Application Research of Substation Power Inspection Robot Based on

Image Recognition Technology

Zeyu Li

Zhonghuan Information College, Tianjin University of Technology, TianJin, 300380, China

Keywords: Faster Region-CNN(RCNN), YOLO, Image Recognition Technology, Substation Intelligent Inspection

Robot, Image Recognition Algorithms.

Abstract: With the continuous development of the economy, substation intelligent inspection robot is widely used.

Currently, the substation inspection robot has problems such as low accuracy and poor pointer accuracy in

recognition technology. Therefore, this paper takes image recognition technology as the breakthrough point

in instrument recognition, aiming to explore the application of instrument recognition in power inspection

robot and analyze the factors affecting the image recognition of inspection robot. At the same time, this paper

summarizes and explores the advantages and disadvantages of image recognition technology in application.

This paper concluded that the traditional recognition technology algorithm needs the support of prior

information in the recognition to be able to accurately identify and sensitive to the data quality, while in the

noise environment, it will be seriously disturbed and the generalization ability is weak. The image recognition

algorithm has high accuracy and fast speed in identifying small images. Faster Region-CNN(RCNN) can

accurately identify multiple targets and save time. Through continuous improvement, the You Only Look

Once(YOLO) algorithm has high accuracy and strong anti-interference in single target recognition. In the

future, the development of image recognition technology in various fields of instrument recognition can be

more accurate.

1 INTRODUCTION

With the development of science and technology,

industrial production and people 's lives rely more and

more on electricity. Therefore, it is necessary to

ensure the smooth operation of the power systems.

The failure of the power systems often brings

unpredictable losses to the national economy, and the

state puts forward higher requirements for the stable

operation of the power systems. In order to ensure the

stable operation of the power systems, the substation

plays a key role as a transfer station for transmission

and distribution in the power systems. Therefore,

ensuring the normal operation of power equipment in

substations is particularly important.

With the continuous development of science and

technology, substations have gradually become

intelligent. Power systems nationwide have carried

out comprehensive transformation and innovation of

existing old substations. However, due to various

unfavorable factors, there are still many problems in

the process of replacing traditional substations with

modern intelligent substations. For example, a large

number of old instruments are still retained in the

traditional substation. These instruments cannot

automatically upload digital data because they cannot

be connected to the computer. This problem leads to

the traditional substation needs to manually copy, and

manual inspection will inevitably lead to missed

detection and omission. In addition, due to the large

number of old-fashioned instruments, risk of manual

inspection increases. For example, in the harsh

environment, the labor intensity of manual climbing

is high, and the scope of work is significantly

increased. At the same time, it will also cause

personal issues safety, visual fatigue and

misunderstanding caused by inattention, which will

lead to accidents. Therefore, the development of

intelligent inspection robots is of great significance to

compensate for traditional inspection methods'

limitations.

Many scholars have conducted profound research

on the application of intelligent robots in substations.

Wei Chao 's (2021) research mainly focuses on the

basic principle and application of intelligent

inspection robot inspection systems. He pointed out

Li, Z.

Application Research of Substation Power Inspection Robot Based on Image Recognition Technology.

DOI: 10.5220/0012866500004547

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Data Science and Engineering (ICDSE 2024), pages 589-593

ISBN: 978-989-758-690-3

589

that the inspection robot reduced the staff's workload

and improved the work efficiency in substation

inspection, infrared temperature measurement,

instrument data reading and equipment defect

tracking and detection. However, the existing

inspection robots have some problems in the accuracy

and efficiency of instrument recognition technology,

which is difficult to meet the actual needs. According

to the algorithm with You Only Look Once(YOLO)

v3 as the core, Song Dongmei (2023) improves a

single target's detection efficiency and accuracy by

fewer network layers. Wang Xinran et al (2024)

proposed an improved Faster Region-CNN(R-CNN)

instrument recognition algorithm using residual

network (Resnet) 101 instead of the original Visual

Geometry Group(VGG). The improved Faster RCNN

has higher recognition accuracy for single and

multiple targets. Sun Hui et al.proposed to quickly

identify the position of the instrument by introducing

the Convolutional Neural Network(CNN) detection

model and perform targeted denoising on the

identified image using the noise capture algorithm.

For the influence of the external illumination

environment, an improved single-parameter

homomorphic filter is proposed, which reduces the

amount of parameter adjustment and shortens the

operation time. Finally, Wan Jilin et al.improved the

image detection network RetinaNet and added a

higher resolution fusion feature map on the original

basis to solve the problem of too little information in

the identification of small parts and used the

correction method of the small probability of the

transformer to correct the detection of small parts.

