Intelligent Deepwater Energy Development: flow assurance monitoring and smart decision-making system
🎉 Our paper has been accepted and published by The Innovation Energy 🎉
The detection of hydrate blockages in deep-sea pipelines has long been a challenge for the industry. With the rapid development of machine learning and deep learning technologies, some efforts have been made to apply these methods to the prediction of hydrate blockages.
However, most existing work relies on supervised machine learning techniques for predicting hydrate blockages, which results in models being highly dependent on labeled datasets. This dependency limits the application of supervised learning models. Therefore, we innovatively transform the hydrate blockage detection problem into an unsupervised one.
- 🔍 Unsupervised Framework: First transformed the hydrate blockage detection problem into an unsupervised learning task to adapt to varying data distributions under different working conditions and overcome data labeling difficulties.
- 🧠 TLV Model: Proposed a Transformer-LSTM model within the VAE framework, which combined the long-range sequence information extraction ability of Transformer with the sequence generation capability of LSTM.
- 📊 Dynamic Weight Adjustment for Threshold Setting: Proposed a dynamic weight adjustment mechanism that increases the model's sensitivity to subsequent data. In practical industrial applications, this improvement facilitates earlier warnings of potential blockages.
- 📁 Loop Data for Hydrate Blockage Detection: Open-sourced a large-scale loop dataset that can simulate the blockage process of deep-sea pipelines under real-world conditions.
Our model achieves state-of-the-art (SOTA) performance on this proposed dataset, outperforming several unsupervised methods.
We compare our model with 5 baselines, including KMeans, Anomaly Transformer, etc. Generally, our model achieves SOTA.
Especially, we abandon the adjustment operation used in Anomoly-Transformer and DCdetector for fair model evaluation.
-
Environment. We support the python environment created by conda. You can prepare the environment by the following instrution. Make sure that your cuda version is suitable with your tensorflow version, otherwise it would cause "GPU not available" problems.
conda create --name tlv python==3.10 # use python3.10 conda activate tlv pip install -r requirements.txt pip install openpyxl -
Download data. You can obtain the proposed loop dataset by cli 76D2 ck here or you can use data from
data/database.xlsxdirectly. -
Train and evaluate. You can use the
run.pyto train and evaluate model. In Linux, you can directly execute thetrain.shand thetest.shto do the same thing. You can custom your own config inconfig.json.python run.py --mode [train | test] --config_path "config/config.json" # For Linux bash script/train.sh bash script/test.sh
-
Model Checkpoint. You can obtain the existing model checkpoint by click here, which is identical to the model used in our experiments.
-
Inference Demo. You can quickly start an inference demo using the
infer_demo.py. In Linux, you can directly execute theinfer.shto do the same thing. Relying on this file, you can further proceed with production-level deployment.python infer_demo.py --config_path "config/config.json" # For Linux bash script/infer.sh
- ✅ Open source our database
- ✅ Organize and open source our codebase（include train, test, deploy demo）
- ✅ Add the README profile
- ✅ Add the paper's citation
@article{wang2025intelligent,
title={Intelligent Deepwater Energy Development: flow assurance monitoring and smart decision-making system},
author={Wang, Jiguang and Yang, Bowen and Jin, Kaiming and Zhang, Lunxiang and Wang, Linhe and Jiang, Bo and Yao, Haiyuan and Song, Yongchen},
journal={The Innovation Energy},
year={2025},
publisher={The Innovation Energy}
}