Machine Learning Validation of Time Series Signals to Reduce Mistakes in Digital Algorithms for Maintenance, Optimization, and Automation

Abstract

The objective of the paper is to show how dynamic machine learning modeling can help drillers and operators validate signals from their sensors. Data and signal quality are a big problem in the industry when it comes to digitization. The method will show the importance of having a validation pipeline, and how it can help other algorithms make better decisions. Our approach uses statistical principles, machine learning and advanced analytics. The method is ISO 8000 compliant and can provide a framework in data management and data quality for companies to use. Depending on the application the accuracy of our method will vary. Results are anywhere in the 88% - 99% range of accuracy. The process has been validated by a major drilling contractor in signals ranging from blow out prevention, dynamic positioning systems, and tripping. The process can save upwards of 50% of time spent cleaning, mapping, and validating sensor signals. The end product allows the user to understand problems in the data collection system from the sensor all the way to the enterprise historian. It will also reduce false positives and false negative that are present in maintenance, optimization, and automation.