Abstract: Aiming at the inefficiency of gas turbine combustion chamber fault warning due to the scarcity of sample data, a method based on transfer learning and convolutional neural networkbidirectional gated recurrent network (CNN-BiGRU) is proposed. Firstly, the CNN-BiGRU network is pre-trained in the source domain using K-fold cross validation and the optimal model is filtered out; then the target domain fault warning model is obtained using transfer learning; then the target domain data is imported into the model and a sliding window is used to reduce the false alarms caused by the abnormal data points; finally, the warning thresholds computed using the triple standard deviation (3-sigma) method are used to determine whether a fault occurs. The experimental results show that compared with the non-transfer learning model, the proposed method reduces the root mean square error by 87.5%, the mean absolute error by 89.05%, and improves the r-squared by 6.39% when there is insufficient sample data; and compared with the system alarm moment, it is able to detect the signs of faults earlier, and warn the gas turbine combustion chamber of faults 82 min in advance.

Key words: gas turbine, combustion chamber, fault warning, transfer learning, convolutional neural network, bidirectional gated recurrent unit network