How AI and ML can improve sensor integrity

The integrity of sensors and actuators is critical to the safe and profitable operations of industrial processes. However, the lack of visibility into the heath of those sensors and actuators makes it challenging to ensure their integrity.  Intentional actions (e.g., tampering, data manipulation, data injection, sensor replacement) and unintentional events (e.g., miscalibration, sensor drift) impact integrity.  

The slightest sensor variation can have a rippling effect on production rate, scrap, and waste. Sensor integrity affects consumer-facing issues such as safety, customer satisfaction, and higher warranty costs. Nielsen conducted a survey for Advanced Technology Services and founded that the average cost of poor-quality calibration costs manufacturers $1,734,000 each year. Large companies with revenues of more than $1 billion cost balloons to an average of $4,000,000 annually. 

An ARC Advisory Group 2017 study estimated the process industry losses are as much as US$ 1 trillion per year due to unplanned downtime[1].  Misconfiguration or inconsistent sensor configurations were significant contributors to this unplanned downtime. An extreme example is the March 23, 2005 explosion at the BP Texas City Refinery in Texas City, Texas that killed 15 workers and injured 180 more. One of the major contributing factors was inoperative alarms and level sensors in the ISOM process unit[2]. Financial losses in the incident totaled $1.5 billion.[3].

Sensor Integrity Challenges

Since sensor integrity is paramount to the industrial world, why is it so challenging to ensure integrity?  The main challenges are the following.

Limited Infrastructure

Most sensors have extremely limited computing power and storage – if any at all.  Industrial facilities cannot implement standard techniques on the sensor level, such as encryption, digital certificates, and embedded digital asset management tools because of this lack of infrastructure.

Sensor Network Complexity

The wide range of physical networks (e.g., analog, RS232, RS485, Ethernet, bus) and network protocols (e.g., 4/20 Ma, MODBUS, HART, Foundation Field Bus) makes it challenging to have one technology that can work with such a wide range of networks. 

Process Areas Are Costly

Another factor is the high cost of installing additional hardware in the process area. Not only is there potential loss of production while installing the additional hardware but in areas such as oil and gas, the devices must be intricately safe – significantly adding to their cost. 

Ever-Increasing Number of Sensors

By 2020, 80% of industrial manufacturing companies will be adopting IIoT technology.[4] This adoption of IIoT is causing sensor numbers to increase exponentially, making it hard for traditional methods to scale to these IIoT numbers.

Lack of Visibility

Once data is digitalized into the TCP/IP realm, there has been tremendous progress in securing the industrial world.  However, there has been little success in identifying integrity threats at the sensor and PLC levels.  The ANSI/ISA-95 reference model can best explain this concept.  ISA-95 defines various levels within the process industry and the associated functions performed in those levels.  The definitions of the levels are the following[5]:

Level 0 

This level is where data creation begins and represents the actual production process.

Level 1    

Defines the activities involved in sensing and manipulating the physical processes such as sensors, actuators, I/O devices.

Level 2    

Defines the activities of monitoring and controlling the physical processes such as Supervisory control and data acquisition (SCADA), Human-Machine Interface (HMI), distributed control systems (DCS), programmable logic controller (PLC), remote terminal unit (RTU).

Level 3    

Defines the activities of the workflow to produce the desired end products such as Work management systems (WMS), laboratory information management system (LIMS), plant historians

Level 4    

Defines the business-related activities needed to manage a manufacturing operation, such as Enterprise resource planning (ERP).

Most systems start working when the data passes through ISA-95 Level 2 into ISA-95 Level 3 (e.g., Programmable Logic Controller (PLC) digitizes an analog signal and places the reading value onto Ethernet). However, they cannot detect compromised data on Level 2 (e.g., PLC, controllers, RTU) or Level 1 (e.g., sensors and actuators).  The lack of visibility into level 1 and level 2 is why the Stuxnet cyberattack on the Natanz nuclear facility was successful.

Making Actionable Information

Finally, to ensure integrity, detections must be made actionable.  Simply recording in a log file will not work.  To easily interpret detections, a solution should interface those detections into existing industrial systems such as control systems, cybersecurity centers, and asset management systems.

Using AI and ML

Using artificial intelligence (AI) and machine learning (ML) can help overcome those challenges to improving sensor integrity. An excellent example of how industrial users can apply AI and ML is a company called iXDen. It uses a patented "biometric" multifactor behavior technology to identify sensor integrity threats.  The first step is to build a model of "normal" behavior for a collection of sensors and actuators.  t models individual behaviors (e.g., the temperature never goes past 100 degrees C). The model also considers interactions and interdependencies between the system elements (such as when a valve closes, the discharge pressure decreases).

