Improving performance of one-class classifiers applied to anomaly detection in oil wells

Abstract

The prompt detection and diagnosis of anomalies in oil wells are fundamental to reduce production losses, maintenance costs and to avoid environmental damage. In this paper, a proposal for detecting such anomalies using Long Short-Term Memory (LSTM) autoencoder and one-class Support Vector Machine (SVM) is presented. A public dataset with instances of eight types of common anomalies characterized by eight process variables is used for training and performance evaluation. A methodology is proposed for training and performance evaluation. Faulty data are used as the target class for one-class classifiers. A time-shift of labels is proposed to improve the discrimination of normal and faulty data during the training phase. Criteria for selecting the time-shifting of the labels are proposed as well. A performance metric is proposed by computing the average of Specificity, F1, and instance identification performance. The detection time compared to the transient time of the fault was also computed. The analyses are performed with emphasis on two classes of anomalies, with slow and fast dynamics. For the fault with fast dynamic, performance was increased by 14% for LSTM with no improvement for SVM. For the slow fault, the performance was increased by 41% for LSTM and 21% for SVM. A noticeable reduction of detection time was observed for the slow fault, from 100% to 59.34% for LSTM and 87.7% to 64.51% for SVM (100% corresponds to the transient time of the fault). A comparison with random forest and decision tree classifiers presented in the literature and applied to the same classes of anomalies shows the superiority of the proposed approach and methods. The special attention given to the metrics used for evaluating classifier performance and the time to detect the anomalies in the instances show that apparently good results measured by F1 and accuracy sometimes hide some weakness.