基于示功圖對抽油機井下工況進行自動診斷是數(shù)字油田不可或缺的環(huán)節(jié)。通常通過人工提取示功圖的特征向量，然后輸入機器學(xué)習(xí)算法分類器識別井下工況。然而，特征的選擇需要借助經(jīng)驗和先驗知識，并且直接影響后續(xù)分類器的最終性能。而人工提取特征易受知識干擾，且在特征提取的過程中存在信息丟失，這決定了識別結(jié)果的上限。為此，受深度神經(jīng)網(wǎng)絡(luò)自動特征提取的啟發(fā)，提出基于卷積神經(jīng)網(wǎng)絡(luò)的示功圖的離線訓(xùn)練與在線診斷的方法。首先將挑選后的信號數(shù)據(jù)轉(zhuǎn)換為圖像數(shù)據(jù)，然后將圖像二值化降低計算復(fù)雜度，最后基于改進的LeNet-5網(wǎng)絡(luò)探究最適合模型的網(wǎng)絡(luò)結(jié)構(gòu)。最終通過實驗與目前先進的算法進行對比，驗證了本方法的有效性和可行性。

Research on diagnostic method for working conditions of pumping unit wells based on deep learning

Abstract： It is an indispensable link to automatically diagnose downhole working conditions of pumping unit wells on dynamometer cards in digital oilfields. The feature vector of the dynamometer card is usually extracted manually and then identified downhole working conditions by the machine learning algorithm classifier as input. However, the selection of features requires experience and prior knowledge, then it directly affects the final performance of subsequent classifier. Features extracted manually are easily disturbed by knowledge, lead to the loss of key information, which determines the upper limit of the recognition result. Therefore, inspired by the automatic feature extraction of deep neural networks, this paper proposes the method of offline training and online diagnosis of dynamometer cards based on convolutional neural networks. Firstly the selected signal data is converted to image data. Then the image is binarized to reduce the complexity of computation. Finally based on the improved LeNet-5 network, we explore the network structure that is most suitable for the model. Compared with the current advanced algorithm, the validity and feasibility of this method are verified.

Key words : dynamometer card；downhole working conditions；automatic diagnosis；convolutional neural networks