Unity AI Insights: Predictive Fault Detection (PFD)

TABLE OF CONTENTS

• Introduction 
• Process
  • Model training and modeled temperature
  • Insight logic: Events, Severity, and Decay
    • Event 
    • Severity
    • Decay
• Results
• Complementary information 
  • How PFD differs from a vibration-based monitoring system
  • Example of components monitored by PFD
• Inputs
  • Signals - Wind Turbine
  • Insights - Wind Turbine
  • Constants
• Outputs
  • Signals - Wind Turbine
  • Insights - Wind Turbine
• How to activate

Introduction 

Predictive Fault Detection (PFD) is a temperature-based AI system that models the thermal equilibrium inside wind turbine nacelle using SCADA data. It compares actual and predicted component temperatures to identify deviations caused by mechanical or electrical issues. By analyzing ambient and nacelle temperatures, power output, and rotor/generator RPM, PFD can accurately detect maintenance needs and notify users of emerging component faults.

The results of PFD are displayed as insights in the AI & Insight Insight Log tab along with other AI & Insight feature insights as shown in the image below.

Process

Model training and modeled temperature

The AI models are trained on a one year period. Training is done per turbine and per component. The required inputs are listed in the Inputs section below. Once training is completed, the feature outputs a modeled component temperature for each 10-minute timestamp. This modeled data is a visible timeseries in the platform for reference such as shown below in Data & Trends. 

Insight logic: Events, Severity, and Decay

Event 

The modeled temperature is compared to the actual value and an insight is generated when the deviation between the actual and modeled temperatures deviate by a significant amount. 

An event is created when the difference between the modeled signal and the actual signal is greater than the threshold value for at least half an hour:

(Modeled Signal – Actual Signal) > Threshold Value.

The threshold value is calculated based on the training performance of the model. The threshold value is defined as the Mean + 3*Standard deviation of the training error, with a lower limit of 0.5 applied to avoid overly sensitive models (as also noted under ‘Constants’). This method of calculating the threshold value avoids insights being generated due to high training errors as a result of relatively bad data during the training period.

A severity metric is calculated for each event, which depends on the amplitude of the deviation and its duration. This severity value cumulates or decays as explained below. 

Severity

If the cumulative severity measure crosses the severity threshold of 1200, an insight is generated. These thresholds are based on extensive testing done during the development phase of PFD. The thresholds will be updated if necessary.  

Decay

If no new events are detected by the model, the severity level begins to decay at a predefined rate. Once it drops below the set threshold, the insight is automatically closed, and a close date is assigned, but the insight remains visible for reference and historical tracking.

Results

The results are formatted as an insight, as previously mentioned. The information available is shown in the Insight Details, and Wind PFD Insights Dashboard. 

Insight Details: PFD detects deviations from normal operational behavior with high sensitivity, often providing ample lead time to address potential issues before they escalate. When an insight is triggered, we recommend starting with a quick review of the PFD dashboard. From there, check the list of recommended actions to guide your initial analysis and determine the appropriate next steps, as illustrated below.

Dashboard - Insight Log: This shows if other insights have been triggered on this same turbine, allowing the user to assess if the issue is isolated, or impacts other components. 

Dashboard - Actual Component Temperature with Modeled plot: The plot displays the modeled temperature vs the actual temperature for the component. The time frame to be displayed is chosen automatically between the First anomaly date – 1 week and the Last anomaly date + 1 week. The purpose of this plot is to show when and by how much the temperature has deviated from the expected temperature for the component.

Complementary information 

How PFD differs from a vibration-based monitoring system

A vibration-based CMS utilizes vibration data provided by specially installed sensors on the turbine. However, PFD works on measurements that are already available in the turbine SCADA system. This simplifies the onboarding process because it does not require the installation of any new sensors. Vibration-based CMS is good at detecting faults in the drivetrain system, whereas PFD is effective in detecting faults in many sub-systems of the turbine, such as the converter system, the hydraulic system, etc. PFD is especially effective in detecting issues leading to a change in the operating temperature of the component. 

Example of components monitored by PFD

• Transmission

• Drivetrain system

• Rotor system

• Power generation system

• Hydraulic system

• Nacelle

• Cooling system

Inputs

Signals - Wind Turbine