Anomaly Detection on Compressor Application: Detailed Evaluation of Statistical and Machine Learning Techniques
Suradech Kongkiatpaiboon, Sarita Laosuwan, Warinphak Suwanpong, Polake Kaivalkritiyakul, Chain Sopitviriyaporn, Songkiet Manoharn, Siriwan Payaksiri · 2024
Abstract The ability to identify, and manage unexpected events is essential for improving productivity and minimizing downtime. Recently, there have been significant advancements in statistical and machine learning techniques for anomaly detection. Despite this, aging and increased downtime issues have been observed in facilities equipment in oil and gas fields. There are numerous opportunities to apply anomaly detection techniques to enhance efficiency. This study assesses techniques and their implementation to reduce compressor downtime and enhance plant productivity. This study analyzes effective anomaly detection initiatives as well as their difficulties and solutions throughout data cleansing, feature engineering, model training, back-testing, and deployment processes. Best practices are identified through a series of interviews. The challenges are that there are overwhelming cases to make a predictive task, simple univariate measures seldom result in satisfactory solutions, and model accuracy greatly drops when the forecast horizon gets longer. This paper examines the necessary steps to address the challenges for smooth execution. Statistical and machine learning methods are evaluated to assess accuracy, taking into consideration of many remote compressor operating parameters. Thousands of shutdown events covering five-year operating records are investigated and categorized into three groups. These are planned controllable, unplanned controllable, and uncontrollable cases. Only the unplanned controllable group is focused on the next steps, as the other two groups present challenges in devising effective mitigation strategies. Data size, processing time, and signal smoothing techniques are assessed. The study shows that data resampling and signal smoothing are crucial and affect results. Although the time series data post challenges in deploying a good visualization tool, the right amount of sampling frequency and smoothing technique is required in this research. Next, several machine learning models are tested, both supervised and unsupervised, followed by back-testing. It is found that the supervised method demands a substantial amount of data with similar shutdown root causes; hence, the usage is limited to common failures. On the other hand, unsupervised methods provide many encouraging results. The anomaly score of both options is compared with the investigation data and found to be effective. Both statistical and rule-based techniques are also evaluated. Some strategies work in certain situations. In the end, a hybrid method that combines machine learning and statistical techniques is chosen for deployment, which leads to the timely mitigation of over seventy percent of unplanned controllable cases. This study investigates multiple options and presents an effective methodology for handling real-time data, recognizing anomalous events, and determining potential underlying factors in the operation of offshore remote compressors. The approach and conclusions of this study, including best practices and lessons learned during the evaluation and implementation, have significant benefits for both the petroleum industry and academic research, particularly within the domain of operation and maintenance.