AI Model Deployment Platform
Yggio can train an AI model on your own sensor history and then score every new reading against it, without the data leaving the platform. Training and inference run as ordinary Yggio services next to the data they use, so there is no export step and no separate machine-learning stack to operate.
Note: this is not switched on by default. Contact technical support to have it enabled for your installation.
Anomaly monitoring: start with an alarm translator
Most anomaly monitoring in Yggio does not need AI. It is done with the
alarm translators, which raise a true/false alarm when a
decoded field crosses a threshold you set. They cover a plain limit, warning and critical bands, a
rate of change, a condition that must hold over several reports, time-of-day filtering, and
algorithms such as mould risk. They are available to every installation, they need no training
period, and the reason an alarm fired is visible in the threshold.
Use an AI model where the normal pattern is not something you can express as a threshold: a value whose usual range differs per device, or drifts with season or use. The rest of this page covers that case.
Anomaly detection with a model
The model type available today is anomaly detection. A model is trained on one sensor's own history, learns what is normal for that sensor, and from then on flags readings that do not fit the pattern. That covers cases such as a leak, a blocked filter, failing equipment, or a door opening at an hour it normally does not.
A model can be trained on any device measurement that is numeric: temperature, light, CO2, RSSI and so on. Non-numeric fields cannot be used.
Because each model learns one sensor's own behaviour, there is nothing to tune per device. Yggio trains more than one candidate model, compares them on data they were not trained on, and keeps whichever performed better. The comparison is recorded with the model.
The one decision left to you is how sensitive the result should be. It is set as an expected anomaly frequency: how much of the history should be treated as unusual enough to flag. That is a question about how many alarms you want to act on rather than a property of the data.
Training a model
Models are managed from the AI Fleet page, which is added to the navigation once the feature is enabled. Training runs as a short wizard:
- Model type - anomaly detection.
- Datastreams - the device and the measurement to learn from. Several can be added.
- Settings - the expected anomaly frequency and the window size, which is how many recent readings the model looks at.
- Train - Yggio fetches the history, trains and compares the candidates, stores the winner and scores the historical period so there are results to look at straight away.
From then on the model runs against every new reading from that device.
What you get back
The results arrive as ordinary readings on an ordinary Yggio device. Nothing about them is a special AI format, so everything the platform already does applies to them unchanged: chart them on a dashboard, raise an alarm on them, use them in the Rule Engine, export them, and control who can see them with access rights.
Each training run also stores its own record of what data and settings were used and how the candidates scored, so a model's provenance can be shown later.