Computer and Modernization

Bearing Fault Diagnosis Based on CWGAN-GP and CNN

JIANG Lei, TANG Jian, YANG Chao-yue, LYU Ting-ting

2023, 0(07): 1-6. doi:10.3969/j.issn.1006-2475.2023.07.001

Asbtract ( 291 )

PDF (2107KB) ( 51 )

References | Related Articles | Metrics

Abstract： Aiming at the problem that the number of bearing fault samples is small and unbalanced in the actual work process， a bearing fault diagnosis method based on Conditional Wasserstein Generative Adversarial Network （CWGAN-GP） and Convolutional Neural Network （CNN） is proposed. First， a CWGAN-GP generative adversarial network is constructed by combining conditional generative adversarial network （CGAN） and gradient penalized Wasserstein distance-based generative adversarial network （WGAN-GP）. Then， a small number of bearing fault data samples are input into CWGAN-GP， in order to obtain high-quality samples similar to the original samples. When the network reaches the Nash equilibrium， the generated samples and the original samples are mixed to generate a new sample set. Finally， the new sample set is input into the convolutional neural network to learn the sample features for fault diagnosis. The experimental results show that the diagnostic accuracy of the diagnostic method proposed in this paper exceeds 99%， which is higher than other diagnostic methods， effectively improving the diagnostic accuracy and enhancing its generalization ability.

Natural Gas Load Forecasting Based on FCGA-LSTM and Transfer Learning

ZHANG Zhi-xia, XIE Bao-qiang

2023, 0(07): 7-12. doi:10.3969/j.issn.1006-2475.2023.07.002

Asbtract ( 191 )

PDF (1521KB) ( 53 )

References | Related Articles | Metrics

Abstract：High precision natural gas load forecasting is of great significance to the smooth and efficient operation of natural gas pipeline network. Most of the existing natural gas load forecasting methods are based on the condition of sufficient historical data， and there is little research on the problem of natural gas load forecasting in areas lacking historical data. To solve these problems， a short-term natural gas load forecasting method based on long and short-term memory（LSTM）neural network optimized by Fuzzy Coded Genetic Algorithm (FCGA) and transfer learning is proposed. First， the source domain and the target domain are selected， and the FCGA-LSTM prediction model is constructed by using a large amount of historical load data in the source domain. After model training and testing， the source domain model is moved to the target domain lacking data as a whole， and then a small amount of data in the target domain is used to fine tune and retrain the model. Finally， the target domain load prediction model is obtained. Taking a new residential area in Xi’an as an example， the results show that the prediction accuracy of the prediction method based on FCGA-LSTM and transfer learning is improved by 15.6 percentage points and 35.2 percentage points respectively compared with the combination method of LSTM and transfer learning， LSTM under non transfer learning， which proves the effectiveness of the model. The proposed method has certain guiding significance for the prediction of natural gas load in new urban areas lacking historical data.

Insertion/Deletion Genomic Variations Detection Method Based on Regional Read#br# Segments Classification#br#

LI Lan-lan, GAO Jian-long, ZHU Xiao, MU Pei-zheng

2023, 0(07): 13-19. doi:10.3969/j.issn.1006-2475.2023.07.003

Asbtract ( 124 )

PDF (1147KB) ( 22 )

Related Articles | Metrics

The long-read sequencing data produced by long-read sequencing technology， especially the accurate long-read， provides a good data basis for genome variation detection. Insertion/deletion variation is a common genomic variation and an important source of pathogenic variation. The diploid characteristics and highly repetitive structure of the human genome led to some difficulties in the detection of some complex forms of heterozygous insertion/deletion variations， and there is still room for improvement in the sensitivity and accuracy of variation detection. In order to solve the problem that the previous methods did not work well for the detection of heterozygous insertion/deletion variations in complex forms， an insertion/deletion genomic variation detection method based on reginal read segments classification is proposed. This method is based on accurate long-read. The read segments classification algorithm based on pairwise alignment is used to divide the read segments in the region into two groups at most according to the diploid characteristics of the human genome， so as to detect insertion/deletion variations more accurately. The proposed method is compared with five other common variation detection methods on two simulated datasets and one real dataset. Experimental results show that this method can improve the sensitivity of complex heterozygous insertion/deletion variations detection， and has a good effect of insertion/deletion variations detection.

