Fault Diagnosis for Wind Turbine Systems by Using Neural Network and Deep Learning Techniques

Rahimilarki, Reihane (2021) Fault Diagnosis for Wind Turbine Systems by Using Neural Network and Deep Learning Techniques. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) rahimilarki.reihane_phd_17024318.pdf - Submitted Version Download (8MB) | Preview

Abstract

Concerning the fact that the number of wind turbines is increasing worldwide, it
seems necessary to implement monitoring systems. To respond to this demand,
this PhD thesis studies different fault diagnosis techniques in order to improve the
reliability and reduce maintenance costs. Based on the fact that a considerable
amount of data is stored via SCADA in every industry nowadays, the methods
developed on historical data (called data-driven methods) can be very beneficial.
By analysing the historical data, the changing trends of a nonlinear dynamics, such
as a wind turbine, can be predicted. Moreover, by applying suitable approaches,
one can distinguish different faults based on the output of the system.

The first part in this research reviews a neural network identification method
by decoupling linear and nonlinear parts of a wind turbine model. As for the
linear part, a Luenberger observer is designed, while for the nonlinear part, a
neural network observer is proposed. By having an identification model for a wind
turbine system, residual-based fault detection is studied.

The second part in this research proposes a novel neuro-robust fault estimation
method to deal with the occurred faults on actuators or sensors. The challenge in
this method is environmental disturbances and sensor noises. To overcome these
problems and simultaneously estimate the faults and the states, an augmented
system is proposed in different scenarios of actuator faults or sensor faults. Then, a
neural network updating rule is calculated along with the robust performance index
to fully achieve this goal. The stability of the augmented system is guaranteed by
having a Lyapunov function and input-to-state stability criteria.

The third and final part in this research studies different structures of Convolutional
Neural Networks for the problem of fault classification in a wind turbine.
As working with time-series signals is challenging in deep learning classification,
a pre-processing analysis is applied to prepare the data of system outputs for the
input of the model.

Each proposed method is applied to a 4.8 MW wind turbine benchmark and obtained
results are illustrated and discussed to validate the accuracy and performance
of the approach.

Actions (login required)

View Item

Downloads