Over the past decade, wind energy has become increasingly significant in
the global energy sector. Nonetheless, operation and maintenance (O&M)
account for at least one-third of the overall energy generation cost.
Condition-based maintenance (CBM) provides a remedy for this issue:
instead of scheduling maintenance, it monitors turbine components and
performs maintenance only when warnings of possible malfunctions are
All strategies related to fault detection and diagnosis of wind turbine
generators (WTGs) can be categorized as model-based approaches and
data-driven approaches. Model-based techniques rely mostly on a precise
mathematical model of the WTG and its subsystems. In contrast,
data-driven systems do not require physical or exact mathematical models
but infer the defect detection system from observed sensor data. The
latter techniques have shown to be particularly successful in recent
years for modeling complex interactions associated with wind turbines.
However, the existence of various nonlinearities in the examined issues
and measurement noise requires the adoption of complex and robust
This thesis proposes a framework for data-driven condition-based
maintenance: the objective of this work is to develop anomaly and
failure detection algorithms, that can be later used to provide
maintenance on condition. To this end, an unsupervised learning method
is provided, involving several feedforward (FNN) or recurrent (RNN) type
auto-encoder (AE) neural network models.
The dataset of this work was provided by Sirius s.r.l., a partner of
important companies in the world of renewable energies. The SCADA data,
belonging to various wind farms located in southern Italy, is collected
every 10 minutes from mechanical and thermic sensors measurements. The
considered problem also includes different turbine designs, with
distinct geometrical and mechanical features.
In the process, data from SCADA systems is acquired and clustered based
on WTGs performances, relying on key performance indicators, states of
the turbines, and alarms. The best-performing time sequences are then
selected as inputs for the subsequent training phase. In an unsupervised
learning manner, several AE models are trained in a multivariate time
series reconstruction task. During this phase, the models learn a robust
latent representation of the time series key features. When used on
unseen data, the algorithm will reconstruct the provided input sequences
and the reconstruction error is then analyzed for anomaly detection.
In this study, the different autoencoder models will then be exposed to
different regularization approaches, such as dropout and de-noise
autoencoder (DAE), to evaluate the different robustness of the models
produced. The various AE architectures are then tested in a simulated
benchmark environment, in which anomalies, noises, and faulty behaviors
are injected in order to be detected. The most promising models are then
employed in a real-world test case, where previously labeled WTG
critical events must be detected. The objective of the analysis will not
only be to detect adverse events, but also to correctly identify the
measures and subsystems implicated in the anomalies.
With this study, the efficacy of data-driven AI-powered ways to acquire
evaluations on the existence and nature of anomalies has been
demonstrated. Also, the efficacy of the method permits an evaluation of
the performance of wind farms and their subsystems.