This algorithm improves the recognition accuracy of

small parts.

Based on the above research background, this

paper aims to explore the application of image

recognition technology in the instrument recognition

of power inspection robot in substation. First of all,

this paper analyzes the development process by

understanding the basic principles of image

recognition technology. Secondly, this paper focuses

on the specific application of image recognition

technology and further explores its advantages.

Finally, this paper analyzes the advantages and

limitations of image recognition technology based on

image recognition technology in instrument

recognition technology, and then gives some

prospects and suggestions.

2 IMAGE RECOGNITION

TECHNOLOGY

2.1 Rationale

Image recognition technology, also known as

computer vision technology, is a technology that can

transform the input image information into

meaningful digital or text information through digital

image processing technology (Wang and Liu 2023).

It mainly relies on computer technology to analyze,

identify and process digital images. Through these

steps, image recognition technology can quickly and

accurately identify the information contained in

countless images. Image recognition technology is

based on the object's shape, size, color, texture and

other features, according to these features to analyze

and identify the objects in the image. The working

principle of image recognition technology includes

image acquisition, image preprocessing, feature

extraction, pattern matching, recognition and

classification.

Image acquisition refers to converting images in

the real world into electronic signals through

cameras, scanners and other devices and storing them

in computers (Zhang et al 2023). Image preprocessing

is the performance of operations such as denoising,

enhancement, and edge detection on the acquired

image during the preprocessing process to improve

the image quality and highlight the characteristics of

the target object. Feature extraction is an important

part of image recognition. By extracting features from

the image, the target object in the image can be

distinguished from other backgrounds. The

commonly used feature extraction methods include

color features, texture features, shape features, etc.

Classification and recognition is that the computer

uses machine learning, deep learning and other

algorithms to classify the extracted features in the

classification and recognition stage, so as to realize

the automatic recognition of the target objects in the

image. The commonly used algorithms are VGG

network, edge detection, YOLO, Faster R-CNN and

other algorithms.

2.2 Application of Image Recognition

Technology

With the rapid development of artificial intelligence

and computer vision technology, the current image

recognition technology mainly includes face

recognition, object detection, scene classification,

text recognition, image content editing, medical

ICDSE 2024 - International Conference on Data Science and Engineering

590

image analysis, video surveillance, virtual reality and

augmented reality, intelligent driving assistance and

agricultural pest hazard detection and other fields.

Face recognition technology is the use of

computer algorithms, through the analysis of facial

features, to achieve automatic recognition and

classification of faces. This technology is widely used

in road safety, finance and other fields, such as access

control systems and mobile payment. Object

detection is the use of computer vision technology to

realize the automatic detection and recognition of

objects in images. This technology is used in

intelligent construction, smart home intelligent

security, and other aspects such as defect detection on

automated production lines and intelligent vacuum

cleaners. Scene classification is to automatically

classify and label images according to the scene

information in the image. Scene classification

technology is reflected in intelligent photo albums,

intelligent search engines and other software, such as

automatic classification of photos according to the

scene, intelligent recommendation related content.

Image content editing is to realize the automatic

editing and modification of image content. Image

content editing is mostly used in image programming

software such as Meitu Xiuxiu and Photoshop, which

can automatically repair photo defects and realize

intelligent whitening. Image recognition technology

is widely used in the field of intelligent transportation.

For example, traffic surveillance cameras can learn

vehicle identification, vehicle counting and vehicle

violation identification through image recognition

technology, so as to improve the efficiency and

accuracy of traffic management. In the field of

medical imaging, image recognition technology can

automatically analyze and recognize medical image

images to help doctors make early diagnosis and

treatment of diseases. For example, the early

detection of breast cancer can automatically identify

potential tumor areas through image recognition

technology to improve the speed and accuracy of

diagnosis. In the field of security monitoring, security

cameras can realize human face recognition, behavior

recognition and other functions through image

recognition technology, helping monitors to detect

anomalies in time and provide effective security

warnings. In the field of Unmanned Aerial Vehicle,

image recognition is mainly used for target tracking,

terrain recognition and so on. UAV can automatically

track the target object, carry out real-time shooting

and monitoring, and carry out ground recognition to

realize the autonomous navigation and flight control

of UAV. In intelligent robots, image recognition

technology is an indispensable part of robot

intelligence. Robots can identify objects and people

in the environment for autonomous navigation, target

tracking, human-computer interaction and other

tasks. In addition, in the field of service robots, robots

can improve the service quality of robots in terms of

face recognition and emotion recognition. At present,

there are robots in major amusement parks to identify

players in real time through cameras. They can also

analyze facial expressions to achieve emotional

interaction and improve the game experience. The

intelligent inspection robot of the substation monitors

and maintains the power equipment through the

ability of image recognition technology, such as

autonomous perception, autonomous planning,

autonomous execution and autonomous learning,

such as abnormal oil level of a transformer, excessive

winding temperature, insulation damage, instrument

aging and fouling. The intelligent inspection robot

can independently analyze the corresponding

decision-making according to the problems that arise.

