應用機械學習於高壓瓶閥音洩故障診斷研究

Fault Diagnosis of High-Pressure Bottle Valves Based on Acoustic Emission Using Machine Learning

本文應用音洩技術於高壓瓶閥體的洩漏偵測，發展機械學習方法進行模型和參數的訓練，以達到閥體各種故障的自動診斷目的。在不同測試壓力和故障模式下，將麥克風裝設在閥門外感測閥體洩漏訊號，再利用資料擷取卡擷取訊號到電腦號進行特徵處理。時域的特徵值包括均方根值（Root mean square）、平均值（Mean）和變異量（Variance）。利用離散傅立葉轉換（Discrete Fourier Transform，DFT）將訊號轉換到頻域，為了減少頻譜特徵值的個數，將DFT頻譜中的振幅由大到小排列，取前最大50個振幅和所相對應的頻率共100個，做為機械學習訓練的特徵值。以人工智慧機器學習方式，建立加權k-近鄰演算法（Weighted k-nearest-neighbors , wkNN）、人工類神經網絡（Artificial Neural Network, ANN）和二次支持向量機（Quadratic support vector machine, qSVM）三種機械學習模型和相對應的訓練演算法，藉以辨識出高壓閥體在不同通氣壓力下，閥體墊圈半破裂、墊圈全破裂、破裂盤和閥芯髒污等四種故障，並且比較這三種音洩學習模型的優劣。最後，透過各種實驗驗證本文所提方法的有效性，閥體故障辨識率整體可達98%以上。

Machine learning models were used to classify acoustic emissions from 700-bar high-pressure bottle valve bodies to identify leakage and diagnose various valve faults under different test pressures and failure modes. A microphone was installed outside a valve to sense acoustic leakage signals that were then captured by a data acquisition card for transmission to a computer. For classification, time-domain features were selected, including the signal’s root mean square, mean, and variation values. Signals were converted to the frequency domain by using the discrete Fourier transform DFT））. To reduce the number of spectral characteristic values, the 50 frequencies with the greatest amplitudes in the DFT spectrum were selected. Weighted k-nearest-neighbors, artificial neural network, and quadratic support vector machine models were trained to classify four of the following types of faults in high-pressure valves under different pressure scenarios: half-ruptured valve body gasket, fully ruptured gasket, ruptured disk, and dirty valve core. The method was experimentally validated and achieved an identification rate for valve body faults of over 98%.