Recurrent Auto Encoders for Automatic Sensor Validation; Tomorrows Data with Yesterday's Sensors

Abstract

Since the early 1920s, sensors have been used on drilling rigs to enable making operations decisions. Drilling operators have always relied on human intervention whenever the data were questionable. Automation, remote operations, smart systems and digital technology are all words which describe the recent directional change for well construction. These changes can have a huge impact on the efficiency, safety and cost of operations. The drilling industry needs real time validation of data. With the volume of data being generated on today's rigs, this requires measures beyond human capacity. There are many gaps which need to be addressed if the industry is to make this standard in the future. For example, to make automation to work, the data fed into the smart systems for decision making must be cleaned, corrected and calibrated. This paper presents how machine learning techniques from other industries can be modified and applied to the data from drilling rig sensors. There are numerous strategies employed in medical, space and government agencies where validating data has been important to success. Of these strategies, a few were selected for drilling operation applications. An auto-encoder was used to study low dimensional data representations in an unsupervised manner. This algorithm was then adjusted to allow for partial reconstruction and handling processed drilling data. Reconstruction errors were used as a metric to identify potential errors and highlight them for the domain experts. Additionally, to identify errors, artificial potential errors were injected into the system and then the system was tested to understand if the errors could be identified by our methods. The results show that we can correctly identify anomalies such as missing data, outliers and sensor drift, using reconstruction errors. The model presented in this paper is based on drilling data from thousands of wells and tested on additional data as well as simulated data.

Sensor Validation for Indoor Air Quality using Machine Learning

“…Another line of future work is to compare point-wise detection as done here, where a single data point must be classified, to temporal classification, where a series of data points is passed to the model (e.g. [Loy-Benitez et al 2020] and [Gupta et al 2020]). The latter requires larger models and can introduce lag to the detection, which is a trade-off that can impact its usefulness in embedded and real-time settings.…”

“…More recently, several works propose the usage of Recurrent Neural Networks or Convolutional Neural Networks to analyse a time-series data to detect faults [Loy-Benitez et al 2020, Gupta et al 2020, Eren 2017, with promising results. These models are however quite complex, often requiring dedicated hardware for training and inference.…”