One-Dimensional CNN-Based Intelligent Recognition of Vibrations in Pipeline Monitoring With DAS

Abstract

The vibration recognition along the fiber is still a challenging problem in pipeline monitoring with distributed optical-fiber acoustic sensor (DAS), because the burying environments in a wide range are complicated, and there are many different vibration sources interfering at different fiber locations, which are unpredictable and changing from time to time. Conventional machine learning methods with fixed hand-crated feature extraction are always time-consuming and laborious, and the recognition is relying heavily on expert knowledge, which has poor generalization ability. Thus, deep learning algorithms have been tried in this area. However, in this paper, it is found that one-dimensional (1-D) CNN can extract the distinguishable properties of the vibration signals of DAS with better performance and efficiency than the 2-D CNN through real field data experiments. And there are two main increment of the work: First, we try to use an efficient 1-D CNN to replace the 2-D CNN for feature extraction, which can improve the computation efficiency by directly feeding raw or the denoised data without any transformation or other manual work, and using simpler network structure; second, we optimize the classification further by replacing the softmax layer by the support vector machine (SVM) classifier, which is selected optimally from several typical classifiers, such as SVM, random forest, and extreme gradient boosting. Finally, the proposed method (1-D CNN+SVM) can achieve an average recognition accuracy of over 98% for five main classes of typical DAS signals in the oil pipeline monitoring application, which is superior to the conventional machine learning methods with fixed hand-crated feature. At the same time, both accuracy and efficiency of the method are better than those of the 2-D CNN.