A novel approach of wind turbine blade root load monitoring based on
bearings vibrational data and a new AE-LSTM-Attention neural network
Abstract
The monitoring of wind turbine blade root load (BRL) has always been a
valuable yet highly challenging problem. BRL monitoring is often
measured directly using strain sensors, which is expensive to install
and poses safety risks. Thus, it is necessary to explore an indirect
measurement method that establishes a mapping relationship between BRL
and easily measurable data. This paper presents a non-intrusive BRL
monitoring method based on the vibrational signals at bearings
(including main bearing, gearbox bearing, and generator bearing) and a
new proposed AE-LSTM-Attention neural network. The proposed neural
network incorporates autoencoder (AE) pre-training process, long
short-term memory (LSTM) neural network, and channel attention mechanism
to overcome difficulties in extracting and fusing long-term,
high-frequency, and multi-channel signal features. Additionally,
considering the wind turbine transmission mechanism, an optimal LSTM
step number selection criterion is proposed. The wind turbine’s physical
feature is innovatively bridged with the neural network training
parameters, e.g., setting the transmission ratio of the gearbox as the
number of LSTM steps, by which the most cost-effective prediction
performance can be achieved. Experiments were conducted using data from
actual operating wind turbines, and the results show that the proposed
method can successfully achieve non-intrusive monitoring of BRL
characteristics and outperform traditional methods in terms of
predictive performance. The mapping relationship between wind turbine
vibration signals and BRL characteristics can be established efficiently
and accurately.