CT Image Generation of Pneumonia Based on Generative Adversarial Network

WANG Jia-chen, ZHANG Hong-xin, LIU Qing-hua,

2023, 0(07): 20-24. doi:10.3969/j.issn.1006-2475.2023.07.004

Asbtract ( 118 )

PDF (1823KB) ( 37 )

Related Articles | Metrics

Aiming at the problem of fuzzy image generated by generative adversarial networks with random noise as input， difficult convergence in training， and the similarity and diversity of data features cannot be fused by traditional physical data expansion methods， this paper proposes a generative adversarial network-UG-DCGAN based on feature pyramid and attention mechanism CBAM. The method first takes the masked and denoised CT image as input to enhance the robustness of the network. Then use the feature fusion pyramid and attention mechanism to jointly establish a generative network to extract and reconstruct CT images， in which the feature fusion pyramid only retains the maximum scale fusion， and the residual structure is added in the down-sampling process to improve the feature extraction ability. Finally， the convolution layer of the discriminator network is added to improve its supervised judgment ability. After experimental verification， the results compared with the StyleGAN2 model， the FID value of the CT image generated by the algorithm is reduced by 21.98%， and the IS value is increased by 12.44%. It shows that the algorithm has obvious effects on improving the clarity of CT image generation， the similarity and diversity of features.

A Temporal Convolutional Knowledge Tracking Model Based on#br# Multiple Feature Extraction#br#

XIE Shi-bin, LIU Meng-chi, TANG Shi-qi, ZHOU Rui-ping,

2023, 0(07): 25-29. doi:10.3969/j.issn.1006-2475.2023.07.005

Asbtract ( 154 )

PDF (2268KB) ( 51 )

Related Articles | Metrics

Knowledge tracing （KT） is a key technology in the field of educational data mining. It uses students’ historical learning records to predict students’ next answer performance. Aiming at the problem that the deep knowledge tracking model based on time convolution network （TCN） only uses students’ answer sequences and answer results， and ignores other behavior characteristics of students， a deep knowledge tracking model based on multi feature extraction （TKT-PCA） is proposed. The model uses principal component analysis （PCA） method to automatically extract hidden features in a variety of students’ answer behavior and learn their representation. It not only reduces the feature dimension and redundant information， but also fully evaluates students’ knowledge mastery. The experimental results show that the TKT-PCA has the better prediction performance compared with other knowledge tracking baseline models.
Key words: deep learning； knowledge tracking； temporal convolution network； educational data mining； intelligent education

Traffic Light Control Optimization Based On D3QN

ZHANG Guo-you, SONG Shi-feng

2023, 0(07): 30-35. doi:10.3969/j.issn.1006-2475.2023.07.006

Asbtract ( 169 )

PDF (1524KB) ( 44 )

Related Articles | Metrics

Traffic lights play a vital role in controlling traffic at intersections. At present， the traffic lights at urban intersections mostly adopt the control strategy of fixed timing and fixed phase transformation， which is difficult to meet different traffic flow conditions. It has become one of the research hotspots in the field of intelligent transportation to design a control scheme that can adjust the traffic light transformation in real time according to the traffic flow at the intersection. However， the traffic flow at urban intersections is dynamic， so it is difficult to study it directly. In order to design an appropriate traffic light dynamic control scheme， the deep strong learning technology is introduced. The intersection traffic light control problem is abstracted into a reinforcement learning model， which is solved by D3QN algorithm. On this basis， considering the vehicles in different states， the state input and reward function are improved. Finally， the simulation experiments under different traffic flows are carried out on the traffic simulator SUMO. The experimental results show that after the model training becomes stable， the average queue length of the D3QN algorithm with improved reward function and state input is significantly improved compared with the traditional fixed control strategy and adaptive control strategy under three traffic flows， and the control effect is better then DQN algorithms and DDQN algorithms.

Bridge Health Monitoring Data Prediction Model Based on ICEEMDAN-BiLSTM-ARIMA Combined Model

LI Shi-jia, HOU Li-juan, TANG Bin, YANG Liu, LIU Heng,

2023, 0(07): 36-42. doi:10.3969/j.issn.1006-2475.2023.07.007

Asbtract ( 110 )

PDF (3187KB) ( 37 )

Related Articles | Metrics

Aiming at the various types of time series data collected by the current bridge structural health monitoring system， in view of the bridge structural response and the additional impact of the environment on the data， in order to achieve bridge structural safety early warning， based on the principle of integrated algorithm， this paper adopts ICEEMDAN （Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise）method studied by the improved current empirical modal decomposition method to decompose the bridge monitoring stress data， and uses the multi-scale permutation entropy algorithm to sort and reorganize the decomposed components. Finally， we combine the classical time series analysis theory BiLSTM (Bidirectional Long Short-Term Memory) network with differential ARIMA (Autoregressive Integrated Moving Average Model) to make predictive analysis for the reconstituted component and combine the results to get the final predicted value. By verifying the stress data collected by the health monitoring system of the Dadu River Bridge on the Yakang Expressway， the results show that this method effectively improves the prediction effect compared with the single model， with an overall increase of about 60%~70%. The method achieves accurate prediction of bridge monitoring data and lays a strong foundation for future bridge structure health state prediction， digital construction and safety early warning.