2.3 The Characteristics and

Classification of Image Recognition

Technology Algorithm

The traditional algorithms of image recognition

technology include: edge detection, morphological

processing, linear fitting and other algorithms.

Edge detection is used to identify significant

changes in the image, usually representing the

boundary of the object. Instrument detection can be

used to identify the edge of the scale line.

Morphological processing This is a method of

analyzing shapes. Some shape features can be

extracted or enhanced from the image through

morphological processing. For example, it can be

used to remove noise, connect broken lines or find

objects of a specific shape. Line fitting : This may be

necessary for recognizing lines (e.g., scale lines) in an

image. The algorithm can determine which points

should be regarded as a straight line and give its

parameters (such as slope and intercept).

The CNN algorithm can directly use the

instrument image as input, output the dial, pointer and

scale after feature extraction and feature mapping, or

output the indicator directly.

Image recognition technology algorithms are

divided into two categories of algorithms: one-stage

and two-stage.

The Faster R-CNN algorithm belongs to the two-

stage algorithm. Its structure mainly includes

convolution layer, RPN layer, region of interest

pooling layer and classification regression layer.

Faster R-CNN has superior performance, high-

Application Research of Substation Power Inspection Robot Based on Image Recognition Technology

591

precision detection performance, can solve multi-

scale, small target problems, versatility and

robustness.

The Scale-Invariant Feature Transform algorithm

uses the convolution of the original image and the

Gaussian kernel to establish the scale space, and

extracts the scale invariant feature points on the

Gaussian difference space pyramid. This algorithm

has certain affine invariance, perspective invariance,

rotation invariance and illumination invariance. The

first step is the pyramid, which greatly reduces the

amount of computation. The second step is the search

of feature points. The third step is the feature

description (Shi 2022).

The Speeded Up Robust Feature algorithm uses

the approximate Harris method to extract the feature

points for the shortcomings of the SIFT algorithm,

which is too slow and computationally intensive. By

using the integral image on different scales, the

approximate Harris value can be effectively

calculated, simplifying the construction of the

second-order differential template and improving the

efficiency of feature detection in the scale space

(Gao2018).

Solid State Disk target detection algorithm is an

end-to-end target detection method. SSD can achieve

target detection through a single neural network and

can detect multiple targets at the same time. The SSD

algorithm only needs one forward transmission to

detect the target. Therefore, the detection speed of the

algorithm is fast and the accuracy is high.

The RetinaNet algorithm is mainly composed of

backbone network ResNet, feature pyramid Feature

Pyramid Networks, classification sub-network and

regression sub-network. This algorithm generates

feature maps of different scales through FPN, and

constructs a feature pyramid from the feature map so

that the network can detect the target.

3 APPLICATION OF IMAGE

RECOGNITION TECHNOLOGY

IN INSTRUMENT

RECOGNITION IN

SUBSTATION

Traditional instrument recognition technology

requires reading manually according to the direction

of the instrument pointer and the scale on the dial.

When identifying the reading of the pointer

instrument, it is first necessary to shoot the instrument

through the camera. Then, according to the

information collected of the image, the area

containing the instrument pointer and the scale is

segmented. Finally, the meter reading is calculated by

the angle of the pointer rotation.

The reason why the intelligent inspection robot

can accurately identify the instrument is that it

depends on the existence of the original information.

For example, information such as categories and

range units contained on the instrument can be input

to the intelligent inspection robot, which can reduce

the error probability and also significantly reduce the

universality. The traditional recognition method is

fast and accurate for low-resolution images, but the

traditional recognition technology lacks

generalization in the face of interference. At the same

time, it needs prior information when detecting and

segmenting images, so it is difficult to apply it to the

external environment (Zhang et al 2023).

At present, instrument recognition faces two main

problems instrument detection and pointer

segmentation. In order to solve these problems, target

detection algorithms in image recognition

technology, such as Faster R-CNN and YOLO, are

applied to instrument detection. Through Mask R-

CNN, the meter pointer's area and the dial's scale is

segmented and the reading is performed. The Faster

R-CNN algorithm is optimized on the basis of the

original algorithm R-CNN, which solves the problem

of long calculation time, inability to be applied in

practice, and low accuracy in detecting small targets.

