Physical AI
Most of the recent progress in artificial intelligence has concerned information that already exists in digital form: text, images, code, audio. Physical AI concerns systems whose subject is the material world. Its inputs come from sensors, its conclusions concern objects and machines, and its errors have physical consequences.
That last point shapes everything. A system that summarizes a document badly wastes a reader's time. A system that misjudges the condition of a generator can cost a facility its power. Physical AI therefore has to be built with a different temperament: conservative about what it claims, explicit about its confidence and designed to fail safely.
Three layers
Sensing
The first layer converts physical phenomena into data. AFFERENT's approach is multimodal: observing an asset through several kinds of sensing at the same time, on a common clock, so that signals can be compared with one another. Read about sensing.
Edge intelligence
The second layer interprets signals where they are captured. Models that run on the device are designed to learn the normal behavior of the asset and to detect departures from it without depending on a continuous connection. Read about edge AI.
Cloud intelligence
The third layer works across devices. It is designed to compare assets, maintain their histories and train the models that are sent back to the edge. Read about AFFERENT Cloud.
Design commitments
Four commitments run through the engineering.
- Synchronized capture. Signals of different kinds are only comparable if they are aligned in time. Synchronization is treated as a first-order requirement of the device.
- Useful without a connection. Industrial networks are constrained and sometimes unavailable. The device is designed to keep observing and analyzing on its own.
- Explainable findings. A recommendation should be traceable to the signals behind it. An operator must be able to ask why.
- Security by design. A device attached to critical equipment is part of that equipment's risk. Security is designed in from the start.
Where this sits among existing approaches
None of the individual ingredients is new. Vibration analysis is a mature discipline. Industrial sensors are widely available. Machine learning for anomaly detection is well studied. Industrial IoT platforms have connected equipment to dashboards for years.
What has been missing, in our view, is integration. Sensors from one vendor, gateways from another and analytics from a third produce data that is hard to align and models that are hard to trust. AFFERENT's position is that the device, the edge model and the cloud model should be designed together, so that each can rely on the others. That is a harder way to build a company. We believe it is the way to build a system whose conclusions deserve confidence.
