Anomaly Model Properties

The Properties tab is where you configure the anomaly model before building it. It is organized into two sections in the left panel: General Settings and Advanced Properties.

General Settings

General settings define the data scope used for training.

Setting	Description
Model Name	Display name for the anomaly model.
Entity	The entity type for which anomaly detection will be performed.
Fields	Numeric telemetry fields to analyze, e.g. temperature, vibration, energy consumption. All fields must belong to the same entity type — combining fields from different entity types is not supported.
Items	Specific devices or assets to include in training. Use the most representatively normal items for better accuracy.
Time Range	Historical data period to use for training, e.g. past 3 months or 1 year.

General Settings panel showing model name, entity selection, telemetry fields, items, and date range

General Settings panel with entity and telemetry field selection highlighted

Advanced Properties

Trendz uses unsupervised machine learning for anomaly detection — no labeled data or pre-defined fault conditions are required. The model learns what “normal” looks like from historical telemetry and flags deviations automatically.

  Raw telemetry
       │
       ▼
  ┌────────────────┐
  │  Segmentation  │  Split data into fixed-length windows
  └────────┬───────┘
           │
           ▼
  ┌────────────────┐
  │  Normalization │  Scale values for comparability
  └────────┬───────┘
           │
           ▼
  ┌────────────────┐
  │ Feature extract│  Describe the shape of each segment
  └────────┬───────┘
           │
           ▼
  ┌────────────────┐
  │   Clustering   │  Group segments into "normal" clusters
  └────────┬───────┘
           │
           ▼
  ┌────────────────────────────────┐
  │  Anomaly Score = distance from │
  │  nearest cluster centroid      │
  └────────────────────────────────┘

Segmentation — continuous telemetry is split into fixed-length time windows. Configure window size, type, and gap tolerances in Segmentation Settings.
Normalization — values within each segment are scaled automatically so that fields with different units and ranges can be compared fairly.
Feature extraction — each segment is described either by statistical attributes (mean, std, slope) or by its overall shape. Configure the representation in Anomaly Detection Settings.
Clustering — segments are grouped into clusters representing distinct normal operating behaviors. Choose the algorithm and tune its parameters in Clustering Settings.
Anomaly scoring — each segment receives a score equal to its distance from the nearest cluster centroid. Configure the score threshold and filtering in Anomaly Detection Settings.

Clustering Settings

Clustering settings determine which algorithm groups data segments into “normal” clusters. Anomalies are segments that don’t fit well into any cluster.

  Segments
     │
     ├── K-Means ──────▶  Fixed # of spherical clusters
     │                    Anomaly = far from centroid
     │                    Best for: clean, consistent data
     │
     ├── DBSCAN ───────▶  Density-based clusters, any shape
     │                    Anomaly = isolated / low-density point
     │                    Best for: noisy data, varied shapes
     │
     └── GMM ──────────▶  Probabilistic mixture of Gaussians
                          Anomaly = low probability point
                          Best for: flexible, overlapping clusters

Setting	Description
Cluster Algorithm	K-Means, DBSCAN, or GMM (Gaussian Mixture Model).
Cluster Count	Number of clusters to create (K-Means and GMM). Set this to the number of distinct typical operating behaviors in your data — for example, use 3 if your equipment has three recognizable normal states.
Max Iterations	How many refinement passes K-Means performs.
Max Epsilon	DBSCAN — maximum distance between points to be considered neighbours.
Min Points in Cluster	DBSCAN — minimum neighbours needed to form a valid cluster.
Distance Function	How similarity between segments is measured: DTW (handles similar shapes with different timing), Euclidean (straight-line), Chebyshev (largest single difference), Manhattan (sum of differences), or Canberra (emphasizes small values).

Use DBSCAN for noisy or irregularly shaped data, K-Means for clean well-separated clusters, and GMM when cluster boundaries overlap.

Clustering Settings panel with K-Means algorithm selected, showing cluster count and distance function fields

Segmentation Settings

Segmentation splits continuous telemetry into fixed-length time windows before clustering.

