IEEE Systems, Man and Cybernetics Magazine - January 2023 - 7

configures four fully connected layers. The artificial neural
network configured in this article also needs a series of
iterations to adjust the weight and threshold nodes of the
network layer so that the artificial neural network converges
to the expected minimum loss.
To filter the input vectors effectively, the artificial neural
network takes the convolutional layer containing 16 convolutional
kernels as the first-layer network, and the size of the
convolutional kernels and the input vector are
1 × 3 and 1 × 4, respectively. The second-layer network is the
convolution layer containing 32 convolution kernels. The
input of this layer is the output of the convolution layer of the
previous layer, and the size of the convolution kernel is 16 × 3.
The third layer is the convolution layer containing 64 convolution
kernels. The output vector of the second convolution
layer is taken as the input vector of this layer, and the size of
the convolution kernel is 32 × 3. In this neural network, the
convolution layer takes one node as the step. The neural network
is activated by a zero-fill method to match the number of
features. Batch normalization and rectification linear units are
configured at the back of each convolutional layer to accelerate
the convergence of the neural network. The output of the
convolutional layer is fully connected to four fully connected
layers, and each fully connected layer has 64, 32, 16, and one
neuron, respectively. A rectified linear unit was installed at the
back end of the first six network layers, namely, the three convolutional
layers and the first three fully connected layers. It is
worth noting that the neural network does not place a pooling
layer between convolutional layers.
There are 26,688 parameters in this neural network, and
the parameters between the three convolutional layers
account for 61.39%, as shown in Table 1 of the supplementary
mater ial avai lable at https://doi.org/10.1109/
MSMC.2022.3211690. Input/output parameters are input
into the neural network to adjust the network layer weights
and threshold nodes through a series of iterations to
achieve the lowest loss. The loss function of the neural network
is mean-square error, and the initial learning rate is
0.00001. The weight b of the CNN is optimized by rootmean-square
propagation (RMSProp). RMSProp has a
momentum of 0.75 and a smoothing factor of 10-15. The
updating principle of b is as follows:
Eg .. g=+11
66
bb1
=-t+
t
@@
Eg
Eg k
[] +
2
t
where t, h , and k are the iteration index, learning rate, and
smoothing factor, respectively, and gt
the current iteration exponent.
We use a Xavier uniform initializer or Glorot uniform
initializer to initialize the weight of each network layer.
The initial weight is uniformly distributed in the interval of
,,
6-ff@ where
f =
ll
6
in + out
(5)
Figure 2. A security prediction structure diagram of a
wind turbine rotor system based on the CNN.
January 2023 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE
7
gt
tt t
22
075025 2
(3)
(4)
is the gradient of
where in
k
and out
k
are the number of input and output units
of the weight tensor, respectively. The neuronal bias of the
convolution layer and the full connection layer is initialized
to zero. We speed up the early learning process by giving
positive input to the rectilinear unit during initialization.
Security Decision Approach
After completing the modeling of the wind turbine rotor
system and the training and verification of the CNN, the
security prediction dataset is used to train the estimation
value of the rotor pitch angle by the CNN, and the security
decision method is studied. Figure 2 shows the structural
block diagram of the wind turbine rotor system security
prediction based on the CNN, where the input values of the
wind turbine rotor system are m
parameters are M ,r
a and v ,g and the output
~
g , ~ and .a
gm ,
output values of the system are M ,r
The input vector of the CNN is the input value a and
gm . The CNN can
m ,
~
g , ~
accurately estimate the estimated value of the pitch angle
without security risks. The residual can be obtained by
comparing the estimated value with the actual output pitch
angle of the wind turbine rotor system model. Wang and
Lum [18] point out that the security decision indicators
are defined as the following residual-related functions:
Jt ce tc^^ #hh ^^exp ()=+ -t
2
1
where
c
a
2
t0
residual signal, c
1 02 represents the weight of the instantaneous
2 02 represents the weight of the historical
residual signal, and m 2 0 is the forgetting factor representing
the decision indicator. The specific selection
method of these three parameters can be referred to in
[18]. When the security decision index is less than the
domain value of a given accuracy, the detected system is
judged to have no security risks. Otherwise, the detected
system is judged to have security risks.
Simulation Results
Analysis of Prediction Results of the Pitch Angle
As shown in Figure 3, randomly selected from 150 groups
of pitch angle of the real value of the projections and intuitive
fitting comparison, the figure can visually see that
whatever pitch angle value fluctuations, real value and the
Actuator Fault
αm
Actuator
System
CNN
Decision
Sensor Fault
Sensor
Mr
ωg
ωgm
a
mxhh x
te td (6)
2
https://doi.org/10.1109/MSMC.2022.3211690 https://doi.org/10.1109/MSMC.2022.3211690
IEEE Systems, Man and Cybernetics Magazine - January 2023

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2023
