US 2021/0201155 A1 · INTELLIGENT CONTROL METHOD FOR DYNAMIC NEURAL NETWORK-BASED VARIABLE CYCLE ENGINE

Abstract

An intelligent control method for a dynamic neural network-based variable cycle engine is provided. By adding a grey relation analysis method-based structure adjustment algorithm to the neural network training algorithm, the neural network structure is adjusted, a dynamic neural network controller is constructed, and thus the intelligent control of the variable cycle engine is realized. A dynamic neural network is trained through the grey relation analysis method-based network structure adjustment algorithm designed by the present invention, and an intelligent controller of the dynamic neural network-based variable cycle engine is constructed. Thus, the problem of coupling between nonlinear multiple variables caused by the increase of control variables of the variable cycle engine and the problem that the traditional control method relies too much on model accuracy are effectively solved.

An intelligent control method for a dynamic neural network-based variable cycle engine, comprising the following steps:

Step 1:
constructing a training dataset of a dynamic neural network
Step 1.1:
taking 0.01 s as a sample period, collecting the operating parameters of the variable cycle engine at the set height and PLA, including an actual value and a desired value of a relative rotary speed of high pressure rotor nh of a controlled variable, a desired value and an actual value of a pressure ratio π, and control variables, wherein the control variables include the actual operating values of fuel flow Wƒ, nozzle throat area A8, high-pressure turbine vane area AHTG, angle of outlet guide vane αƒ, angle of core driven fan guide vane αCDFS, angle of compressor guide vane αc, angle of high-pressure compressor guide vane αturb, and core driven fan mixer area ACDFS;
Step 1.2:
performing data processing on the operating parameters of the variable cycle engine collected in the step 1.1, and deleting an outlier, i.e. exception value;
and for a missing value, performing mean interpolation of the same class;
Step 1.3:
taking the deviation Δnh between the desired value and the actual value of the relative rotary speed of high pressure rotor nh in the operating parameters of the variable cycle engine on which data processing is performed and the deviation Δπ between the desired value and the actual value of the pressure ratio π of the engine as input parameters of the dynamic neural network, and taking the change value of the control variables as the target output of the dynamic neural network;
constructing a training dataset of the dynamic neural network:
x=[Δnh,Δπ]
y=[Wƒ,αturb,A8,αCDFS,ACDFS,AHTG,αc]
h=[x,y]
where x represents the input parameter of the dynamic neural network, y represents the target output of the dynamic neural network, and h represents the training dataset of the dynamic neural network;
Step 1.4:
normalizing the training dataset of the dynamic neural network:
h
norm
=
h
-
h
min
h
max
-
h
min
where hnorm, hmin and hmax respectively represent data normalized value of the training dataset h of the dynamic neural network, minimum value and maximum value;
Step 2:
training the dynamic neural network
Step 2.1:
taking 80% of the data normalized value hnorm of training dataset h of the dynamic neural network obtained in the step 1.4 as the training sample, and 20% as the test sample;
Step 2.2:
initializing the parameters of the dynamic neural network:
number of neurons in an input layer, number of neurons in all hidden layers, number of neurons in an output layer, iteration number, training accuracy ε and learning rate μ;
Step 2.3:
calculating the state and activation value of each layer of the dynamic neural network:
{
a
(
l
)

(
k
)
=
x

(
k
)
,
l
=
1
a
(
l
)

(
k
)
=
f

(
z
(
l
)

(
k
)
)
,
2
≤
l
≤
L


z
(
l
)

(
k
)
=
Θ
(
l
)

(
k
)

a
(
l
-
1
)

(
k
)
+
b
(
l
)

(
k
)
,
2
≤
l
≤
L
where L represents the total number of layers of the dynamic neural network, ƒ(⋅) represents the activation function of the neuron;
Θ(l)(k)∈Rnl×nl−1 represents the weight matrix of the l−1th layer to the lth layer when learning the input x(k) of the kth sample;
nir represents the number of neurons in the lth layer;
b(l)(k)=(b1(l)(k), b2(l)(k), . . . , bnl(l)(k))T∈Rnl represents the bias of the l−1th layer to the lth layer when the input is x(k);
z(l)(k)=(z1(l)(k), z2(l)(k), . . . , znl(l)(k))T∈Rnl represents the state of the neurons in the lth layer when the input is x(k);
and
a(l)(k)=(a1(l)(k), a2(l)(k), . . . , anl(l)(k))T∈Rnl represents the activation value, i.e. output value, of the neurons in the lth layer when the input is x(k);
Step 2.4:
calculating the error and loss of the output layer of the dynamic neural network:
e(k)=y(k)−o(k)
E(k)=½∥e(k)∥2
where y(k) represents the desired output of the kth sample, o(k) represents the actual output generated by the dynamic neural network for the input x(k) of the kth sample, e(k) represents the error of the output layer at this time, E(k) represents the loss of the output layer when learning the kth sample currently, and ∥□∥ represents the norm operation;
calculating δ(l) as the intermediate quality of error back propagation:
{
δ
(
l
)