Some scholars have proposed a pointer meter reading

method R-YOLOv5 based on the rotating target

detection algorithm. This method is an improvement

on the YOLOv5 algorithm (Zhang 2023). It does not

need to detect the direction and scale of the pointer.

This method can simultaneously predict the direction

of the pointer and locate the scale. Finally, the angle

method is used to read the instrument, and then the

data is collected. Some scholars have improved the

Faster R-CNN algorithm to solve the problem that the

accuracy and efficiency are still not very high in

complex scenes such as instrument overlap and

occlusion detection. Due to the long training time,

large storage capacity, loss of information and loss of

VGG16 network, ResNet101 is introduced to replace

the original VGG16 feature network. Through

experiments on different feature extraction networks,

it is concluded that the recall rate, accuracy and time

of single target image detection of ResNet101 are

better than VGG16 (Feng 2022). Then, the original

recursive feature pyramid is improved to improve the

ability of the backbone network to extract features

and make the positioning of the target more accurate.

ICDSE 2024 - International Conference on Data Science and Engineering

592

4 THE FUTURE DEVELOPMENT

OF IMAGE RECOGNITION

TECHNOLOGY

With the further optimization of deep learning

algorithms in the future, especially the continuous

advancement of image recognition technology, more

optimization algorithms will be proposed. These

algorithms can be used to improve the speed and

accuracy of image recognition, while reducing the

consumption of computing resources. For example, a

new model structure is developed to accelerate the

training of the model. At present, image recognition

is mainly based on single mode, but in many

application scenarios, researchers need to identify

cross-modal images. For example, the radar signal or

infrared image is converted into a human-

recognizable image, and the data collected by

multiple sensors is fused to improve image

recognition accuracy. With the rise of technologies

such as artificial intelligence and unmanned driving,

image recognition technology is developing in the

direction of multi-modal fusion. For example, the

traditional image data, sound, video and so on are

fused to provide users with reliable information.

However, image recognition technology will face the

challenge of heterogeneous data. For example, in the

field of unmanned driving, image recognition

technology needs toidentify external weather, road

conditions and traffic signs accurately. This requires

image recognition technology combined with sensor

technology and real-time data processing technology

for research. In the use of image recognition

technology, it is inevitable to encounter privacy

leakage and lack of security. Researchers need to

consider how to use image recognition technology

efficiently to improve security performance in the

future.

5 CONCLUSION

This paper introduces the application of intelligent

inspection robot in substation based on image

recognition technology. It mainly takes image

recognition technology as the breakthrough point in

instrument recognition, and points out that the

accuracy of traditional instrument recognition

technology in identifying complex images is not high.

At the same time, prior information is needed when

detecting and segmenting images, so the impact on

the external environment is difficult to apply in

practice. The basic principles of image recognition

technology are described. For example, the traditional

image recognition algorithm has high accuracy and

fast speed in identifying small images. Faster RCNN

can accurately identify multiple targets and save time.

YOLO algorithm has been continuously improved.

This algorithm has high accuracy and strong anti-

interference in single target recognition. Image

recognition technology algorithms can be used to

improve the speed and accuracy of image recognition

and reduce the consumption of computing resources.

However, when identifying and collecting a large

number of images, information is sometimes lost, and

privacy protection is insufficient. The application of

cross-modal image technology is more promising

because with the development of new energy trams,

more and more people are proposing driverless

technology, which can help people with disabilities

bring convenience when they travel.

REFERENCES

C.Wei. Light source and lighting, (09) (2021)

M.D.Song. China Instruments, (01) (2023)

X.R.Wang B.Zhang and M.Zhan etc. Electromechanical

engineering (2024-02-04)

X.J.Wang, X.L.Liu. Electrical technology and economy,

9,(2023)

K.Zhang, J.Wang, J.W.Shao,etc. Power safety technology,

25 (09), (2023)

B.H.Shi.Research on automatic meter reading of substation

inspection robot.Southwest Jiaotong University,

(2022).

J.W.Gao.Research on intelligent reading method of

substation inspection robot. University of Electronic

Science and Technology of China, (2018).

W.Zhang, F.Huang, J.B.Liang. Mechanical design and

manufacturing, (05) ,(2023).

Y.R.Zhang. Pointer instrument recognition technology and

its application in inspection robot. Wuhan University of

Science and Technology, (2023).

L.Feng. Research on automatic reading method of

industrial instrument based on machine vision.Taiyuan

University of Science and Technology, (2022).

Application Research of Substation Power Inspection Robot Based on Image Recognition Technology

593