How AI predicts turbine component breakdowns: Predictive Maintenance Part 2

In Part One of our Predictive Maintenance series we talked about philosophies and the history of maintenance (Click here for Part 1). The second part of the series focuses on wind turbines in specific and includes examples of how Tarentum’s Wind Suite solution identified a possible major component breakdown in advance to avoid production loss and equipment costs totaling in the hundreds of thousands of dollars range.

Predictive maintenance module of Wind Suite is an interactive SaaS platform where users are able to keep track of the condition of their turbines and also see the previously generated alarms. The platform also sends automated emails to warn users of a potential breakdown of a component.

A snapshot of the predictive maintenance module is shown in the figure below. Fault alarms are organized in the form of hierarchical drop down lists.

Multiple sensory sources are employed together by the ensemble of smart fault detection algorithms of Tarentum AI to give automated fault alarms. Some of these used sources are:

  • Accelerometer (vibration) sensors
  • SCADA sensors

Vibration sensors are devices that measure the vibration or acceleration of motion of a component (e. g . Generator drive-end bearing). Generally these have a transducer that converts the mechanical force from vibration or movement change into electric current using the piezoelectric effect. 

They are mostly sensitive when it comes to mechanical degradation in rotating components. Nowadays most critical rotating equipment such as wind turbines, gas turbines, steam turbines, gearboxes, centrifugal compressors, high voltage motors and large centrifugal pumps are equipped with these sensors and monitored continuously.

SCADA is the abbreviation for Supervisory Control And Data Acquisition system which has been installed in most medium voltage scale wind turbines. It provides a rich source of continuous time monitoring that can be used to monitor overall turbine performance. Fault detection based on data from these systems has brought significant benefits for wind farm operators.

Fault Detection with the help of AI

Tarentum AI’s data-driven approach is able to capture the normal behaviour of a turbine by using its historical SCADA and vibration data. During the training session, a set of anomaly detection models are created for each sub-component (e.g. Gearbox rotor bearing) of each turbine and trained with the historical data. The model parameters are optimised with the help of system logs of fault / part replacement / service dates.

Depending on the sensors available, Tarentum AI models are able to detect faults in the following parts of a wind turbine:

  • Gearbox
    • 1st planetary stage
    • 2nd planetary stage
    • Low speed stage
    • Intermediate speed stage
    • High speed stage
    • Rotor bearing
  • Generator
    • Drive end bearing
    • Non drive end bearing
  • Main Bearing

Normal behaviour of each sub-component is determined by employing an ensemble of machine learning algorithms. Outputs of these algorithms are put into a smart voting mechanism which then generates the final diagnosis and automatically updates the Tarentum Wind Suite platform’s flags page.

Ensemble of algorithms employs a mixture of unsupervised learning algorithms and multi-dimensional distance metrics on the statistical and spectral features (e.g. skewness, spectral centroid) that are created from the raw data to give how an equipment subcomponent deviates from its normal behavior.

Case Study

We would like to share a case study here in which Tarentum AI algorithms successfully detects a bearing fault in the high speed shaft part of a Vestas V112 wind turbine gearbox.

As seen in the figure below, starting from March 2020, condition indicator of the gearbox subcomponents increases dramatically which leads the system to raise a fault flag on March 16th.

The graph below shows the heat map of 15 day aggregation of the condition indicators, which is actually the final output to determine the condition of a system. In the graph, the darker the cell the more severe the condition is. Tarentum AI algorithms raise flags on March 16th for the 2nd planetary stage (GBX_2PS), high speed stage (GBX_HSS), and intermediate speed stage (GBX_ISS) subcomponents as they exceed the learned and optimized thresholds.

Following an endoscopy and regular checks in June, maintenance personnel confirmed that the Gearbox HSS bearing is faulty and requires replacement. This shows that the Tarentum algorithms detected the fault three months before a costly failure occurred. The algorithms also captured the degraded performance of GBX_2PS and GBX_ISS, due to their proximity to the Gearbox HSS and the sensitivity of the accelerometer sensors.

Financial Impact:

As this fault was prevented, financial impacts can be calculated as follows. 

  • A 45 days production loss due to unplanned downtime was eliminated, which is equivalent to ~USD 130,000
  • Reducing the equipment and maintenance costs, which is roughly ~USD 35,000. 

These figures represent more of a conservative scenario, in case the failure was not prevented and would affect the whole gearbox, the financial impact would be much higher than these figures.

In general, each main component failure prevention would lead to savings of USD 30,000-250,000 loss production and USD 10,000-400,000 equipment and maintenance cost. These ranges depend on the severity of fault, wind regime, spare part availability, maintenance crew and crane availability, farm location and turbine type.

Data Pipeline

This section gives a summary of the extract, transfer and load (ETL) process. Tarentum AI predictive maintenance module builds on a serverless architecture. All computations and processes (e.g. feature extraction, model predictions, ensemble voting mechanism) run on cloud functions in parallel. 

The process flow can be summarized as follows:

  • Get raw vibration & SCADA sensor data from customer pipeline, periodically 
  • Process the vibration & SCADA sensory data
  • Extract features from sensor data for model training & prediction
  • Run sub-component level predictions by using the sensor data of the corresponding part of the turbine
  • Run smart voting mechanism to obtain final diagnostic flags, if any
  • Store the results & show updates on the module

This brings us to the end of this post. We will keep adding new posts when new case studies are prepared.


 

Ömer Eker