US 2026/0253198 A1 · Visual Inspection Method for Crimping Quality of Automotive Wiring Harnesses

Abstract

A visual inspection method for crimping quality of automotive wiring harness. The method comprises: performing denoising, grayscale conversion, and binarization processing on a wiring harness image to determine white connected regions corresponding to wiring harnesses; calculating the number of wiring harness strands based on the white connected regions, and comparing the number with a preset standard value; if the number is consistent with the preset standard value, performing edge extraction to determine a length of the wiring harness, and determining whether the length is within a preset dimensional range; after the condition is satisfied, segmenting the wiring harness image into single-wire-strand images according to a preset order, and calculating a first-order color moment and a second-order color moment in an HSV color space for each single-wire-strand image, and comparing the calculated moments with template wiring harness values; if differences are within a threshold range, segmenting a wiring harness crimping area and generating a binary image thereof; detecting defects in the crimping area using a pre-trained defect instance segmentation model, generating a defect mask, and calculating compactness of a contact surface; if no defect is detected and the compactness conforms to a standard, completing inspection of a qualified wiring harness.

A visual inspection method for crimping quality of automotive wiring harnesses, sequentially performing wire strand sequence inspection, dimension inspection, and crimping defect inspection on a wiring harness based on an image thereof, the method comprising:

processing an image of the wiring harness through grayscale conversion, denoising, and binarization to identify white connected regions corresponding to wiring harnesses within the harness image;
determining the number of wire strands based on the white connected regions corresponding to wiring harnesses;
in response to the number of wire strands matching a first preset standard value, performing edge extraction on the harness image to obtain edge pixels of the wire strands and determining a length of the wiring harness based on the edge pixels;
in response to the length conforming to a preset dimensional range, segmenting the harness image into a plurality of single-wire-strand images according to a preset initial sequence;
extracting first-order and second-order color moments for each single-wire-strand image in an HSV color space, and comparing differences between the extracted color moments and template values for corresponding positions of single wire strands;
in response to all the differences being within a preset threshold range, cropping a wiring harness crimping area from the harness image and generating a binary image of the wiring harness crimping area;
processing the binary image using a pre-trained defect instance segmentation model for crimping areas to generate a defect mask, wherein the defect mask marks crimping defects within the wiring harness crimping area and is used to determine a compactness of the crimping area based on a perimeter and an area of a contact surface thereof, and
in response to processing no crimping defect and the compactness matching a second preset standard value, completing quality inspection for a qualified wiring harness;
wherein the pre-trained defect instance segmentation model for crimping areas is a YOLACT instance segmentation model, using ResNet50 as a backbone network;
wherein training the defect instance segmentation model for crimping areas comprises:
acquiring a plurality of images containing various types of crimping defects to construct a crimping area defect dataset, wherein the types of crimping defects comprise crack defects, abnormal crimping shape defects, crimping surface oxidation defects, and wire strand exposure defects in crimping areas:
inputting the crimping area defect dataset into a configured YOLACT instance segmentation model, and extracting feature information through the backbone network, the feature information being expressed as:
F
∈
R
C
×
H
×
W
(
9
)
where F represents the feature map dimension, C represents the number of feature channels, H represents a height of the feature map, and W represents a width of the feature map;
performing dimensionality reduction on the feature tensor F, using a 1×1 convolution kernel to reduce the number of feature channels C in the feature tensor F to 32 to serve as the number of shared mask prototypes;
determining a shared mask dimension P, expressed as:
P
∈
R
M
×
H
×
W
(
10
)
performing normalization on an original prediction value Ci of the YOLACT instance segmentation model through an activation function to obtain mask weight coefficients cim, wherein the mask weight coefficients characterize a contribution degree of shared mask prototypes to a specific target instance, expressed as:
C
i
=
{
C
i
⁢
1
,
C
i
⁢
2
,
…
,
C
i
⁢
M
}
(
11
)
c
i
⁢
m
=
sigmoid
⁢
(
C
i
)
performing a linear combination on the shared masks to generate a final instance mask {circumflex over (M)}i(x, y), thereby obtaining the trained defect instance segmentation model for crimping areas, expressed as:
M
ˆ
i
(
x
,
y
)
=
∑
m
=
1
M
⁢
C
i
⁢
m
×
P
m
(
x
,
y
)
(
13
)
where M represents the number of shared masks, {circumflex over (M)}i(x, y) represents a final mask for an i-th instance, Pm(x, y) represents an m-th shared mask prototype, and Cim represents an m-th weight for the instance i;
and
optimizing the network using a binary cross-entropy loss function to enhance an ability of the defect instance segmentation model for crimping areas to distinguish between different types of crimping defects, the loss function being expressed as:
L
m
⁢
a
⁢
s
⁢
k
=
-
1
N
⁢
∑
i
=
1
N
⁢
∑
(
x
,
y
)
[
M
i
(
x
,
y
)
⁢
log
⁡
(
M
ˆ
i
(
x
,
y
)
)
+
(
1
-
M
i
(
x
,
y
)
⁢
log
⁡
(
1
-
M
ˆ
i
(
x
,
y
)
)
)
]
(
14
)
where Mi (x, y) represents a ground truth mask annotated using Labelme software, and {circumflex over (M)}i (x, y) represents a predicted mask.

182 more elements

Visual Inspection Method for Crimping Quality of Automotive Wiring Harnesses

Life of this patent

Visual Inspection Method for Crimping Quality of Automotive Wiring Harnesses

Life of this patent

Abstract

Claims9