Granted Patent B1

US 12,314,149 · App. 18/747,957 · Granted May 27, 2025

Inventors: Yaohua Deng (Guangzhou, CN); Zilin Zhang (Guangzhou, CN)

Assignee: GUANGDONG UNIVERSITY OF TECHNOLOGY

G06F11/2733

Abstract

Claims (47)

1. A complex device fault diagnosis method based on multi-dimensional features, comprising the following steps of:

acquiring fault data X of a target complex device, preprocessing the fault data X, and carrying out input coding mapping and position coding;

acquiring fault features covering global and local deep information in combination with a feedforward neural network;

fusing the fault features to acquire a multi-dimensional feature vector; and

acquiring a fault type probability score matrix according to the multi-dimensional feature vector, and outputting a fault diagnosis result of the target complex device;

wherein, an input X o1 ′ is obtained after the position coding is imported, an attention score of each of multiple attention heads is constructed and calculated, and an output X oh of each of the multiple attention heads is obtained according to the attention score; and

outputs of the multiple attention heads are combined and spliced, and mapped through linear transformation to obtain a final multi-head self-attention output X oH , and the final multi-head self-attention output X oH is subjected to layer standardization and then subjected to residual connection with the input X o1 ′ to obtain a global feature X o2 output;

wherein, the global feature X o2 is input, and a last dimension of the global feature X o2 is segmented, and subjected to dilated convolutions with dilation rates of 1, 2 and 3 respectively; and

feature matrices of three different scales of dilated convolution outputs are added with equal weights to obtain multi-head outputs, the multi-head outputs are combined and spliced, and subjected to linear mapping to obtain X oH d , and finally, layer normalization and residual connection are carried out to obtain a local feature X o3 output;

the local feature X o3 is input for fully connected mapping and nonlinear transformation, and layer normalization and residual connection are carried out to obtain an output X o4 ;

X o4 =LN (ReLU( X o3 ·W 1 F )· W 2 F )+ X o3 ;

wherein, W 1 F and W 2 F respectively represent weight parameters of the two fully connected layers, LN refers to normalization calculation, and ReLU refers to activation function calculation; and

the outputs X o2 , X o3 and X o4 are acquired, and skip residual connection between the output and the input X o1 ′ is added to finally acquire an output X o5 as the fault features; and

the fault features are fused to acquire the multi-dimensional feature vector;

the output X o5 is input, so as to fuse the fault features into a vector with a fixed length to acquire the multi-dimensional feature vector Output GAP ;

Output

GAP

=

1

M

2

⁢

∑

i

=

1

D

e

⁢

m

⁢

b

⁢

e

⁢

d

⁢

X

o

⁢

5

i

;

wherein, X o5 i ∈ 1xM 2 represents that an i th line of the X o5 has a total of M 2 feature values, M 2 represents a sequence length, and D embed represents a feature dimension size.

2. The complex device fault diagnosis method based on multi-dimensional features according to claim 1 , wherein the preprocessing the fault data X, and carrying out the input coding mapping and the position coding, specifically comprises:

acquiring the fault data X by using a sensor for sample preprocessing of normalization and slide window segmentation, reconstructing each of multiple parts of sample data into a two-dimensional feature map X 2D through an extended operation, segmenting the two-dimensional feature map X 2D into a plurality of patches through a convolution operation, and outputting a feature vector X o1 after rearrangement; and

constructing a learnable position coding matrix, and embedding the position coding matrix into the feature vector X o1 in a form of addition for the position coding to obtain the input X o1 ′.

Assignments (1)

Priority Claims (1)

References Cited (23)