機器學習於熱交換管渦電流缺陷訊號的處理及缺陷分類與預測

Abstract

本研究使用機器學習方法，對熱交換管渦電流非破壞檢測之一維訊號進行訊號處理及分析，渦電流檢測設備在單一頻率條件下，在缺陷處可產生阻抗的實部及虛部之兩組訊號變化。研究分為兩個主要部分：實驗和資料分析。實驗部分是收集資料，包括使用雙頻渦電流設備檢測銅鎳和鈦標準管的缺陷渦電流訊號，以及使用四頻設備檢測鈦標準管，以直接獲取四個頻率的缺陷渦電流訊號。本論文使用渦電流差異式探頭之訊號作為研究。 在資料分析部分，由於測量使用的管件每隔若干距離，有外部鐵磁性材料的支撐，這些支撐板會對渦電流訊號，造成非線性的訊號干擾，使檢測人員難以直接判讀訊號的缺陷類型。為解決此問題，本研究應用類神經網路對該訊號進行濾波處理，不僅能消除支撐環的訊號並保留完整的缺陷訊號，還能解決訊號漂移問題。 接著，由濾波後的訊號中，以物理模型所定義的各種特徵進行特徵提取，這些特徵主要從訊號之二維阻抗圖中提取，包括八字圖案的相位角、面積、半徑、寬度等。此外，我們也考慮到並非所有的特徵都是必要的，且某些特徵之間可能因存在高度的相依性，而降低辨別準確度，因此本研究通過單一特徵分類表現及特徵相關圖對各種特徵進行排序，以有效地精簡的特徵數目，達到準確且有效的缺陷評估。然後，將這些特徵輸入隨機森林分類器，對每個缺陷訊號進行學習分類，另外也通過神經網路進一步預測缺陷的深度。 此外，由於支撐環與缺陷重疊的相對位置不同，支撐環對缺陷訊號產生的非線性訊號影響會不同。我們利用少數取得的實際資料，以兩個特徵:缺陷種類和與支撐環相對位置，通過類神經網路學習，以模擬出各種缺陷與對應相對位置的渦電流訊號，該模擬器可做為協助訓練檢測人員判讀缺陷的資料庫。 最後，本研究考慮到未來有新的資料加入時，該模型需要重新學習建立新模型。為了減少重新學習以建立訓練模型之浪費，我們採用增量式學習方法，使其不需在重新完整訓練的情況下，僅採用少部分就資料並加入新資料進行訓練學習，有效學習新資料，更新舊模型。 本研究的主要目的是結合人工智慧與專家系統，使專家能依據原有的物理模型所定義的缺陷特徵，通過機器學習建立缺陷評估系統，以輔助其做更準確且有效地的缺陷判讀。

This study uses machine learning methods to process and analyze one-dimensional signals in eddy current non-destructive testing of heat exchanger tubes. Eddy current testing equipment generates two sets of signal changes, the real and imaginary parts of impedance, at defect locations under single-frequency conditions. The research is divided into two main parts: experiments and data analysis. The experimental part involves data collection, including the use of dual-frequency eddy current equipment to detect defect signals in copper-nickel and titanium standard tubes, as well as the use of four-frequency equipment to detect titanium standard tubes, thereby directly obtaining defect eddy current signals at four different frequencies. This study focuses on the signals obtained from eddy current differential probes. In the data analysis part, the tubes used in the measurements are supported by external ferromagnetic material plates at certain intervals. These support plates cause nonlinear signal interference with the eddy current signals, making it difficult for inspectors to directly identify defect types from the signals. To solve this problem, this study applies neural networks for filtering the signals. This approach not only eliminates the support ring signals and preserves the complete defect signals but also addresses the issue of signal drift. Next, features defined by physical models are extracted from the filtered signals. These features are mainly derived from the two-dimensional impedance diagram of the signals, including the phase angle, area, radius, and width of the figure-eight pattern. Additionally, we consider that not all features are necessary, and the high interdependence among some features might reduce the accuracy of identification. Therefore, this study ranks the features through single-feature classification performance and feature correlation diagrams to effectively streamline the number of features, achieving accurate and efficient defect assessment. These features are then input into a random forest classifier to classify each defect signal. Moreover, neural networks are used to further predict the depth of defects. Furthermore, due to the different relative positions of the support rings overlapping with the defects, the nonlinear signal impact of the support rings on the defect signals varies. Using a small amount of actual data, we employ two features—defect type and relative position to the support ring—and use neural networks to simulate the eddy current signals for various defects and their corresponding relative positions. This simulator can serve as a database to assist in training inspectors to interpret defects. Finally, considering the future addition of new data, the model needs to relearn to establish a new model efficiently. To reduce the waste of retraining, we adopt incremental learning methods, allowing the model to learn new data effectively and update the old model without a complete retraining. The main objective of this study is to combine artificial intelligence with expert systems, enabling experts to use defect features defined by physical models to establish a defect assessment system through machine learning, thereby assisting in more accurate and efficient defect interpretation.