基於機器視覺之閥門內部洩漏檢測:應用於石油產業設備之開發

Abstract

在石化產業中，工業閥門必須要能在高溫、高壓或低溫等極端條件下密封流體。但在物理條件限制下，無法達成完全密封，仍會產生極為小洩漏，考量到這些極微小洩漏可能對環境帶來的危害，美國石油協會(API)及其餘相關協會各自定義關於洩漏檢測的標準，將工業閥門的洩漏依部位區分為內部與外部洩漏，並使用氣泡為單位來計算洩漏量。 首先本論文使用符合規範定義的洩漏收集瓶，並設計黑箱外殼罩住外部形成一個封閉的成像環境，以隔絕外部雜訊干擾，接著安裝嵌入式系統及相機，在內部收集洩漏影像，供機器視覺辨識演算法進行計數。接著本論文建立一個人機互動介面，方便測試人員確認測試環境及設備參數，同時也將規範整併至數位化資料庫內部，以減少測試人員反覆查閱多份規範表單所消耗的時間成本。 本論文共開發兩種氣泡辨識計數的演算法，首先是基於側邊投影的生成式計數法，利用影像的亮度特徵進行計數，實驗結果顯示使用此方法能有效的進行計數，正確率可達99% ; 其次是使用底部投影的完整追蹤計數法，透過結合YOlO-V5模型及SORT演算法，對氣泡動態進行完整追蹤，並計算符合洩漏動態的氣泡序列，實驗結果顯示使用完整追蹤計數法在小洩漏的準確率上較生成式計數法佳，但是由於計算速度不夠快而無法在邊緣運算裝置上實現real-time檢測。 最後，本論文結合正規化的的規範資料庫與人機互動介面，達成一站式的檢測，並實際收集洩漏影像以驗證其效果。透過計算可知使用此檢測設備可以降低閥門製造商的人力成本至原先的三成。

In the petrochemical industry, industrial valves must be able to seal fluids under extreme conditions such as high temperature, high pressure, or low temperature. Hence, most industrial valves use metal sealing components. However, due to some physical conditions, it is impossible to achieve perfect sealing, and there may still be tiny leaks. Considering the potential environmental hazards caused by these tiny leaks, the American Petroleum Institute (API) and other relevant associations have defined standards for leak detection. They classify valve leaks into internal and external leaks and use bubbles as units to calculate the leakage rate. Because current leakage detection depends on examiners, this thesis aims to develop automatic counting algorithms to relieve the labor work. We develop two bubble recognition and counting methods: generative counting and tracking counting methods. We built leak collection bottles that complied with the specified standards, and we designed a black box enclosure to create a closed imaging environment to isolate interferences from external noises. We then developed an embedded system with a camera to collect leak images for automatic counting. We also constructed a human-machine interaction interface to confirm the test environment and equipment parameters. The specifications are integrated into the digital database to satisfy the multiple standards by different associations. The generative counting method applied side projection, counting bubbles by the brightness features of images. Experimental results showed that this method could effectively count with an accuracy rate of up to 99%. The tracking counting method applied bottom projection, tracking the dynamic movements of bubbles and calculating bubble sequences. Experimental results showed that the tracking counting method achieved better accuracy than the generative counting method when leakage was little. However, due to its processing speed, the tracking counting method could not achieve real-time detection. Finally, we integrated the system as a portable leak detection robot for leakage detection in the petroleum and petrochemical industries. It uses machine vision to develop algorithms for leak counting and combines a normalized specification database and a human-machine interaction interface to achieve one-stop detection. Actual leak images are collected to verify the effectiveness of the device. We also estimated that this detection robot could reduce the labor cost for valve manufacturers by two-thirds.