Application of Novel SN-1DCNN-LSTM framework in small sample oil and gas pipeline leakage detection

Abstract

In this article, a novel ensemble framework of improved siamese network (SN) is proposed to address the small sample issue that deep learning approaches encounter, as well as to enhance the precision of pipeline leakage detection (PLD) under small sample conditions. Firstly, training samples are input in pairs to the feature extraction network, and a combination of one-dimensional convolution neural network (1DCNN) and long short-term memory (LSTM) network is introduced to extract features of the time-series data, thus enhancing the effectiveness and robustness of feature extraction. Then, an improved relational metric network is designed to measure the similarity of features, to further strengthen the discriminative nature of the whole framework. In addition, the framework has been augmented with a classification network that can be used directly for PLD. The proposed SN-1DCNN-LSTM framework not only increases the number of training samples from the side, but also fully exploits the similarity information and data features between samples, overcoming the problem of instability and overfitting of deep learning models under small sample conditions. Finally, the experimental results verify the validity and superiority of the method under small sample conditions.

Keywords

• Pipeline leakage detection
• Small sample
• Siamese network
• Long short-term memory
• One-dimensional convolution neural network
• Time-series