5 minute read 21 Jun 2019
Close up of robotics engineer fitting sensors to traditional engineering lathe in robotics research facility

How organizations can calibrate to trust IoT sensors

By

Aleksander Poniewierski

EY Global IoT Leader and Partner, EY EMEIA Advisory Center

Leading-class knowledge in IT/OT security and IT systems risk management. My passion for IoT is probably paralleled only by my interest in photography.

Contributors
5 minute read 21 Jun 2019

Show resources

IoT and associated technologies such as AI depend on accurate data. But can every sensor be trusted?

We are coming out of an era where cleaning poor quality data has been a costly exercise – the era of ERP implementations and migrations.

Now, with IoT, machines are exchanging data messages with other machines via storage in the cloud, and organizations are using AI techniques including Machine Learning (ML) to process that data at incredible speeds. 

So how, in this new era, are we managing the quality of IoT data?

The short answer is: we’re not. And we’re potentially sleepwalking into a situation where we ask AI to learn things based on flawed, poor quality data, which can very quickly lead to exponentially flawed results.

For example, at a basic level, IoT sensors are designed in fundamentally different ways according to the vendor: different nationalities of vendors will program sensors to read temperature in Celsius or Fahrenheit. And when you put data from these disparate sources into a single cloud, it’s unable to merge. That may be simple enough to fix.

But what if individual sensors start to drift in their readings? In the world of physical machinery there are specific rules for calibrating equipment unit by unit. But imagine you have thousands or millions of sensors – no human calibration or maintenance operative can ever keep up. And when you factor in a failure rate of 10-15% of IoT sensors in any given installation, the task looks even more daunting.

As things are becoming more connected and machines rely on accurate data to perform the right actions, calibrated sensors are the prerequisite to deliver on the promise of IoT. 

Where could inaccurate sensor calibration have a major impact?

Wherever there is a decision based on data coming from sensors, there is a need for calibration.

Inaccurate sensors could undermine quality and productivity for an entire sector too: manufacturing. So it would be incredibly valuable for a sensor calibration system to understand the typical operating conditions and behavior, as 99% of production environments should be very similar.

Some other examples:

  • In a refinery or chemical plant, if the recipe for a final product depends on sensors reading parameters such as pressure and temperature, faulty readings result in a lower quality product
  • In a smart city, faulty light detection sensors could mean street lights come on and go off too early or late
  • Uncalibrated sensors detecting the humidity of concrete could lead to problems when building a skyscraper
  • An airport camera wrongly calibrated could fail to spot someone exhibiting signs of a contagious illness
  • Connected fridges in retail stores that do not read the correct temperature could mean drinks are not sufficiently chilled, leading to a loss of sales

In summary, organizations that increasingly rely on sensors to manage the business need a way to achieve trust in their implementations.

Make it right at the data level before scheduled maintenance

This is exactly the issue EY wanted to address when innovating a system to provide Calibration-as-a-Service (CALIBRaaS).

EY began experimenting with data from a refinery to identify bad sensors and found that 10%-15% of data was based on incorrect readings. The neural network that was applied identified all sorts of different sensors in categories such as speed and pressure, but also “others,” which included some sensors that were not working. This grouping of sensor “families” opens more possibilities to play with the sensor data. 

Since then, EY has been evolving a proprietary methodology to develop custom AI models, which allow operators to spot that something is wrong with a sensor or clusters of sensors. Operators can then make adjustments to correct the sensors at the data level before scheduled maintenance takes place. This is an important emphasis: CALIBRaaS does not replace manual calibration, but alerts operators (and proposes predictive values) when sensor readings are incorrect, based on the self-learning model, or starting to drift by tiny increments, which could indicate the beginnings of failure.

As more and more businesses are looking to connect everything to everything, to compete effectively on customer experience, serving physical assets, and operational excellence, calibration becomes an important process.

Supporting organizations on their IoT journey

The use of unsupervised machine learning in calibration enabled EY to gain unique insights into oceans of sensors like never before. And for organizations, sensor calibration can support several different journeys, for example:

  1. Operational improvement: running scans of the sensor landscape and provide insights into the quality of data, and giving estimates of what correct data and values should look like, to provide virtually corrected data until sensors are physically calibrated or replaced
  2. Digital transformation: companies that are on the move from selling assets to services must be able to provide quality measures in an objective and independent way
  3. Autonomous business: smart factories or intelligent grids depend on correct data to perform corrections and decisions. Continuous monitoring and correction becomes essential to avoid fall-outs or safety risks

Sensor calibration: the new critical process issue of connected businesses

As more and more businesses are looking to connect everything to everything, to compete effectively on customer experience, serving physical assets, and operational excellence, calibration becomes an important process.

Calibration can take up to 30% of maintenance time, and this may increase as AI systems are becomes integrated into the data streams to monitor and handle production processes. Furthermore, multiple sources of streaming data in real-time push businesses to implement governance and processes to make certain the data is correct and can be trusted.

Uncalibrated sensors will not only impact business performance, but also an organization’s risk exposure while moving toward a more connected business. A self-learning calibration system can reduce the time to identify wrong sensors, and can also feed gaps with measures based on historical sensor behaviors. 

This is a direct route to building the necessary trust into IoT implementations, and the data they produce, from the very beginning.

Summary

In the era of IoT, machines are sharing data with each other without human intervention – or checks. What if the sensors supplying that data were faulty? And what if AI was learning from flawed data? Where IoT implementations involve thousands or even millions of sensors, there is an urgent need for constant calibration.

About this article

By

Aleksander Poniewierski

EY Global IoT Leader and Partner, EY EMEIA Advisory Center

Leading-class knowledge in IT/OT security and IT systems risk management. My passion for IoT is probably paralleled only by my interest in photography.

Contributors