Create Anomaly Model

Trendz provides a guided three-step wizard to build an anomaly model — no manual algorithm configuration required. Pick a detection intent, scope your data, and fine-tune sensitivity. The wizard configures everything else automatically.

Two entry points are available on the Anomaly Detection page:

• Create model — button in the top-right corner, always visible on all tabs.
• + Create — inline button on any profile row in the Profile Coverage table. Opens the wizard with that profile pre-selected.

The Select Intent step asks what you want to detect. Choose the strategy that matches your goal — each option automatically pre-configures the optimal clustering, segmentation, and scoring settings:

Intent	Best for
Quick Scan (recommended)	Rapid baseline scan for obvious deviations in recent data. Best starting point for most use cases.
Operational Monitoring	Continuous alerting for deviations from known seasonal or operational patterns. Uses a rolling window for ongoing detection.
Comparative Analysis	Identifies units that behave differently from the peer group. Useful for fleet-wide benchmarking.
Health Monitoring	Detects long-term gradual drift in overall system health. Uses progressive tracking across extended time ranges.

The Configure step defines the entities, telemetry keys, and time range the model will train on.

Entity Selection — select the entity type and profile to analyze:

Item Selection — choose which devices or assets from the profile to include:

Mode	Behavior
All entities	Includes every device or asset in the selected profile.
Manual selection	Pick specific items from a searchable list.

Telemetry Selection — pick the numeric telemetry keys the model will monitor. Type to search all keys available for the selected profile:

Date Range — set the historical period for training. Use a quick preset or enter a custom range:

Data Readiness — the panel on the right updates in real time as you fill in each field:

Field	Description
Status	Overall assessment: Pending → Poor → Fair → Good → Excellent.
Entities	Number of devices or assets selected.
Telemetry keys	Number of keys selected.
Approximate points	Estimated total data points available for training.
Step frequency	Approximate interval between consecutive readings.
Time range	Duration of the selected training period.

Aim for Good or Excellent before proceeding. A Poor status typically indicates too few data points, too short a time range, or too sparse a telemetry key.

Once the status is acceptable, click Next: Configure to advance to the final step.

The Review step lets you fine-tune two parameters before the model trains.

Sensitivity — controls how aggressively the model flags anomalies:

Option	Behavior
Strict	Minimizes false positives. Only flags extreme deviations. Best when alert fatigue is a concern.
Balanced (recommended)	Best default for most industrial IoT environments.
Don’t miss anything	Maximum recall — flags every potential deviation for expert review. Expect more false positives.

Detection Granularity — the time window the model evaluates data in. Trendz suggests up to five options based on your data’s step frequency. Shorter windows catch brief spikes; longer windows surface gradual drift.

When satisfied, click Finish to submit the model for training.

The model enters the Queued → Running → Ready lifecycle. Training typically completes within seconds to a few minutes depending on data volume.

Once training completes, Trendz redirects you to the Result tab of the newly built model. It shows an immediate overview of what the model found: key metrics, an activity heatmap, ranked items by anomaly impact, and a list of top anomalies.

Review the results to verify the model is detecting meaningful anomalies before enabling continuous monitoring or alarms. See Anomaly Model Results for a detailed guide on analyzing results.

The wizard covers the most common detection scenarios. For full control over every parameter — clustering algorithm, distance function, segmentation window, score thresholds, and limits — use Advanced Mode instead.

Access it via Advanced configuration on the intent selection step. Full parameter documentation is available in Anomaly Model Properties.

Situation	Recommendation
Too many anomalies	Switch to Strict sensitivity. The model is flagging normal variation — tighten the threshold to reduce noise.
Too few anomalies	Switch to Don’t miss anything or choose a shorter detection granularity. The model may be too conservative.
Selecting telemetry keys	Only combine keys that physically correlate — e.g. vibration and temperature on the same pump. Mixing unrelated keys degrades model accuracy.
Training data quality	Use a period of normal operation. Avoid ranges covering outages, commissioning, or maintenance — they skew what the model learns as “normal.”
Using Advanced Mode	Only switch if you understand clustering and segmentation. Misconfigured parameters can produce worse results than the defaults.
Training data volume	Aim for at least 3 months of regular telemetry. Sparse or short datasets result in Poor readiness and unreliable detection.