最佳卷積神經網路架構在水流表面流速之探討

Abstract

流量是設計及管理水利工程中的重要考量因子，準確的流量能在最低的施工成本中達到其興建目的，製造安全及經濟雙贏的結果，而準確的流速才能計算出準確的流量。在流速量測中，若以人工進行接觸式的量測，需要花費較多時間，在大流量時亦不易量測，因此非接觸式的量測成為近年來重要發展的項目。其中粒子影像測速法(PIV)是一常被討論的非接觸式量測法，透過影像對的匹配計算單位時間的位移量以推估出流速。在影像匹配方法大多採用互相關性演算法(DCC)，透過計算影相對間之強度分佈來進行匹配影像，然連續影像之強度分佈易受到攝影設備對焦、環境光照、陰影等環境及人為影響。故本研究嘗試採用卷積神經網路(CNN)提取二維特徵的優勢，透過模式訓練學習以辨識匹配影像對。 本研究將使用人造粒子影像以確保有真實值能夠評估DCC法及CNN法的差異，在流場選擇上使用穩態均勻流、渦流、射流，並在三種流場分別加入不同的雜訊及光照，再利用DCC法及CNN法進行流速計算，其中考慮了子圖像的尺寸、輸入資料的尺寸、不同層數及深度的卷積神經網路配合不同的激活函數。根據本研究所使用的方法，子圖像為 時兩方法皆有最佳表現；若輸入資料的尺寸選用原始子圖像的尺寸，則需要四層卷積層的神經網路才能良好的計算出流速，而若先將子圖像縮放為 時，則僅需要兩層卷積層的神經網路；改變激活函數的影響在本研究方法中並不明顯；而擁有四層卷積層的神經網路在面對不同尺寸的子圖像時，會表現的較僅有兩層卷積層的神經網路穩定。並且在擁有四層卷積層的神經網路中，加深卷積層深度對其計算結果沒有明顯幫助。在面對雜訊時，DCC法的誤差急劇上升，而CNN法相對穩定，如在穩態均勻流中添加35%高斯雜訊時，DCC法較CNN法誤差增加3.6倍。依本研究所使用的案例，CNN法可以取代DCC法在PIV中的計算。

The discharge of flow is an important consideration in the design and management of water conservancy. An accurate flow can achieve its construction purpose at the lowest cost, and create a safe and economic win-win result additionally. Since the contact flow measurement techniques is performed manually, it not only takes much time but also difficult to perform at a large flow rate. Therefore, non-contact measurement techniques has become a hotspot in recent years. Particle image velocimetry (PIV) is widely used in measuring the surface flow velocity and estimating the discharge in the field. The method of PIV to estimate the flow velocity is to calculate the displacement by matching image pairs. Direct cross-correlation algorithm (DCC) matches the images by calculating the intensity distribution between the shadows. However, the intensity distribution of the continuous image is susceptible to environmental and human influences such as camera equipment focusing, ambient lighting, and shadows. Therefore, this study attempts to use the advantages of Convolutional Neural Network (CNN), extract two-dimensional features, to identify matching image pairs. In this study, the artificial particle images are used to ensure that there are real velocity values to evaluate the difference between the DCC method and the CNN method. Uniform and steady flow, vortex flow, and jet flow are used in the flow field selection. The results show that the CNN method is more stable than the DCC method when noise and light are adding in the images. In this study, we choose three types of convolutional neural Networks with combinations of different depths and convolutional layer, and changing the activation function. The aim is to find a suitable convolutional neural Network framework. According to the method used in this study, the effect of the activation function is not obvious. A neural Network with four convolutional layers will perform better than a neural Network with two convolutional layers when faced with different size sub-images. With a neural Network with four convolutional layers, deepening the depth of the convolutional layer does not help significantly. In the face of noise, the error of the DCC method rises sharply, and the CNN method is relatively stable. For example, when 35% Gaussian noise is added to the steady uniform flow, the DCC method has an error of 3.6 times that of the CNN method. According to the case used in this research, the CNN method can replace the calculation of DCC method in PIV.