Deep Learning Based Anomaly Detection on Natural Gas Pipeline Operational Data

In oil and gas industry, production monitoring is a crucial yet challenging problem. Small changes or issues in the production process should be detected and taken care of swiftly to avoid unwanted consequences. Some machine learning tasks can be applied to potentially helping monitoring process. One of these tasks is anomaly detection. However, existing studies have not conducted anomaly detection properly in the field of oil and gas industry due to tendency of using plain techniques or unsuitable paradigms. To handle this, deep learning model of autoencoder type is implemented to detect anomalies on gas pipeline operational data. Specifically, the data used is 2 years long of hourly time series data which contains 17 features and 17520 datapoints. The autoencoder model will be trained to reconstruct the data with a minimum error. Once trained, the reconstructed data can be compared to the original data to detect any potential anomaly. This paper considers 6 different cases of model architecture while comparing the results using mean squared error to obtain the optimal model setup in the reconstruction process. To evaluate the model, we compute Euclidean distance between the reconstructed and original data as the anomaly score. By determining a threshold value of the score, we can identify the anomalies in the data. The resulting identifications are analyzed qualitatively and show good prediction of the presence of abnormalities in data.