A hybrid model of LSTM neural networks with a thermodynamic model for condition-based maintenance of compressor fouling

Abstract

Compressor fouling is one of the critical gas path faults of gas turbines, and the fouling process is significantly influenced by the quality of the inlet air coming from the air intake system with filters. The maintenance strategies for compressor fouling mainly consist of online/offline washing and replacement of filters, where optimizing the washing cycles and the replacement of filters is essential for the economy and safety of gas turbine operation as of the trade-off between the performance improvement and the corresponding costs. By considering the filtration effects of the air intake system, the gas path analysis of the gas turbine is carried out to tackle the coupled fouling process, and a hybrid framework is presented to predict the washing cycle (remaining useful life prediction for washing) and detect filter leakage (diagnosis for filter) via integrating the thermodynamic model and long short-term memory (LSTM) neural networks. The proposed scheme is applied in a field dataset and the results show that: (i) a deterioration index based on the thermodynamic model can be used to evaluate the compressor fouling degree, which is independent of ambient conditions and control factors. (ii) A prediction model based on the LSTM-Hankel method demonstrates good performance in long-time washing cycle prediction. (iii) Air filter leakage will significantly increase the degradation rate of compressor efficiency, which can be identified by the diagnosis model to predict the new washing cycle.

Prescriptive maintenance for complex products with digital twin considering production planning and resource constraints

“…Yet, its training process often suffers from gradient disappearance or explosion. For this purpose, the long short-term memory (LSTM) was constructed to deal with the long-term dependence of the time series [41]. As its evolutionary version, the gate recurrent unit significantly reduces training time by combining input gates and forgetters into update gates [42,43].…”

A Novel Data-Driven Approach for Predicting the Performance Degradation of a Gas Turbine

“…Real field data from a gas turbine power plant were used for training and testing, and it was found that this method had a smaller relative mean square error compared to other prediction methods. Y. Jin et al [21] developed a hybrid framework that integrates thermodynamic models and Long Short-Term Memory (LSTM) neural networks to predict washing cycles (remaining service life prediction) and detect filter leakage (filter diagnosis). The prediction model based on the LSTM-Hankel method exhibited good performance in long-term washing cycle prediction.…”