Machine Learning Models for Analyzing FBG Pressure Sensor Data in Monitoring Leak in Water Pipeline
Aging water and sewage pipelines require effective monitoring as leakage and burst incidents pose environmental, economic, and public safety risks. Many existing studies rely on laboratory or simulated data from electrical sensors installed in harsh sewer environments, which require regular cleaning...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , , , |
| التنسيق: | article |
| منشور في: |
2026
|
| الوصول للمادة أونلاين: | http://hdl.handle.net/11675/14529 https: |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | Aging water and sewage pipelines require effective monitoring as leakage and burst incidents pose environmental, economic, and public safety risks. Many existing studies rely on laboratory or simulated data from electrical sensors installed in harsh sewer environments, which require regular cleaning and calibration. The novelty of this work lies in presenting a systematic, real-world evaluation of machine learning models for pipeline leak detection using fiber Bragg grating-based pressure sensors deployed in an operational pipeline test facility for the first time. The models evaluated include supervised models, such as Random Forest and Extreme Gradient Boosting (XGBoost), and unsupervised models, such as Long Short-Term Memory (LSTM) and LSTM Autoencoder. These models are compared under identical experimental conditions and were evaluated under both serial (single sensor) and parallel (multi-sensor) configurations, using sliding windows of varying lengths. Weighted precision and F1-scores were used to address strong class imbalance across leak sizes. Results show that XGBoost achieved the best overall performance, reaching 77% precision and 52% F1-score at a 45-point window length. The results obtained are better than those reported in previous work, most of which rely on simulations or laboratory datasets and critically do not evaluate models under real operational pipeline conditions, as is done here. The findings demonstrate the robustness of supervised XGBoost when applied to small, noisy real-world datasets and they highlight the feasibility of integrating ML-based FBG sensors into cloud-enabled pipeline monitoring systems. |
|---|