Why sensors alone are not enough
A connected device measures something: a temperature, a vibration, a door opening, a heartbeat. On its own, one reading says very little. Ten thousand sensors reporting every few seconds produce far more numbers than any person can read. The hard part of the Internet of Things is often not collecting data but deciding what to do with it.
Artificial intelligence is a broad label for software that makes judgments or predictions that would otherwise need a person’s attention. In connected systems it usually means a model that has been trained on past data and then applied to new readings. The sensors act as the senses, the network carries the signals, and the model supplies the interpretation.
From readings to decisions
Most uses follow the same basic chain.
- Sense. Devices measure the physical world.
- Transmit. Readings travel over a network to a gateway, a server or a nearby computer.
- Prepare. The data is cleaned, timestamped and organized. Poor input leads to poor output, which is why data quality matters so much.
- Interpret. A model looks for patterns, estimates a value, or assigns a label such as “normal” or “unusual”.
- Act. The system sends an alert, adjusts a setting, or hands a recommendation to a person.
The interpretation step is where AI sits. The rest of the chain still has to work well, or the model has nothing reliable to learn from.
What kinds of AI are used
Several different techniques fall under the same heading, and they do different jobs.
- Classification assigns a label. Is this sound from a machine a normal hum or a worn bearing? Is this camera frame showing a person or a pet?
- Prediction estimates a number or a future event, such as how much heating a building will need tomorrow morning.
- Pattern discovery groups similar behavior without being told the categories in advance, which can reveal types of usage nobody planned for.
- Unusual-event spotting flags readings that do not fit what normally happens.
Many of these rely on machine learning, where a model is fitted to examples rather than programmed with explicit rules. Instead of writing “if the temperature exceeds a set limit, alert,” a developer shows the model lots of past data and lets it find the relationships.
Plain examples
Homes. A smart thermostat can learn when a household is usually away and lower the heating accordingly. Voice assistants turn speech into text and then into commands. In both cases the device has to cope with messy, varied real-world input.
Factories. Vibration and heat sensors on motors can feed models that estimate whether a part is wearing out. This is the idea behind predictive maintenance: fix something shortly before it fails instead of after, or on a fixed schedule regardless of condition. It builds on basic condition monitoring, adding a model that weighs several signals together.
Farms and the environment. Soil, weather and image data can help estimate when fields need water. Models are also used to sort camera images, for example to separate healthy leaves from damaged ones, though results depend heavily on lighting and local conditions.
Health and wearables. A wristband can estimate activity or sleep stages from motion and pulse signals. These estimates are approximations, not diagnoses, and a consumer device should not be treated as a medical instrument.
Cities and buildings. Traffic counts, occupancy readings and energy meter data can help forecast demand so that lighting, heating and signals respond sensibly.
Where the processing happens
A model can run in a data center or on or near the device. Each choice has tradeoffs. Running in the cloud allows larger models and easy updates but needs a connection and adds delay. Running locally keeps data close and responds quickly but is limited by the small processor and battery of the device. The choice between these is covered in the discussion of edge computing, and many real systems split the work: a small model on the device screens the data, and a larger one in the cloud handles the harder cases.
What can go wrong
AI does not remove the ordinary problems of IoT; in some ways it magnifies them.
- Bad or drifting sensors. A sensor that slowly loses accuracy produces readings that a model may treat as truth. Regular sensor calibration and checks still matter.
- Not enough or unrepresentative data. A model trained in one climate, building or factory may behave poorly in another.
- False alarms and missed events. No model is right every time. Too many false alerts and people stop listening; too few and real problems slip through.
- Hidden bias. If training data reflects only some people or conditions, decisions can be unfair or less accurate for others. This is explored under algorithmic bias.
- Privacy. Learning from data about homes, bodies and movements raises real questions about who holds that information and for how long. See the broader discussion of IoT privacy.
- Opaque reasoning. Some models cannot easily explain why they reached a conclusion, which is a problem when a decision affects safety or people’s lives.
A sensible way to approach it
Start with the problem, not the technique. Sometimes a fixed threshold or a simple average solves the job more cheaply, more reliably and more transparently than a trained model. AI tends to earn its place when patterns are too complex or too variable for hand-written rules, when there is enough good data, and when there is a clear way to check whether the output is right.
It also helps to keep a person in the loop for decisions with real consequences, to test a system on a small scale first, and to plan how the model will be monitored and updated once it is running. Connected devices operate in changing conditions, and a model that was accurate at installation can slowly become less so.