Modified spatio-temporal neural networks for failure risk prognosis and status forecasting of permanent downhole pressure gauge

Abstract

Permanent Downhole Pressure Gauges (PDPGs) are sensors installed in the wellbores to measure the downhole pressure. Due to the abrupt changes in pressure, the PDPG can fail, and this will result in difficulties in production optimization. To manage the operational risk levels (negligible, low, mild, high and severe), a signal processing technique that utilizes anomaly detection, Fisher's Discriminant Function (FDF) and Mean Deviant Function (MDF) is developed. The operational risk levels were established with the Bayesian Gaussian Mixed Model (BGMM) and a Spatiotemporal Graphical Model (STGM) helped to compute the risks transition probabilities at different time intervals. After balancing the resulting input-output data with Synthetic Minority Oversampling Technique (SMOTE) and using Multi-layer Perceptron Artificial Neural Network (MLP-ANN) classifier, an online PDPG monitoring strategy was developed. The model was tested with Coal Seam Gas (CSG) PDPG data from thirty-eight wells and the results showed an average accuracy of ~93%. This result indicates the potency of the model for managing the risk profile of PDPGs and other telemetric equipment used for critical missions in the oil and gas, space, medical and defence industries.