In addition to sensor readings, the model also examining other system elements and parameters, such as doing a checksum on software to detect software changes (e.g., potential tampering).  iXDen then adjusts the model for changes in the operating mode (e.g., the system behavior might be different on the weekends) and natural drift (e.g., engine oil temperature raises as it gets dirtier). After model creation, IXDen interrogates for current conditions. Comparing how closely current conditions match the model determines the system's health – the closer the match, the healthier the system. To express the match between modeled and observed, iXDen computes a health grade. The higher the health grade – the healthier and more reliable the industrial system under surveillance. 

The solution has robust detection rates and low false positives because it examines behavior in context. Since it has modeled all the sensors and their related behaviors, it understands that when the valve closes and the discharge pressure does not drop, this is not "normal" behavior.  Examining the sensors readings in isolation could not detect such failures. Using AI and ML to analyze the complete environment provides the "multifactor" component of the solution.

How Is Using AI and ML Different?

Multifactor Approach

A key advantage of using AI is its ability to take a multifactor approach. An AI model considers many factors and their interdependencies, just not sensor readings. This multifactor approach views sensors in context, resulting in a much higher detection rate and a much lower nuisance alarm rate. Most solutions only look at one aspect – such as the sensor value – which means they cannot detect the more complex integrity issues. Applying AI and ML on the sensor level can identify integrity risks that other systems operating higher on the ISA95 stack would miss.

Full Spectrum

Another key difference is that AI can provide the detection of a full spectrum of threats.  The method works for both intentional and unintentional acts.  It also works for any device, on any network, using any protocol rather than implementing various niche solutions (e.g., one for MODBUS, one for 4/20ma).


Systems such as IXDen take the AI model results to express the system's health in an easy-to-understand health grade. The more "normal" the system behavior, the higher the grade. This health grade allows for the rapid assimilation of information to quickly assess an industrial system's health. A health grade provides for easy integration into SCADA/DCS, asset management systems, and cybersecurity systems. 


Unlike many other systems that use historical or "data at rest," AI models use real-time interrogation to assess current conditions.   Users can configure the interrogation frequency to reflect the resource's criticality – the more critical the asset, the more frequent the interrogation.

How do AI and ML generate value?


The most outstanding value is that industrial facilities can trust their data.  Data is worthless to decision-makers using data riddled with errors.  Those errors may lead to wrong decision-making results.[6] An estimated 40% of the industrial sensor has data integrity issues.  With AI, one can detect those issues and resolve them, allowing for making decisions with useful, reliable data.


Using AI lowers costs since it does not need additional hardware and does not need any software on the sensors. Data interrogation can be either an agent-based solution (putting a small agent on the sensor or controller) or an agentless solution (using protocols like OPC to interrogate) to match the various industrial situations. And since it is a full-spectrum solution, there is no need to implement multiple solutions – lower both implementation and on-going support costs.

Improved Reliability

AI identifies "abnormal" behavior in the early stages of failure. This approach dramatically reduces unplanned outages. 

Reactive to Proactive

AI can detect integrity issues early on while there is still time to correct the problem. Without this detection capacity, operations are always reactive, responding to outages and malfunctions in more of a "run to failure" approach.

Condition-based maintenance: A quick win

However – perhaps one of the most immediate benefits is how AI can move an organization to condition-based sensor maintenance and calibration strategy.  Condition-based maintenance is one of the quickest ways to reduce costs, improve reliability, and increase availability.  Condition-based maintenance is when maintenance is performed based on the asset's actual condition rather than using some interval (e.g., preventive maintenance). 

The problem with preventative maintenance is that it is difficult to determine the right maintenance/calibration frequency.  Select an interval that is too frequent; then, one is performing unneeded maintenance and calibrations. Eliminating this unnecessary maintenance can significantly reduce costs by reducing labor (performing less maintenance), eliminating third-party calibration services, and decreasing production losses due to maintenance activities (e.g., taking the production offline).  

Selecting an interval that is not frequent enough can also result in losses such as producing products out of specification, loss of revenue (e.g., flow meters in oil and gas), loss of production due to reliability issues, and increase safety/environmental risks. Historically, determining the correct frequency has been a very manual process that involved trial and error. AI can eliminate the cost and complexity of selecting the appropriate maintenance frequency by identifying "abnormal" deviations while understanding when variations are "normal."

