本研究主要利用神經網路機器學習之方法運用於船模阻力試驗的初始訊號數據判讀。船模阻力試驗中，阻力計輸出其電壓訊號，會被紀錄且繪製成圖。本研究比較以人工判讀方式、時間序列訊號分析、神經網路訓練時間序列訊號分析(NN)以及卷積神經網路訓練影像辨識(CNN)之結果，期望能將阻力計紀錄結果之判讀自動化，使阻力試驗數據更不受實驗人員之操作影響。
傳統判讀方式是試驗人員根據經驗，判定臺車的穩定等速度段，再挑出雜訊或是其他因素導致的誤差。本研究中將其結果與臺車速度紀錄比較，以了解是否準確地取出穩定等速度段。時間序列訊號分析是設計一套更完善之運算方式，運用統計學讓程式運算出各項數據之關係，將數據分成加速段、穩定段以及減速段，且同樣將其結果與臺車速度紀錄比較。神經網路機器學習是本研究的重點項目。將阻力計之訊號與臺車時間繪製成波形影像，利用神經網路學習判讀時間序列訊號，以及用卷積神經網路學習判讀訊號影像中各段數據之種類。其中會運用不同原始訊號處理方式以及不同的神經網路架構，比較其學習效率及準確度。
研究結果發現：目前神經網路與卷積神經網路方法之準確率約在70%左右，且其原始訊號數據處理時，需將數據分割成適當大小；而四種方法比對時，在各種情況下被以人工判讀方式以及時間序列訊號分析可信度較高，神經網路與卷積神經網路存在使加速段誤判為減速段之問題導致可信度降低，未來須先改良此部分後才可適用於船模阻力試驗之訊號分析。

This study focuses on the application of neural network machine learning in the data interpretation of the initial signal in the ship-model resistance test. In the ship-model resistance test, the resistance gauge outputs the voltage signal, and these were recorded and graphed. The study compares the results of practicing time series signal analysis via manual interpretation, time series signal analysis, and the neural network (NN), with practicing image recognition via the convolutional neural network (CNN). The study attempts to automatize the interpretation of the results recorded by the resistance gauge in order to minimize the human influence of the experimenters on the resistance test data.

Traditionally, in the process of interpretation, the experimenter would deduce the constant velocity section of the vehicle based on past experience, and select a datum that deviated from the average by more than one standard deviation, which would be regarded as the noise or an error caused by other factors. This study compares such results with the record of the velocity of the vehicle in order to examine whether the constant velocity section has been acquired precisely. The time series signal analysis is an even more comprehensive calculation method and enables the program to calculate the connections between the data by means of statistics, dividing the data into an acceleration section, a constant section, and a deceleration section. The results of this analysis were also compared with the record of the velocity of the vehicle. A key element of this study was neural network machine learning. The signal from the resistance gauge and the time period of the vehicle was converted into a waveform image. The time series signal was interpreted by means of the neural network, while each type of data from each section in the signal image was interpreted by means of the convolutional neural network. Different treatments of the original signal and different neural network structures were also applied in order to compare the learning efficiency and accuracy that each produces.
The results of the study indicate that the neural network and the convolutional neural network can achieve an average accuracy of 70 percent, and that the data has to be divided into proper-sized units during the treatment of the original signal data. Moreover, the comparison results of the four methods suggest that manual interpretation and time series signal analysis have a higher degree of credibility in all circumstances. On the other hand, the neural network and the convolutional neural network can result in the acceleration section being misinterpreted as the deceleration section, thus indicating a lower degree of credibility. However, this is not applicable in signal analysis in the ship-model resistance test at the present stage.

李宏毅 (2017) ML Lecture 0-1: Introduction of Machine Learning，國立台灣大學電機工程學系，https://www.youtube.com/watch?v=CXgbekl66jc， (2017/09/29)
李宏毅 (2017) Machine Learning (Hung-yi Lee, NTU) ，國立台灣大學電機工程學系，https://www.youtube.com/watch?v=CXgbekl66jc&list=PLJV _el3uVTsPy9oCRY30oBPNLCo89yu49， (2017/09/29)
陳維揚 (2014) 拖航水槽主要船模試驗項目之技術改善與比較分析，國立 成功大學系統與船舶機電工程學系碩士論文。
福島邦彥 (1980) Neocognitron, A. self-organizing neural network model for a mechanism of visual pattern recognition. Competition and cooperation in neural nets. Springer, Berlin, Heidelberg, 1982: 267-285.
Alex, K. (2012) Imagenet classification with deep convolutional neural Networks. Hinton G E. Imagenet classification with deep convolutional neural networks[C]//Advances in neural information processing systems. 2012: 1097-1105.
Aäronvanden, O. & Sander, D. (2016) SoundNet: Learning Sound Representations from Unlabeled Video. Advances in neural information processing systems.
Jake, B. (2006) Notes on Convolutional Neural Networks. Center for Biological and Computational Learning Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology Cambridge, MA 02139.
ITTC (2002) ITTC-Recommended Procedures 7.5-02 -02-01 Testing and Extrapolation Methods Resistance Resistance Test. 7.
ITTC (2008) ITTC-Recommended Procedures 7.5-02 -02-01 Testing and Extrapolation Methods Resistance Resistance Test. 7.
Li, D. & Dong, Y. (2014) Deep Learning: Methods and Applications 6.2. Foundations and Trends in Signal Processing. Now Publishers Inc. 252.
Lecun (1989) Generalization and Netwoek Design Strategies. Department of Computer Science University of Toronto. Technical Report CRG-TR-89-4
Oord, A. & Dieleman, S. (2016) WaveNet:A Generative Model for Raw Audio. arXiv:1609.03499v2 [cs.SD]