Select an algorithm
- Last UpdatedMar 24, 2023
- 3 minute read
The following provides more information on Process Anomaly Detection and Asset Anomaly Detection algorithms. Use this information to help determine which algorithm to use.
|
Process Anomaly Detection |
Asset Anomaly Detection |
|
|
How often does the algorithm execute? |
Every 15 minutes. |
Every 15 minutes. |
|
What is the algorithm's strength? |
This algorithm is best for process anomaly detection, although you may also use it to find reliability anomalies. |
This algorithm is best for equipment performance anomaly detection. |
|
How does it handle equipment that regularly starts and stops several times an hour (or other irregular intervals)? |
This analyzes data on a rolling 2 hour window. Any filtered conditions (such as offline conditions) during that period prevent execution of the model. |
This is a better choice because it evaluates only on the current conditions at each execution interval. |
|
How does it handle per-product segmentation? |
This provides more accurate results when operating conditions vary by product/recipe/SKU. The Process Anomaly Detection algorithm can segment the training data according to product (assuming you provide a tag containing the product ID/name*). |
The Asset Anomaly Detection algorithm does not allow for user-specified segmentation based on product type. |
|
How does it help identify the root cause of problems? |
It indicates the top three contributors to the anomaly, including the type of variation for each sensor. |
It calculates a predicted value for each input sensor, which identifies tags that deviate from training data and suggests where the problems are. |
|
What happens when executed? |
Every 15 minutes, it compares the most recent 2 hours of data to the historical training data set. |
Every 15 minutes, it compares current conditions to the historical training data set. |
|
How much computer time does training take? |
It can take from several minutes to up to 4 hours, depending on number of tags and data size. |
It should take a few seconds. |
|
What overall anomaly score is considered "normal" training period conditions? |
0 |
0 |
|
What do higher scores indicate? |
They indicate that current conditions are seldom found in historical behavior. 100 reflects the rarest condition in the training period. Greater than 100 indicates that current conditions have deviated farther from any historical behavior. |
They indicate that current conditions are deviating farther from historical behavior. |
|
What score is considered an anomaly (and generates a News story)? |
85 |
Five anomaly scores above 10 (within 90 minutes of the first anomaly detection). |
|
After a News story is generated, how is the anomaly "reset?" |
A score below 65. |
90 minutes elapse after the first anomaly score over 10. |
*Note that this tag should be an integer or string with a set of discrete values, not an analog/continuous tag. Do not include this tag in the modeled points. Only use it as the operational mode indicator.
Using both algorithms
-
When deploying a model across a large fleet of similar assets, you can run models from both algorithms against a single asset.
-
Configure the same tag selection, filter conditions, and training date ranges for each model.
-
Observe each model's results to see which results are better.
-
Roll out models from the preferred algorithm to the same fleet.
-
-
For any algorithm, set filter conditions carefully. Aggressive filters may exclude too much data from the training set. Loose filters might include unwanted conditions (like startups).
-
When your asset or process has many sensors, you can improve accuracy by grouping related sensors and creating multiple models.
For example, configure mechanical sensors (such as bearing temperature, vibrations, lube oil temperature) into one model and process sensors (such as process flows, temperature, pressures, valve positions) into another model. The same sensor can appear in multiple models. The motor tag might appear in both models to indicate the expected operating point of the equipment.