Enhancing the value of low-cost optical fiber gas sensors via machine learning

Evanescent-field optical fiber sensors offer a wide range of benefits for gas and chemical sensing applications, including low cost per sensor node, stability under conditions that would be unfriendly for electronic sensor packages (e.g., high temperature, chemically harsh conditions, electromagnetic interference), and a small “footprint” for space-constrained applications. Optical fiber sensors also offer multiple unique capabilities, such as multi-wavelength interrogation via wavelength multiplexing and compatibility with spatially distributed interrogation techniques for multi-point detection. This approach relies upon the development of materials that demonstrate (1) high sensitivity, (2) selectivity, (3) fast response, and (4) long-term stability under relevant operating conditions. In this work, machine learning tools will be demonstrated to improve the performance in these areas, using simulated sensor data based on a simple, idealized sensor system. The application of recurrent neural networks in the form of a long short-term memory (LSTM) model will be applied to simulated time-series sensor data to demonstrate improvement in selectivity and effective response time. Multi-parameter sensing applications relevant to energy sector applications will be discussed.