Industry is increasingly reliant on widespread and sophisticated instrumentation of assets to monitor performance, optimise operation, and facilitate condition-based maintenance. Some of our previous posts talk about using self-organising maps for simplifying visualisation of complex multivariate systems and human-in-the-loop AI systems for continuous dynamic plant performance optimisation. The potential benefits of machine learning and AI are huge, but ultimately, they all rely on reliable and accurate plant data.
Over time the accuracy of sensors can diminish, with the data they produce drifting away from actual values and providing an increasingly distorted picture to plant operators. In some cases, such sensors may be linked to automated control systems, such as chemical dosing equipment or flow controllers. As increasingly inaccurate data is used to drive plant control, downstream process compliance falls, eventually impacting plant performance. This may result in inadequate quality of plant output, excessive and inefficient chemical dosing, increased power consumption and other operational problems.
As the number of sensors on a plant rises and their recorded data is increasingly relied upon to control automated systems and inform operator decisions, the need to ensure sensors are correctly calibrated also rises. In some cases, particularly with chemical analysis equipment, continuously checking instruments to ensure calibration can represent a major demand on maintenance resources, eating into the benefits provided by robust instrumentation. Where plant data is underutilised, the burden of regular calibration may even outweigh the benefits of instrumentation.
While instrumentation can enable automated plant condition monitoring, the sensors themselves can also benefit from the same approach. Machine learning techniques allow us to create digital twins of plant subsystems featuring accurate simulations of expected plant condition, known as ‘soft sensors’. These can be used as a benchmark against which to assess actual sensor data, thereby allowing for continuous assessment of sensor accuracy / health. Depending on the application, the digital twin may also be supported by physical simulations, logic-based assessments of multiple sensors, and time series data analysis.
By providing a robust picture of the health of each sensor, our approach allows operators to shift to condition-based calibrations. This can result in a reduction in maintenance effort while also reducing the number of plant issues caused by sensor drift, delivering an overall improvement in performance,efficiency and data reliability.
The same technology which allows for assessment of sensor accuracy can also be used to obtain insights into plant performance beyond that which is possible based only on existing instrumentation. Digital twins built to model periods of good plant performance can be used to quantify system degradation, even if the source of degradation is not directly measured.
At Ada-Mode, this technology has been applied to a steam turbine with a suspected steam leak. A machine learning model was developed on an identical but sealed and healthy turbine to predict turbine speed using predictors such as steam temperature, pressure, and inlet valve position. The subject turbine was found to reach speeds lower than expected. The difference between the observed and predicted speeds was used to measure the impact of the leak in terms of lost speed. This proxy indicator of degradation identified when the leak started, how it has been affected by various maintenance activities and how it was propagating – all insights that were not possible based on normal review of plant sensor data. You can read more about this case study here.
Training digital twins to produce accurate soft sensors requires access to historic plant data and sensor calibration records which may not always be available. In the case of newer plants, or other situations where records are not available, unsupervised techniques with greater reliance on physical modelling can be applied to produce similar results.
In some cases, plant design and data availability may not allow for production of an accurate digital twin / soft sensor. A key first step to implementation is working with operators and/or domain experts to build a good understanding of the plant, allowing for initial review of the data to draw conclusions on if current plant configuration is suitable for deployment of soft sensors.
Sensor drift detection systems can be of benefit to any industrial systems which rely on accurate plant data. In particular, they are of great value for ensuring high accuracy of sensors which drive automated control systems where poor quality data directly leads to drops in plant performance.
