Wind turbine gearbox fault prognosis using high-frequency monitoring SCADA data

Abstract

With growing wind energy capacity, especially offshore, reliability of
wind turbines (WT) becomes a relevant concern. Poor reliability directly
affects their cost effectiveness due to increased operation and maintenance
(O&M) costs and reduced availability to generate power because of downtime.
This certainly encourages WT operators to employ
advanced O&M methodologies and focus on the critical components to reduce
failure rate, time to repair and maximizing WT performance. Condition monitoring (CM) of wind turbines for
the purpose of prognostics and health management of critical equipments can
improve the reliability and reduce maintenance costs by identifying failures at
the earliest possible stage and by eliminating unnecessary scheduled
maintenance. In contrast to the expensive purpose-built condition monitoring
systems, a SCADA (Supervisory Control and Data Acquisition System) data-based
condition monitoring system uses data already collected at the wind turbine
controller and provides a cost-effective way to monitor wind turbines.

This research focusses on developing a prognostics framework for WT
gearboxes, which are one of the costliest subsystems to maintain during a
turbine’s life. The framework follows a data-driven approach and combines two
machine learning algorithms – Artificial Neural Network and Support Vector
Machine to capture anomalous operations of the WT gearbox.  A real-time monitoring scheme is developed to
track the degradation and set a maintenance alarm as the first evident
signature of failure is identified.  The
framework was implemented using high-frequency SCADA data and was able to
detect gearbox failure, a month in advance, providing enough lead time to plan
and perform required maintenance activities. Additionally, a sensitivity study
is conducted to determine an optimal sampling frequency of SCADA data which can
be used for CM purposes as the current industry practice of storing it as 10
min averages leads to a loss of information about the condition of a WT
component. 

The results show
that the feed­forward ANN can efficiently learn the complex mapping between the
input and output features. To analyse the error between ANN predictions and the
in-field measurements, four residual error features ­ maximum error, minimum
error, root mean squared error and error distribution are used as inputs for
the OC­SVM model to understand the complex boundary between normal and
anomalous operation. The percentage of anomalies computed for each week of
operation, 4 months before failure, show an increasing trend as the turbine
approaches failure. To determine a threshold for maintenance alert, a real­time
monitoring scheme based on linear regression and bootstrapped confidence
intervals is developed to track the progression of anomalies and alarm a
maintenance alert as the first indication of incipient fault becomes evident.
The scheme alarms for maintenance a month before the actual failure, providing
enough lead time to plan and maintain the gearbox.

A sensitivity
study is carried out for a range of sampling periods ranging from 100 Hz to 10
min. The results demonstrate that high­frequency SCADA data is beneficial for
condition monitoring of the gearbox, but only if the noise in the data can be
excluded. On the other hand, despite the loss of information due to the
averaging effect for large sampling periods, SCADA data aggregated over a 30 s
period could be utilized to predict the gearbox failure a month in advance.
Furthermore, the ANN model performance is found to be sensitive to the number
of data samples available for training.