臺灣博碩士論文加值系統

本論文以深度學習(Deep Learning, DL)與聲音辨識技術(Sound Recognition)為基礎，辨識金屬待測物如鐵罐及鋁罐及其瑕疵樣本。本論文研究過程主要分為三個部分，第一部分為以非接觸式聲波檢測系統對鐵罐及鋁罐進行聲音資料蒐集，能在不對目標物產生結構性破壞的前提下對其進行聲音蒐集，第二部分為將蒐集到的聲音進行聲資料預處理如聲音資料分析、聲音資料標籤、特徵提取、聲音資料拆分等，以完成訓練資料集準備。第三個部分為模型的建立、訓練與識別，將預處理過的聲音資料集輸入至以Tensorflow的高階函式庫建構的卷積神經網路(Convolutional Neural Network, CNN)中加以訓練以取得模型，透過調整激活函數及其它超參數來最佳化模型，並以訓練好的模型辨識金屬待測物及其瑕疵樣本。其中於特徵提取步驟後，嘗試將特徵以三種不同的線性與非線性降維演算法進行特徵分群且增加不同類別特徵向量間的距離，藉此保留特徵的最大變異性，並與一般的高維特徵資料集進行比較，實驗特徵維數的高低是否會對模型的訓練成果與運算時間造成影響。

Defection identification of metal objects such as iron cans and aluminum cans and their defective samples is using sound recognition technology based on deep machine learning in this paper. This thesis is mainly divided into three parts: firstly the sound data of iron cans and aluminum cans with Sound-Wave Non-Contact Detection System (SWNCDS) can be collected without causing structural damage to the target objects; secondly sound preprocessing includes sound data analysis, sound data labeling, sound feature extraction and sound data split, etc. to complete dataset preparation; thirdly models are established for model training and recognition. The preprocessed sound data is as input into the convolutional neural network constructed whit Tensorflow function library for training to obtain the model. Adjustment and tunning of the activation function and other hyperparameters is used to to optimize the model for identifying the iron cans and aluminum cans and their defective samples. After the feature extraction step, we try to add three different dimensionality reduction algorithms including linear dimensionality reduction and nonlinear dimensionality reduction. By reducing the dimension of the input data and retaining most of the characteristics of the original data, we can test the difference in accuracy, and compare with the general high dimensional feature dataset to check the level of feature dimension which affect the training results and operation time of the model.

摘要......i
Abstract......ii
誌謝......iii
目錄......iv
表目錄......vi
圖目錄......vii
第一章 緒論......1
1.1 研究動機與目的......1
1.2 文獻探討......3
1.3 論文架構......5
第二章 深度學習與聲音辨識技術......6
2.1 深度學習(Deep Learning)......6
2.1.1 人工神經網路......6
2.1.2 多層感知器(Multiple Layer Perceptron, MLP)......7
2.1.3 激活函數(Activation Function)......7
2.2 卷積神經網路(Convolutional Neural Network, CNN)......9
2.2.1 卷積層(Convolutional Layer)......10
2.2.2 池化層(Pooling Layer)......11
2.2.3 扁平層(Flatten Layer)......12
2.2.4 全連接層(Fully Connected Layer, FC)......13
2.3 前向傳播演算法(Forwardpropagation)/反向傳播演算法(Backpropagation, BP)......13
2.4 學習率&優化(Learning Rate And Optimization)......14
2.4.1 動量(Momentum)......15
2.4.2 梯度下降法(Gradient Descent)......16
2.4.3 Full&Mini Batch......16
2.4.4 RMSprop......17
2.4.5 Adam......17
2.5 聲音及聲學特徵......18
2.5.1 頻譜質心(Spectral Centroids, SCs)......19
2.5.2 過零率(Zero Crossing Rate, ZCR)......19
2.5.3 色度特徵(Chroma Feature)......19
2.5.4 梅爾倒頻譜係數(Mel Frequency Cepstral Coefficients, MFCC)......20
2.6 降維演算法......23
2.6.1 主成分分析(Principal Component Analysis, PCA)......23
2.6.2 線性判斷式分析(Linear Discriminant Analysis, LDA)......24
2.6.3 t-隨機鄰近嵌入法(t-Distributed Stochastic Neighbor Embedding, t-SNE)...... 25
第三章 研究架構及實驗方法......26
3.1 研究架構......26
3.2 資料蒐集......26
3.3 聲音資料預處理......28
3.3.1 聲音資料分析......28
3.3.2 數據標籤與映射......31
3.3.3 固定聲音長度......31
3.3.4 聲音資料拆分......32
3.4 特徵提取......32
3.5 模型建立......33
3.5.1 GPU......33
3.5.2 Callbacks EarlyStopping......34
第四章 實驗結果與討論......35
4.1 實驗流程與設定......35
4.2 模型資料維數比較......35
4.2.1 未降維資料......36
4.2.2 降維資料......36
4.3 鐵罐鋁罐辨識實驗......39
4.3.1 實驗一：調整優化器(Optimizer)......39
4.3.2 實驗二：調整激活函數(Activation Function)......41
4.3.3 實驗三：調整批次(Batch Size)......42
4.3.4 實驗四：調整執行週期(Epoch)......44
4.3.5 鐵罐鋁罐區分模型辨識比較與結果......48
4.4 鐵罐鋁罐缺陷識別實驗......49
4.4.1 實驗五：調整優化器(Optimizer)......49
4.4.2 實驗六：調整激活函數(Activation Function)......51
4.4.3 實驗七：調整批次(Batch Size)......52
4.4.4 實驗八：調整執行週期(Epoch)......54
4.4.5鐵罐鋁罐缺陷識別模型辨識比較與結果......58
第五章 結論與未來展望......59
5.1 結論......59
5.2 未來展望......60
參考文獻......61
Extended Abstract......64

