Analytical sensors rarely fail out of the blue. They usually leave clues first. A pH sensor's slope creeps down, or its reference junction impedance climbs. A dissolved oxygen sensor slows down as its membrane or electrolyte ages. A conductivity cell reads differently once fouling builds up. If you can see those changes early, you can act before a bad reading becomes a bad batch.
That's where AI is useful in process analytics. It doesn't replace the measurement physics, and it shouldn't turn your pH reading into a prediction. It looks at the data around the measurement: diagnostics, calibration history, process conditions, and maintenance records.
Start with the sensor data you already have
The most practical use of AI today is sensor health and drift detection. Modern digital analyzers can report calibration slope, zero offset, impedance, response behavior, and temperature alongside the process value. With enough history, even a simple statistical model can tell normal aging from something that needs attention.
You don't need a giant model to do this. You need clean history. That means knowing which sensor produced which measurement, when it was calibrated or cleaned, what the process was doing at the time, and when the sensor was swapped out. In many plants, the hard part is lining up historian data with maintenance records after tag changes and sensor replacements. Fixing that is real engineering work, but it pays off, because a trustworthy sensor history helps every project that comes after it.
This is where METTLER TOLEDO's Intelligent Sensor Management (ISM) platform has a real head start. ISM sensors are digital, and they store their own calibration data and history inside the sensor head. Calibration information travels with the sensor instead of living in a spreadsheet or someone's memory. ISM also gives you diagnostics like the Dynamic Lifetime Indicator, Adaptive Calibration Timer, and Time to Maintenance. Those tools estimate remaining sensor life and suggest when to calibrate or service, based on how the sensor has actually been used rather than a fixed calendar interval.
To be clear, these are established diagnostics, not machine learning. They are also the kind of structured, sensor-level data that makes machine learning practical when you want to go further.
The right mix of physics and learning
Compensation is another area where AI can add value, if it's done carefully. Temperature effects on pH and conductivity are well understood, and those equations should stay in place. Nobody should replace working physics with a black box.
What a learned model can do is estimate the leftover error that the physical model misses. Pressure, ionic strength, coating, transport limits, and slow dynamic response can all matter in the right application. This hybrid approach is easier to defend than a black box. It also gives you a natural boundary, because the model only covers the conditions it was validated for.
Those boundaries matter, because a model built for one stream doesn't automatically work on another. Any compensation model needs a defined operating envelope, validation against independent references, and a way to flag when it's outside that range. Regulated applications need extra care. Pharmaceutical water conductivity checks, for example, follow specific compendial procedures with their own rules about temperature compensation. Check those requirements before layering anything adaptive on top.
Seeing the whole process, not just one loop
Most processes involve several related variables: pH, ORP, conductivity, dissolved oxygen, temperature, and flow. Looking at them together can reveal a contamination event or feed change that no single alarm would catch.
But moving together doesn't mean causing each other. Two sensors might shift at the same time because the process changed, or because they share a temperature input, a grounding fault, or a sample-line problem. A useful system has to tell process behavior from instrument behavior. That takes diagnostics and equipment status, not just raw values.
This is another reason smart sensors matter. When each sensor reports its own health, your analytics can rule out an instrument problem before blaming the process. Multi-parameter transmitters such as METTLER TOLEDO's M800 bring several measurements and their diagnostics into one place, which makes that kind of cross-checking much simpler.
Where METTLER TOLEDO fits in
METTLER TOLEDO's Process Analytics portfolio covers the measurements most plants depend on: pH and ORP, dissolved oxygen, conductivity, dissolved CO2, turbidity, and more, plus Thornton solutions for pure and ultrapure water, including TOC analysis. The Thornton line is widely used in pharmaceutical and power applications, where water quality has to be proven, not assumed.
A few things stand out in practice:
- Digital sensors: ISM sensors keep their own calibration and usage history, so replacing or moving a sensor doesn't wipe out its record.
- Diagnostics built in: lifetime and maintenance indicators help teams service sensors when needed instead of on a guess.
- Robust hardware: InPro-series sensors are designed for demanding processes, including sterilizable and CIP/SIP-capable options for biopharma and food and beverage.
- Optical dissolved oxygen: sensors such as the InPro 6860i offer a fast response and reduced maintenance compared with traditional electrochemical designs.
- Flexible integration: transmitters connect to common industrial communication and control systems, so the data reaches the tools your engineers already use.
If you're planning an analytics project, confirm the exact sensor and transmitter combinations for your application with a METTLER TOLEDO specialist.
Edge AI and soft sensors: useful, with conditions
Running models close to the instrument is practical now. For most analyzer applications, a drift detector or change-point method doesn't need a large neural network. Simple algorithms run fine on gateways and edge computers. The real question is whether you can validate, version, secure, and maintain that model the way you maintain the instrument itself.
Soft sensors deserve the same discipline. They can estimate variables that are slow, expensive, or hard to measure continuously, but a calculated value isn't a physical measurement. It needs reliable reference data, validation across the operating range, drift monitoring, and a clear plan for when the process leaves its training conditions. The bar is higher when the output feeds regulatory reporting, product release, safety functions, or closed-loop control.
Autonomy is a different conversation
Monitoring, diagnostics, and decision support are ready today. Handing an AI system authority over a safety- or quality-critical control function is another matter entirely. It requires bounded, repeatable behavior, a defined envelope, safe failure, and cybersecurity that accounts for the risk of a compromised model affecting physical equipment.
The sensible path is gradual: monitor first, advise next, and automate only where validation and safeguards justify it.
A practical place to begin
- Get the data in order. Bring calibration records, diagnostics, process data, and maintenance history together, with dependable sensor identities and timestamps.
- Use simple methods first. Basic drift and anomaly detection go a long way, and engineers find them easier to trust than complex models.
- Add complexity when it earns its place. Sensor fusion, compensation models, soft sensors, and edge analytics can follow as the need appears.
In many plants, the instruments are already good enough. The bigger opportunity is in the data and workflow around them. The wins won't be dramatic: a sensor flagged before it fails, a calibration problem caught before it produces a bad reading, or several small changes recognized as one developing event. That's not a substitute for engineering judgment. It's a way to give it better information.
