Yes and Yes. There is a strong, fact-based argument that many industries are already in a “calibration crisis”, especially if we’re talking about using hydrogen sensor data to feed AI, digital twins, safety systems, and predictive analysis maintenance.
Here is the logic. AI and predictive models only work if the historical operational data is accurate.
But across manufacturing, energy, utilities, chemical processing, battery rooms, and power infrastructure, operators openly acknowledge three recurring problems (see Solution at end):
- Sensor drift and lower reliability over time,
- Sensors that were never properly calibrated in the field, and
- Sensors that fall out of calibration and stay in service anyway because taking them offline is costly.
Everyone is racing to build AI models to promote safety, optimize assets, reduce faluires, and sell “predictive intelligence” to management and regulators.
Per McKinsey (Aug 2024 and Jan 2023 reports) 30% of deployed sensors are not properly calibrated. The result is bad real-time data being logged, not just bad alarms and alerts in the moment. But also this is exactly the scenario where you poison the very training dataset that the future AI will learn from.
You also have a structural mismatch. Everyone is racing to build AI models to promote safety, optimize assets, reduce failures, and sell “predictive intelligence” to management and regulators.
At the same time, plant instrumentation teams are understaffed, many assets are aging in place, lead times for replacements are in the years, and a meaningful slice of critical sensors in the field (nearly 1 in 3) are mis-calibrated.
So, the volume of sensor data is going up, and the level of trust in that data is not keeping up. That gap is the crisis!
The bottleneck in AI for industrial systems is not GPU capacity, not energy, not talent, not math. It is calibration integrity of frontline sensors. If the baseline hydrogen numbers going into the historical operational readings are wrong, then the AI built on top of that plant is training on fiction, and they may not even know it.
That is an existential credibility problem for any company telling investors, regulators, or even acquirers that they have effective and meaningful predictive AI analytics. And accurate sensor data in the first step in any industrial AI innovation.
Solution: Use sensors that autonomously selfcalibrate (auto-calibration is not adequate), have a record of long-term accurate operation (maintenance free for up to 10 years), and stores (logs) historical operational readings on-sensor and in the Cloud (if IoT connected) for future AI training. Contact H2Scan Corporation at www.H2Scan.com for their many options.
Note – also, temperature, moisture and pressure sensor options.
The bottleneck in AI industrial systems is not GPU capacity, not energy, not talent, not math. It is calibration integrity of frontline sensors.
Sources: McKinsey Aug/2024 & Jan/2023, Emerson/
Lilly 2022, NIST Oct/2024, INL/PNNL Sept/2019
Steve Pfrenzinger is an AI Transition Strategist, Education-based Marketing Specialist, Board Member/Advisor, MIT AI Certified, Forbes Author, Certified Exec Coach, HoF Angel Investor. His past clients include: Sony, SpaceX, Disney, Tesla, PwC, Deloitte, Oracle. See Steve’s other AI related posts on LinkedIn.
This article was originally published in the December 2025 issue of the Transformer Critical Components magazine.
View Magazine