[1] Turing, A.M. (1980). Computing Machinery and Intelligence. Creative Computing, 6, 44-53.
[2] Russell, S.J., & Norvig, P. (1995). Artificial Intelligence: A Modern Approach.
[3] Vishnubhatla, A. (2021). Classification of Fresh Vegetables Through Deep Learning and Neural Networks. Cognitive Informatics and Soft Computing.
[4] Munaka, T., Samie, F., Bauer, L., & Henkel, J. (2020). Improved Feature Extraction Method for Sound Recognition Applied to Automatic Sorting of Recycling Wastes. J. Inf. Process., 28, 658-665.
[5] Davis, S., & Mermelstein, P. (1980). Comparison of Parametric Representations for Monosyllabic Word Recognition in Continuously Spoken Se.
[6] Chaithra, & Doddamani, M.B. (2020). Automated Waste Segregator and Visualization for Separation of Wastes using Neural Networks.
[7] Zhang, M., Yan, L., Luo, G., Li, G., Liu, W., & Zhang, L. (2021). A Novel Insect Sound Recognition Algorithm Based on MFCC and CNN. 2021 6th International Conference on Communication, Image and Signal Processing (CCISP), 289-294.
[8] Santiago, R.M., Rabano, S.L., Billones, R.K., Calilung, E.J., Sybingco, E., & Dadios, E.P. (2017). Insect detection and monitoring in stored grains using MFCCs and artificial neural network. TENCON 2017 - 2017 IEEE Region 10 Conference, 2542-2547.
[9] Shin, H. C., Roth, H. R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mollura, D., & Summers, R. M. (2016). Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE transactions on medical imaging, 35(5), 1285–1298.
[10] Winursito,A., Hidayat, R., & Bejo, A. (2018). Improvement of MFCC feature extraction accuracy using PCA in Indonesian speech recognition. 2018 International Conference on Information and Communications Technology (ICOIACT), 379-383.
[11] Roweis, S.T., & Saul, L.K. (2000). Nonlinear dimensionality reduction by locally linear embedding. Science, 290 5500, 2323-6 .
[12] Rachmad, A., Anamisa, D.R., & Bintari, N.P. (2017). Voice Recognition Application by Using Fisher’s Linear Discriminant Analysis (FLDA) Feature Extraction. Advanced Science Letters, 23, 12344-12348.
[13] Potratz, G.L., Canchumuni, S.W., Castro, J.D., Potratz, J., & Pacheco, M.A. (2021). Automatic Lithofacies Classification with t-SNE and K-Nearest Neighbors Algorithm.
[14] Yang, J., Zhang, D., Frangi, A.F., & Yang, J. (2004). Two-dimensional PCA: a new approach to appearance-based face representation and recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26, 131-137.
[15] Schmidhuber, J. Deep Learning in Neural Networks: An Overview. Neural Networks. 2015, 61: 85–117.
[16] Schmidt-Hieber, J. (2020). Nonparametric regression using deep neural networks with ReLU activation function. Annals of Statistics, 48, 1875-1897.
[17] Wang, Y., Li, Y., Song, Y., & Rong, X. (2020). The Influence of the Activation Function in a Convolution Neural Network Model of Facial Expression Recognition. Applied Sciences.
[18] Cortes, C., & Vapnik, V.N. (2004). Support-Vector Networks. Machine Learning, 20, 273-297.
[19] Wang, Y., Li, Y., Song, Y., & Rong, X. (2020). The Influence of the Activation Function in a Convolution Neural Network Model of Facial Expression Recognition. Applied Sciences.
[20] Hochreiter, S., & Bengio, Y. (2001). Gradient Flow in Recurrent Nets: the Difficulty of Learning Long-Term Dependencies.
[21] Schuller, B., Steidl, S., Batliner, A., Vinciarelli, A., Scherer, K.R., Ringeval, F., Chetouani, M., Weninger, F., Eyben, F., Marchi, E., Mortillaro, M., Salamin, H., Polychroniou, A., Valente, F., & Kim, S. (2013). The INTERSPEECH 2013 computational paralinguistics challenge: social signals, conflict, emotion, autism. INTERSPEECH.
[22] LeCun, Y., & Cortes, C. (2005). The mnist database of handwritten digits.
[23] Kim, Y. (2014). Convolutional Neural Networks for Sentence Classification. EMNLP.
[24] Wu, L., & Perin, G. (2021). On the Importance of Pooling Layer Tuning for Profiling Side-channel Analysis. IACR Cryptol. ePrint Arch..
[25] Ruder, S. (2016). An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747.
[26] Ruder, S. (2016). An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747.
[27] Kingma, D.P., & Ba, J. (2015). Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6980.
[28] Gupta, S., Jaafar, J., Fatimah, W., & Bansal, A. (2013). FEATURE EXTRACTION USING MFCC. Signal & Image Processing : An International Journal, 4, 101-108.
[29] Ming, J.T., Noor, N.M., Rijal, O.M., Kassim, R.M., & Yunus, A. (2018). Lung Disease Classification Using Different Deep Learning Architectures and Principal Component Analysis. 2018 2nd International Conference on BioSignal Analysis, Processing and Systems (ICBAPS), 187-190.
[30] Pérezgonzález, Antonio, Vergara, M. , Sanchobru, J. L. , Van, D. M. L. J. P. , Hinton, G. E. , & Shanmugapriya, D. , et al. (2015). Visualizing Data using t-SNE.
[31] Rácz, A., Bajusz, D., & Héberger, K. (2021). Effect of Dataset Size and Train/Test Split Ratios in QSAR/QSPR Multiclass Classification. Molecules, 26.