(
k
)
=
δ
(
l
+
1
)

(
k
)

Θ
(
l
+
1
)

(
k
)

a
(
l
)

(
k
)

(
1
-
a
(
l
)

(
k
)
)
,
2
≤
l
<
L
δ
(
l
)

(
k
)
=
-
(
y

(
k
)
-
a
(
L
)

(
k
)
)

a
(
L
)

(
k
)

(
1
-
a
(
L
)

(
k
)
)
,
l
=
L
updating the weight matrix and bias parameter matrix of the dynamic neural network:
Θ(l)(k+1)=Θ(l)(k)−μδ(l)(k)a(l−1)(k)
b(l)(k+1)=b(l)(k)−μδ(l)(k)
Step 2.5:
performing structure increase judgment of the dynamic neural network
calculating the average loss E:
E
=
1
N

∑
k
=
1
N

E

(
k
)
of all samples after training once, where N represents the total number of samples;
calculating the average value
1
N
w

∑
i
=
1
N
w

E
i
of the loss of the output layer in a current sample window, where Nw represents the size of a sliding window;
if Ei>ε is met after the ith training, adding neurons from the hidden layer 1, i.e. layer 2 of the dynamic neural network, l=2;
initializing the weight θnew of the newly added neurons with a random number, and updating the number of neurons in the lth layer of the dynamic neural network and the weight matrix during the l−1th training:
nl(i+1)=nl(i)+1
Θ(l)(i+1)=[Θ(l)(i),θnew]
after adding neurons to the current lth layer (1<7<L), if after continuous Nw training, the loss change rate of the output layer of the dynamic neural network is less than the threshold η3:
1
N
w

(
E
(
i
-
Nw
)
-
E
i
)
<
η
3
and Ei>ε is met, going to the next hidden layer for neuron increase judgment, and ending structure increase judgment until l=L;
Step 2.6:
after stopping adding neurons to the hidden layers of the dynamic neural network, performing structure pruning judgment;
taking the hidden layer node output a(l)(a1(l), a2(l), . . . , anl(l))T∈Rnl as a comparison serial and the dynamic neural network output o=(o1, o2, . . . , onL) as a reference serial, calculating a correlation coefficient ξnl between a(l)=(a1(l), a2(l), . . . , anl(l))T∈Rnl and o=(o1, o2, . . . , onL):
ξ
n
l

(
k
)
=
min
n

min
k


o

(
k
)
-
a
n
(
l
)

(
k
)

+
ρ


max
n

max
k


o

(
k
)
-
a
n
(
l
)

(
k
)


o

(
k
)
-
a
n
(
l
)

(
k
)

+
ρ


max
n

max
k


o

(
k
)
-
a
n
(
l
)

(
k
)

where ρ∈(0, +∞) represents a distinguishing coefficient, and ξnl(k) (k=1, 2, . . . , N) represents a grey relation serial;
calculating the relativity using the average value method:
γ
n
l
=
1
N

∑
k
=
1
N

ξ
n
l

(
k
)
where γnl∈[0, 1] represents the relativity;
comparing the relativity γnl between various neurons in the lth layer, deleting the neuron with the minimum relativity, updating the number nl of the neurons in the lth layer of the dynamic neural network and the weight matrix Θ(l);
if the loss of the output layer of the dynamic neural network is increased too much so that Ei+1−Ei>η4 after a certain neuron in the lth layer (1<l<L) is deleted, where η4>0 represents the error floating threshold, withdrawing the deleting operation in the previous step and going to the structure adjustment judgment of the next hidden layer, and ending structure pruning judgment until l=L;
Step 2.7:
if the error meets a given accuracy requirement, ending the training, and saving the weight matrix and bias parameter matrix of the dynamic neural network;
and if the error does not meet the given accuracy requirement, proceeding the iterative training until the accuracy requirement is met or the specified iteration number is reached;
Step 2.8:
based on the test sample, checking the currently trained dynamic neural network, and calculating the test error;
Step 3:
constructing a dynamic neural network controller
encapsulating the trained dynamic neural network as a variable cycle engine controller, and normalizing the deviation Δnh between the desired value and the actual value of the relative rotary speed of high pressure rotor nh of the controlled variable and the deviation Δπ between the desired value and the actual value of the pressure ratio π of the engine and then taking same as input parameters of the dynamic neural network controller;
and inversely normalizing the output parameters of the dynamic neural network and then taking same as the control variables of the variable cycle engine, to realize the intelligent control of the variable cycle engine.

451 more elements

Intelligent control method for dynamic neural network-based variable cycle engine

Life of this patent

Intelligent control method for dynamic neural network-based variable cycle engine

Life of this patent

Abstract

Claims1