On-Edge Implemented Machine-Learning Based Synthetic Flame Detector for Gas Turbine Operation

Abstract We have successfully designed and implemented a cutting-edge data-driven flame detection system for gas turbines using a 55-parameter feed-forward neural network. This simple neural network model receives as input seven distinct physical quantities measured at various locations within the turbine, such as fluid temperature, pressure, flow rate, and shaft speed. Basically, a digital twin of the physical flame detector has been developed, which allows to replace the sensor and obtain a more cost-efficient, low-maintenance and accurate solution. The neural network has been integrated into the machine control system (MarkVIe) through edge device implementation, enabling real-time synthetic flame detection. The model has demonstrated exceptional performance, achieving 100% recall and 100% precision on an independent test dataset. Future enhancements will focus on integrating the model onto a MarkVIs device, an enhanced version of MarkVIe specifically designed for managing security-related machine controls automatically, allowing for the closure of the control loop and further optimizing the flame detection process.