This paper proposes a novel heat recovery steam generator (HRSG) early tube leak detection model which leverages a convolution neural network classifier by utilising transfer learning with ResNet50 architecture. The design goal of this model was to achieve high classification accuracy with a minimal amount of leakage data. The model is also intended to be user-friendly and require minimal hyperparameter tuning. The proposed neural network was trained on the drum-specific conductivity time series data of HRSGs encoded in the Gramian Angular Difference Field (GADF). The model yielded a validation accuracy of 96.64%, true-positive rate of 93.28% and precision of 100% in regard to the validation set. The study included experiments on the influence of different encoding algorithms, Markov Transition Field (MTF) and Recurrence Plot (RP), and architectures on the performance of the model. This paper further discusses the viability of adapting the design to other time series classification problems.
Keywords:
HRSG conditional monitoring; machine learning; deep learning; transfer learning; time series classification