Operation status monitoring of reciprocating compressors based on the fusion of spatio-temporal multiple information

Abstract

Due to the complex structure and harsh working environment, it's difficult for the reciprocating compressors of offshore oil platforms to achieve accurate operational diagnosis. A multi-segment attention-based long short-term memory network (MA-LSTMs) is proposed to solve this problem. In order to obtain spatial information, internal attention mechanisms and a comprehensive attention mechanism are designed, three health curves are output to monitor the health status of the internal components of the compressor. A min-weights dropout strategy is added to prevent the model from overfitting. The temporal attention mechanism is used to fuse effective information from different periods of time, and finally the overall health curve is predicted through a decoding operation. Experiments with real industrial data show that the MA-LSTMs model has higher prediction accuracy and generalization ability than other models. On both validation sets, the RMSE, MAPE, and MAE of the MA-LSTMs model are lower than those of the DA-RNN model, which are reduced by 7.3%, 16.1%, 11.6%, and 19.8%, 53.2%, and 22.1%, respectively. In addition, the effectiveness of each module in the model is verified by ablation experiments, and the performance of the model under noisy conditions is verified.