Optimizing Pipeline Leak Detection: Leveraging Attention-based 1DCNN-BiLSTM for Enhanced Accuracy and Minimal False Alarms
Date
2024-09-20
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Pipelines are an essential infrastructure for the transportation of fluids and gases in many industries. Leaks in pipelines present significant environmental and economic concerns, making accurate and timely leak detection crucial. Recent advances in deep learning, particularly sequential models, have shown promising capabilities in anomaly detection for time series data. However, the challenge remains to detect leaks accurately while minimizing false alarms. This paper presents a novel approach combining the CB-AttentionNet model, which integrates a 1D convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and multi-head attention mechanisms to capture local and long time series dependencies. Additionally, we introduce a probabilistic search framework using Monte Carlo methods to optimize window sizes dynamically, addressing the limitations of fixed window sizes in handling variable-length sequential data. Experimental results demonstrate that our method performs better in terms of accuracy and reducing false positives across various simulations with industrial pipeline data. Optimized window sizes, particularly between 45 and 60 seconds, offer an effective balance between reducing misclassified leaks and maintaining high training accuracy. Furthermore, our analysis of resource usage and evaluation time shows that the model’s performance is efficient and manageable within typical operational constraints.
Description
Keywords
Leak Detection, Pipeline, Deep Learning, CNN, LSTM, Attention Mechanism
Citation
Khazali, S. (2024). Optimizing pipeline leak detection: leveraging attention-based 1DCNN-BiLSTM for enhanced accuracy and minimal false alarms (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.