Design and Application of Mixed Natural Gas Monitoring System Using Artificial Neural Networks

Jinlei Wang, Bing Li, Bingjie Lei, Peiyuan Ma, Sai Lian, Ning Wang, Xin Li, Shaochong LeiSensors 2021 · OpenReview

TL;DR

This paper presents a novel, robust, and maintainable online monitoring system for mixed natural gas, leveraging an infrared sensor array and an artificial neural network (ANN) for accurate component analysis. The system addresses the limitations of traditional methods by offering real-time, on-site detection with improved stability and reduced cost, making it suitable for harsh industrial environments.

What Problem Does This Paper Address?

The exploitation and utilization of natural gas require accurate and real-time analysis of its composition. Traditional methods like gas chromatography are accurate but slow, expensive, and require laboratory settings, making them unsuitable for continuous, on-site monitoring. While electrical sensors offer a more practical solution for online monitoring, they suffer from cross-sensitivity issues, where one sensor responds to multiple gases, leading to inaccurate readings. Furthermore, existing systems often lack robustness, maintainability, and proper hardware-software integration for industrial applications, especially in challenging environments like gas fields. This paper aims to bridge this gap by developing a stable, accurate, and maintainable online monitoring system for mixed natural gas that overcomes these limitations.

Methods

The proposed system integrates several key innovations:

  • Novel Hardware Architecture: A core circuit with an FPGA is encapsulated in an anti-static box for enhanced electrostatic protection, crucial for dry environments. An auxiliary board acts as a buffer between the core and the sensor array, improving system stability and maintainability.
  • Data Preprocessing with EWMA: To mitigate noise and improve the stability of sensor readings, the Exponentially Weighted Moving Average (EWMA) method is employed. This technique gives more weight to recent data, effectively filtering out abnormal samples and reducing the impact of non-ideal factors.
  • Tailored Artificial Neural Network (ANN): A Multilayer Perceptron (MLPNN) is designed for calibration to eliminate the cross-sensitivity of the infrared sensors. The network's structure, including the number of hidden layers and neurons, was optimized through extensive experimentation to balance accuracy and hardware implementation scale.
  • Specific Instruction Set: A custom instruction set is designed to facilitate software-hardware coordination, simplifying tasks like parameter updates and system calibration, which are essential for maintaining accuracy over time due to sensor aging.

Datasets

The paper describes the collection of 4128 groups of data in the laboratory. Of these, 4000 groups were used for training the ANN, and the remaining 128 groups were used for testing. The authors also mention retraining the neural network using data collected from the gas field during deployment, indicating a real-world dataset was utilized for fine-tuning.

Key Insights and Ideas

  • System-Level Optimization is Crucial: The paper emphasizes that improving individual components (sensors, algorithms) is not enough; a holistic system design that considers hardware architecture, software integration, and data processing is essential for industrial applications.
  • ANN as a Powerful Tool for Sensor Fusion: ANNs are demonstrated to be highly effective in compensating for the inherent cross-sensitivity of sensor arrays, enabling accurate composition analysis from multiple, imperfect sensors.
  • EWMA for Robust Real-time Data: EWMA proves to be a computationally efficient and effective method for preprocessing real-time sensor data, enhancing stability and filtering out noise without requiring excessive storage.
  • Hardware-Software Co-design for Maintainability: The development of a specific instruction set and the modular hardware design significantly improve the system's maintainability, allowing for easier calibration and parameter updates, which are critical for long-term industrial deployment.
  • FPGA for Real-time Processing: Utilizing an FPGA for ANN computation enables high-speed, parallel processing, crucial for real-time monitoring applications.

Key Results

  • The developed monitoring system achieved high recognition accuracy for alkane gases, with maximum errors dropping to 0.58% for methane, 1.15% for ethane, and 1.18% for propane after ANN calibration in laboratory settings.
  • The system has been running stably for several months in a gas field, demonstrating its reliability and suitability for harsh industrial environments.
  • The ANN calibration effectively reduced the cross-sensitivity of the infrared sensors, with average errors of 0.24% for methane, 0.33% for ethane, and 0.38% for propane.
  • The system offers a significantly faster response time (seconds) compared to traditional methods like gas chromatography (hours to days).
  • The manufacturing and operational costs are lower compared to existing technologies.

Limitations and Future Work

  • Reference Sample Dependency: The paper acknowledges that training the neural network requires accurate reference samples, which are typically obtained through gas chromatography. The process of obtaining large-scale samples for training can be time-consuming and complex.
  • Sensor Aging and Drift: While the ANN calibration helps mitigate the effects of sensor aging and drift, it is noted that parameters need to be updated regularly. The authors suggest that the gradual adjustment of neural network parameters is an ongoing process.
  • Future Work: The authors suggest several directions for future research, including:
  • Improving the accuracy of the test.
  • Extending the service life of the sensor array.
  • Further enhancing the automation level of the system.
  • Integrating wireless communication modules for greater convenience.

Summary updated

Abstract

Natural gas component analysis is one of the significant technologies in the exploitation and utilization of natural gas. A stable and accurate online natural gas monitoring system is necessary for the gas extracting industry. We have developed an online monitoring system of natural gas with a novel hardware architecture. It improves the dependability and maintainability of the system. A specific instruction set is designed to facilitate the coordination of software and hardware. To reduce the sample noise, the exponentially weighted moving average (EWMA) method is used to preprocess the real-time raw data of the sensor array. A tailored neural network is designed for calibration. And the relationship between the performance and the structure of the gas neural network is demonstrated to find the optimal solution for accuracy and hardware scale. The design not only focuses on the optimization of individual components but also focuses on system-level improvement. The system has been running stably for several months in the gas fields. It meets the requirements of stability, ease of use, maintainability, and online monitoring in industrial applications.
BibTeX
@inproceedings{wang2021design,
  author = {Jinlei Wang and Bing Li and Bingjie Lei and Peiyuan Ma and Sai Lian and Ning Wang and Xin Li and Shaochong Lei},
  title = {{Design and Application of Mixed Natural Gas Monitoring System Using Artificial Neural Networks}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=8fx1Q4NiKl},
}
Figure