For example, the AI model would comprehend that it is normal for discharge pressure to increase when the pump RPMs increases.  However, it would also realize that it is abnormal for the discharge pressure not to rise with increased RPMs and would flag that event. A properly tuned AI model can even understand when environmental factors (e.g., temperature, pressure) cause "normal" sensor drift verse when drift is not associated with other factors (i.e., recalibrate is likely needed). 

What are the savings associated with moving to a conditioned-based strategy?  Savings will be highly dependent on individual situations; however, examining work orders can compute those savings. Look for all the sensor-related work orders closed out with "No problem found" or where there was little or no recalibration performed. The condition of the asset did not warrant performing those activities, and these work orders were unnecessary. Compute the cost of those work orders by looking at labor costs, part costs, and production loss.

This computed cost is the savings that could be realized by not performing unnecessary maintenance. Also, examine the work orders for significant recalibrations. Compute the cost of the asset not being calibrated correctly. For example, if an oil and gas flow meter was reading five percent too low, there was an under-reporting of production and most likely resulting in loss of revenue. A proper AI can identify when maintenance/calibrations are needed based on the asset's actual condition and reducing the losses associated with not performing maintenance frequently enough while reducing un-needed maintenance. 

Another way to view the problem is with a cost chart. The chart below shows a curve of the costs related to sensor maintenance and associated losses related to those maintenance activities. On the Y-axis is the cost relative to the optimal solution. Fifteen percent would represent costs that were 15% more than the optimal solution.  The X-axis is the average reliability of the device.  Generally, reliability is measured by how far a sensor is out of calibration but includes other factors such as being inoperable. The curve to the right of the red "X" shows the costs associated with performing maintenance too frequently – the more unneeded maintenance/recalibration performed, the greater the costs.The curve to the left of the red "X" shows the costs associated with too infrequent maintenance– the greater the asset is out of specification, the greater the loss. The red "X" is the optimal frequency of maintenance, and the AI model can determine that spot without the trial and error of conventional methods. 

maintenance and calibration cost curve
Calibration/maintenance cost curve.

The integrity of sensors and actuators is critical to safe and efficient industrial operations. Using AI and ML at the sensor level provides a cost-effective solution that significantly improves sensor integrity over a wide range of sensors, networks, and protocols. Since the AI model operates at the sensor level, it can identify intentional and unintentional threats at their source – all without having specialized hardware in the process area. Improving integrity boosts productivity, increases margins, and improves safety – all of which contribute to safe and profitable industrial operations.

Article sources:

[1] “PAS Releases Sensor Data Integrity”, Automation World, February 5, 2021 accessed February 10, 2021 from

[2] Tom Price 2005, ‘What Went Wrong: Oil Refinery Disaster’, Popular Mechanics, September 14, 2005, accessed January 5, 2021 from

[3] Jennufer Busick 2017, ‘The Cost of Catastrophe: Is There a Business Case for Chewmical Safety?’, EHS Daily Advisor, May 15, 2017, accessed January 5, 2021 from

[4] Gilad David Maayan 2020, ‘The IoT Rundown For 2020: Stats, Risks, and Solutions’, Security Today, January 13, 2020, accessed January 4, 2021 from

[5] ‘ISA 95 Framework and Layers’ ,, accessed on January 12, 2021

[6] Hui Yie Teh, Andreas Kempa-Liehr, Kevin Wang, ‘Sensor Data Quality: A Systematic Review’, Journal of Big Data, February 11, 2020, accessed January 8, 2021 at

Author Bio

Dave Lafferty has over 40 years of experience in the oil and gas industry, David provides business development, pre-sales support, advisory services, and digital transformation consulting for the oil and gas segment, emphasizing its digitalization.  Mr. Lafferty is considered a thought leader in the digitalization of Oil and Gas. He has made over 40 presentations on IIoT, machine learning, analytics, "drones," and blockchain.  Besides technical presentations, roles included serving as conference chairman, providing keynote "landscape setting" presentations, and chairing discussion sessions.  David has also developed an "IoT for Oil and Gas Masterclass" that offers a day-long deep dive into the technology stacks of sensors, communications, big data, analytics, and visualization.  Before STS, David was at the BP Chief Technology Office (CTO), where he spearheaded a wide range of innovative projects focused on wireless sensor systems including WirelessHART, ISA100, corrosion and passive sensors to predictive analytics, including equipment health; and mobility projects, including rounds and readings. David has a BS in Computer Science from the University of Wisconsin – La Crosse.

AI Platform Set To Modernize Process Manufacturing

AI Cloud Apps Advance Smart Manufacturing

IoT In Manufacturing Market Showing Extreme Growth