Wellhead Compressor Failure Prediction Using Attention-based Bidirectional LSTMs with Data Reduction Techniques

In the offshore oil and gas industry, petroleum in each well of a remote wellhead platform (WHP) is extracted naturally from the ground to the sales delivery point. However, when the oil pressure drops or the well is nearly depleted, the flow rate up to the WHP declines. Installing a Wellhead Compressor (WC) on the WHP is the solution [9]. The WC acts locally on the selected wells and reduces back pressure, thereby substantially enhancing the efficiency of oil and gas recovery [21]. The WC sensors transmit data back to the historian time series database, and intelligent alarm systems are utilized as a critical tool to minimize unscheduled downtime which adversely affects production reliability, as well as monitoring time and cost burden of operating engineers. In this paper, an Attention-Based Bidirectional Long Short-Term Memory (ABD-LSTM) model is presented for WC failure prediction. We also propose feature extraction and data reduction techniques as complementary methods to improve the effectiveness of the training process in a large-scale dataset. We evaluate our model performance based on real WC sensor data. Compared to other Machine Learning (ML) algorithms, our proposed methodology is more powerful and accurate. Our proposed ABD-LSTM achieved an optimal F1 score of 85.28%.