A spatio-temporal fault diagnosis method based on STF-DBN for reciprocating compressor

Reciprocating compressor is the core equipment of petrochemical industry and its stable running is very important for productions in the offshore drilling platform. The reason why it is difficult to extract features from vibration signals to reflect the operating state of the compressor is that its internal structure is complex and there are many excitation sources. To solve this problem, a new fault diagnosis method based on spatio-temporal features fusion based on deep belief network (STF-DBN) was proposed, which comprehensively processes multi-source signal features from dimensions of time and space. The temporal features extraction strategy is designed to reflect the data trend by reconstructing the time series according to different period characteristics of fault-related parameters. And the spatial features are extracted to reflect the non-amplitude characteristic of data by breaking down the raw data trend and considering the importance of reconstructed series to various faults. STF-DBN can overcome the deficiency of traditional unsupervised network DBN that cannot extract periodic features, no longer rely on the number of fault data samples, and construct a more comprehensive health curve representing the operation status of reciprocating compressors for fault diagnosis and early warning. The classic Tennessee Eastman (TE) data set in the control field is used for the diagnosis effect test, and the STF-DBN is applied to the operation status detection of the reciprocating compressor for offshore natural gas extraction of China National Offshore Oil Corporation. The experimental results confirm the effectiveness of the proposed method in fault defection and early warning.