Recognition of pre-emergency situations during oil wells exploitation based on telemetry analysis

The paper proposes an AI model for predictive oil well diagnostics. An open 3W dataset with rare adverse events on offshore oil wells was used in present study. The paper proposes to extract features from the telemetry time series using autoencoder based on convolutional neural network. The time series were preliminarily converted into derivatives. Then derivatives were converted into spectrograms and finally were converted into average amplitude spectra. These average amplitude spectra were subsequently subjected to neural network analysis. A transfer training scheme was used to set up the autoencoder. The use of the transfer training scheme made it possible to reduce the amount of telemetry data used to set up an image classification system. After training, the neural network can extract informative features from the telemetry data. A modified Bayes classifier was used in present study with classification accuracy of 0.97.