have the highest accuracy, and the accuracy of the validation set is The accuracy of the test set is 97.7%, the accuracy of the test set is 92.48%, and the accuracy of the actual hitting the floor tile is up to 81.25%. Therefore, this model is deployed on the Arduino Nano 33 BLE Sense and installed on the self-designed diagnosis stick for direct use. The general public or inspectors can use the intelligent diagnostic stick to detect floor tiles at home or during house inspections. Through a simple tap, they can instantly identify whether the floor tiles are hollow and whether they need to be replaced or reinforced to reduce excessive reliance on subjective judgments and problems that cannot be recorded. |