  Raw telemetry stream
  ────────────────────────────────────────────────▶ time

  Fixed Range (non-overlapping):
  ╔══════════╗ ╔══════════╗ ╔══════════╗
  ║ segment1 ║ ║ segment2 ║ ║ segment3 ║
  ╚══════════╝ ╚══════════╝ ╚══════════╝

  Sliding Window (overlapping, step = 50%):
  ╔══════════╗
  ║ segment1 ║
  ╚══════════╝
        ╔══════════╗
        ║ segment2 ║
        ╚══════════╝
              ╔══════════╗
              ║ segment3 ║
              ╚══════════╝

  ▶ Sliding Window produces more segments than Fixed Range
    due to overlap — this increases training time and memory.

Setting	Description
Segment Time Range	Duration of each data window, e.g. 6 hours, 1 day. Match this to natural cycles of your system.
Time Window Type	Fixed Range — non-overlapping windows. Sliding Window — overlapping windows that advance by a fixed step. Sliding produces more segments, which increases training time and memory usage.
Sliding Step Percent	Step size as a percentage of the segment duration (Sliding Window only).
Max Time Gap in Segment	Maximum allowed gap between consecutive data points within a segment. Segments with larger gaps are discarded.
Min Segment Duration	Minimum elapsed time for a segment to be considered valid.
Min Points in Segment	Minimum number of data points required for a valid segment.

Segmentation Settings panel showing segment time range, window type, sliding step, and gap configuration

Anomaly Detection Settings

After segmentation, each segment is scored for how much it deviates from normal clusters.

Comparison Type controls how segments are represented before comparison:

Feature-based — extracts statistical attributes from each segment (mean, std, slope). Focuses on specific measurable properties of the data.
Behavior-based — captures the overall shape and trend of the segment by aggregating it into a fixed number of points.

Behavior-based additional parameters:

Setting	Description
Aggregation	Function used to summarize each sub-interval: AVG, MIN, MAX, SUM, or COUNT.
Aggregation Points	Number of sub-intervals each segment is divided into before aggregation.

Scoring and filtering:

Setting	Description
Score Threshold Percent	Percentile cutoff for classifying a segment as anomalous. For example, 15% means the top 15% of highest-scoring segments are marked as anomalies.
Min Anomaly Duration	Minimum duration an anomalous event must last to be reported. Filters out very short noise spikes.
Cluster Joining Strategy	When enabled, treats anomalies from consecutive segments — even across different clusters — as a single continuous anomaly. Two anomalous segments are joined when the gap between them does not exceed the Max Time Gap in Segment threshold.

Anomaly Detection Settings panel with Feature-based comparison type selected

Limits

Limits cap the volume of data processed during training and detection. When a limit is reached, the data is truncated to the specified count — training continues with the available subset. Set limits conservatively on large datasets to avoid long training times.

Setting	Description
Max Points Count	Maximum number of individual telemetry data points used for training. Excess points are truncated.
Max Segments Count	Maximum number of segments the model will process. Excess segments are truncated.

Limits settings panel showing Max Point Count and Max Segments Count fields

Building the Model

Once all properties are configured, use the Build model button in the top right to train the model for the first time. The button is labeled Rebuild on models that have already been built at least once.

New anomaly model Properties tab with the Build model button highlighted in the top right

Clicking Rebuild opens a confirmation dialog. Rebuilding retrains the model from scratch — all previously detected anomalies are lost. Optionally, check I want to create a copy of this model to preserve the current configuration as a new model before rebuilding starts.

Rebuild confirmation dialog warning that all previously detected anomalies will be lost

Rebuild dialog with the create a copy option checked, showing the Create copy button

Best Practices

Area	Recommendation
Training data	Use data that represents normal behavior. Build on all devices first to identify the most representative ones, then retrain on those.
Segmentation	Match segment duration to natural cycles of your system (e.g. shift length, maintenance intervals).
Clustering algorithm	Use DBSCAN for noisy or irregularly shaped data; K-Means for clean, well-separated clusters; GMM when cluster boundaries overlap.
Score threshold	Start with a higher threshold (fewer anomalies) and lower it gradually as you validate results.
Performance	Set Limits conservatively on large datasets to avoid long training times.