The Tractian platform uses AI models that continuously learn from the data captured. This learning is updated in a controlled manner, based on customer feedback, operational events, and, in some cases, manual adjustments. Understanding how and when the model is adjusted is essential to correctly interpret the insights generated (or not generated).
Customer Feedback
Feedback provided on an insight is the most direct way to interact with the AI. It allows you to calibrate the sensitivity of alerts and adjust the model's behavior based on the asset’s real operation.
When clicking “Check” on an insight, customers can access different types of evaluations. Each type serves a specific purpose — some adjust AI learning, others only document the situation.
Feedback that adjusts learning:
Process Adjustment: Informs that the detected behavior results from a planned change in the process. The AI learns this new pattern as the asset’s new healthy behavior.
No Failure: Indicates that the insight was a false positive. The pattern is learned as normal behavior.
Sensor Adjustment: Indicates that the pattern changed due to a sensor’s installation or position change. The AI will always be retrained in this case.
Feedback that does not adjust learning:
Potential Failure: The asset is not failing but should continue to be monitored. No learning adjustment is made.
Functional Failure: Confirms that the insight was correct. No retraining is needed.
Dismissed: The insight was discarded, and no inspection was performed. The system only records the feedback without altering the model.
Events That Trigger AI Relearning
Important note: To learn more about creating Events, read this article.
Some operational or hardware changes significantly impact the captured vibration data and automatically trigger AI retraining. Below are the main events that generate this behavior:
Sensor Replacement: Different sensors, even of the same model, may have slight variations in frequency response or mounting stiffness. This disrupts data consistency and requires retraining.
Rotation Change in Technical Sheet: The recorded rotation is used as a reference for identifying machine-related frequencies. Changing this value directly impacts the fault analysis algorithms, requiring model adjustment.
Scheduled Downtime (Maintenance): During maintenance, it’s common for sensors to be repositioned or asset parts replaced. This can drastically change vibration behavior. In these cases:
AI waits 5 to 7 days before relearning to ensure that the new behavior is stable.
Process Adjustment: Operational changes like load, speed, or usage mode alterations impact the vibration pattern. The AI learns this new behavior as the new "normal" operation.
AI Retraining via the Status Tab
You can directly select the asset for which AI will be retrained. Follow the steps below:
Go to the "Status" tab.
Select the asset.
Scroll the left-side content until you find the "AI Training" section.
Click "Retrain Model."
Choose between Velocity and Acceleration.
Then, you’ll need to select a period that contains at least 100 samples with the machine running and represents between 5 and 30 days of operation.
To select the period, adjust the dates until you reach the ideal range, measured by the number of samples and days shown on the right side.
