Abstract: Due to the data imbalance problem in the fault diagnosis of oilfield plunger injection pumps, a multilevel Inception-LSTM (long short term memory) network incorporating wavelet packet decomposition and efficient channel attention (ECA) mechanism is proposed in this paper. The wavelet packet decomposes the low-frequency and high-frequency components of the vibration signal. The Inception module extracts the data features from multiple scales. The LSTM module captures the temporal correlation of the data. The ECA mechanism further enhances the model's mining ability for cross-channel data correlation. The experiment is based on the data collected in an operation site of an oil field. The results show that the proposed model performed optimally. The fault diagnostic accuracy can reach 99.38%, which demonstrates the effectiveness and superiority of the proposed model.

Key words: plunger injection pump; imbalance data, convolutional neural networks, long short term memory（LSTM）, attention mechanisms