Glossary · AI engineering and governance
Model monitoring
Also known as: ML monitoring
German: Modellüberwachung
In machine learning operations, model monitoring is the continuous observation of a deployed model's inputs, outputs, performance and technical health in operation, in order to detect drift, errors and degradation and to trigger investigation, retraining or rollback.
- Industrial AI
- AI
In one sentence
Model monitoring continuously observes a deployed model's inputs, outputs and performance to detect drift, errors and degradation.
Example
The monitoring dashboard of a defect detection model shows the reject rate, input image brightness and agreement with manual spot checks per shift.
How it applies
- Engineering: Monitor at several levels: technical health (latency, errors), input data (distribution, missing values), outputs (prediction distribution) and, where ground truth is available, accuracy.
- Operation: Define thresholds, who is notified and what they do. Monitoring without a response procedure only produces unused charts.
- Documentation: Document the monitored metrics, thresholds, responsibilities and actions. Monitoring results are evidence that the model continues to perform as validated (Post-deployment monitoring).
Model monitoring vs. condition monitoring
Condition monitoring observes the state of a machine. Model monitoring observes the state of an AI model. Where an AI model performs condition monitoring, both are needed: the machine is monitored by the model, and the model by model monitoring. For edge deployments, collect monitoring data centrally, so that the same model version can be compared across lines and sites.