Fault diagnosis method and apparatus for four-switch buck-boost converter

Patent metadata

Field	Value
Publication number	US-12314840-B2
Application number	US-202418640874-A
Country	US
Kind code	B2
Filing date	Apr 19, 2024
Priority date	May 9, 2023
Publication date	May 27, 2025
Grant date	May 27, 2025

Abstract

Disclosed is a fault diagnosis method for a four-switch Buck-Boost converter, including the following steps: obtaining circuit parameters when different faults occur, labeling fault parameters with fault types, and combining the fault parameters and labels into a data set; constructing a simplified self-organizing neural network consisting of a plurality of nonlinear operation modules, a weighting module, and an analysis module; performing training by using the data set to obtain a self-organizing neural network for determining the faults of the converter; and inputting the circuit parameters of the four-switch Buck-Boost converter to the self-organizing neural network, so as to determine the fault types of the four-switch Buck-Boost converter. Further disclosed is a fault diagnosis apparatus for a four-switch Buck-Boost converter. The method provided by the present invention is capable of simultaneously identifying and locating the faults of the four-switch Buck-Boost converter, such as an open circuit fault of a switching tube, aging of a capacitor, and aging of an inductor, and has good diagnosis accuracy and generalization capability.

First claim

What is claimed is:

1. A fault diagnosis method for a four-switch Buck-Boost converter, comprising the following steps:
obtaining circuit parameters when different faults occur, labeling fault parameters with fault types, and combining the fault parameters and labels into a data set;
constructing a simplified self-organizing neural network consisting of a plurality of nonlinear operation modules, a weighting module, and an analysis module, performing, by the plurality of nonlinear operation modules, eigenvalue calculation based on the inputted circuit parameters, performing, by the weighting module, weighting calculation based on all eigenvalues to obtain weighted eigenvalues, and predicting, by the analysis module, the fault types based on the weighted eigenvalues; wherein a weight of each nonlinear operator in the plurality of nonlinear operation modules is calculated by using a particle swarm optimization algorithm, and a specific process is as follows:
setting parameters of the particle swarm optimization algorithm, wherein the number of particles is N, and a particle dimension D=the number n of network layers× the number m of nonlinear operators;
adopting an inertia weight linear decreasing strategy for an inertia weight, and using an error between a predicted result and an actual result as a fitness function in the particle swarm optimization algorithm to determine an individual historical optimal position of each particle and a historical optimal position in a whole particle swarm, wherein an expression for the inertia weight linear decreasing strategy is as follows:

ω
=

ω
max

-

iter
*

(

ω
max

-

ω
min

)

iter
max

wherein ω is an inertia weight factor, and “iter” is the number of iterations; based on updated velocities and positions of the particles, recalculating fitness values of the particles and determining an individual historical optimal position of each particle and a historical optimal position in the whole particle swarm, wherein an expression for the updated velocities and positions of the particles is as follows:

v
i_d

k
+
1

=

ω
⁢

v
i_d
k

+

c
1

⁢

r
1

(

a
i_d
k

-

x
i_d
k

)

+

c
2

⁢

r
2

(

B
d
k

-

x
i_d
k

)

wherein 1≤i≤N, 1≤d≤D; k represents the number of current iterations; c 1 and c 2 are learning factors; and r 1 and r 2 are random numbers in [0, 1];
determining whether new fitness meets a requirement, and if the new fitness meets the requirement, outputting an optimal position as an optimal weight value, otherwise, performing repeated iteration until an end condition is met;
the weighting calculation process of the weighting module is as follows:
a node operator Ψ k l of an 1th layer is selected, and the node operator is linear weighting of a sine operator sin(w ki l ,y k l−1 ), an exponential operator e(w ki l ,y k l−1 )−1, and a Gaussian operator w ki l+1 e(−w ki l (y k l−1 ) 2 ), with an expression as follows:

Ψ
k
l

=

a
1
l

⁢
Sin

+

a
2
l

⁢
Exp

+

a
2
l

⁢
Gauss

wherein a 1 l , a 1 l , and a 3 l are hyperparameters obtained by training with the particle swarm optimization algorithm,
a pooling operator P i l is fixed as a summation operator, and an activation operator f i l is a relu function;
an output vector of a channel i of an (l−1)th layer is y i l−1 and is convolved with a (K×1)-dimensional convolution kernel of the corresponding channel i, and this channel is calculated by the node operator and the pooling operator:

x
k
l

⁢

(
m
)

=

b
k
l

+

∑

i
=
1

C

Assignees

• Univ Zhejiang

Inventors

• Zhang Xin
• Liu Xueqi
• Ma Hao

Classifications

• G01R31/54

  Testing for continuity · CPC title

• G01R31/40

  Testing power supplies (testing photovoltaic devices H02S50/10) · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/0464 Primary

  Convolutional networks [CNN, ConvNet] · CPC title