Textile Raw Material Cost Warning Based on Apriori Algorithm of Association Rules

ZHONG Song-ying

2023, 0(07): 43-43. doi:10.3969/j.issn.1006-2475.2023.07.008

Asbtract ( 80 )

PDF (1152KB) ( 25 )

Related Articles | Metrics

In order to solve the problems of low accuracy and success rate of existing early warning methods， this paper proposes a textile raw material cost early warning method based on the Apriori algorithm of association rules. The composition of textile raw material cost is analyzed， including raw material cost， electricity cost， salary cost， other expenses， transportation and loading and unloading cost and period cost. The content of textile raw material cost management is studied. The cost early warning index system of textile industry is established， and the gray correlation analysis method is used to accurately mine the early warning data sequence. The Apriori algorithm is used to calculate the maximum frequent item set of the warning data， and the textile raw material cost warning is completed through the calculation of the confidence results. The experimental results show that the method proposed in this paper has high accuracy， short time-consuming and high success rate of early warning.

Improved DWA Obstacle Avoidance Algorithm in Dense Obstacle Environment

DENG Yun-zheng, HUANG Yi-hu

2023, 0(07): 48-53. doi:10.3969/j.issn.1006-2475.2023.07.009

Asbtract ( 123 )

PDF (2726KB) ( 55 )

Related Articles | Metrics

Aiming at the problems of the traditional dynamic window approach（DWA） algorithm in the dense obstacle environment， such as easy to bypass the obstacle area and poor obstacle avoidance， an improved DWA algorithm based on A* is proposed. Firstly， the offset cost is introduced into the evaluation function of A* algorithm to guide the algorithm to search in the target direction quickly， so as to improve the problem of low planning efficiency， and the global optimal waypoint is obtained by optimizing the waypoint. Secondly， the weights of the evaluation function are dynamically adjusted by the orientation and distance of obstacles in the DWA algorithm to solve the problem of poor obstacle avoidance in the dense obstacle environment. Finally， the global optimal waypoint is incorporated to ensure that the improved DWA algorithm can achieve real-time obstacle avoidance and ensure the optimal path. The simulation results show that compared with the other two algorithms， the improved DWA algorithm can effectively improve the robot’s obstacle avoidance in the dense obstacle environment， the path length and the number of steps can be decreased by more than 15%， and can effectively avoid random obstacles， with higher security and stronger robustness.

Deep Cross Network Recommendation Model Based on Attention Perception

CUI Shao-guo, ZHANG Gang, WANG Ao-di

2023, 0(07): 54-60. doi:10.3969/j.issn.1006-2475.2023.07.010

Asbtract ( 96 )

PDF (2505KB) ( 31 )

Related Articles | Metrics

The combination of low-order and high-order features in the recommendation system is crucial to the predicted click-through rate. This paper designs an Attention Deep Cross IO-awre Factorization Machine （ADCIOFM） model. The traditional recommendation model extracts low-order and high-order features through the attention factor decomposer and deep cross-network respectively. However， the hidden field information is easily ignored when the attention factor decomposer extracts low-order combined features， and the diversity of user interests in deep cross-network mining is weak. Therefore， this paper enhances the representation ability of attention mechanism to estimate the low-order combined feature weight by incorporating the perceptual auxiliary matrix. The feature depth of different subspaces is extracted by integrating a multi-head attention mechanism to solve the problem of user interest diversity in deep cross-network mining. Finally， the low-order and high-order combined features are effectively fused for a recommendation. Through experimental comparison on Criteo and Movielens-100K data sets， the AUC index is used for evaluation， which is 0.0087 and 0.0159 higher than the benchmark model.

Decentralized Online Shopping Data Sharing Scheme Based on Blockchain

ZHONG Lin-feng, LI Yan-feng, ZHANG Gui-peng, LIU Wen-yin

2023, 0(07): 61-68. doi:10.3969/j.issn.1006-2475.2023.07.011

Asbtract ( 113 )

PDF (2488KB) ( 20 )

Related Articles | Metrics

To solve the problems of over-centralized data storage and low utilization efficiency in online shopping， this paper proposes a decentralized online shopping data sharing scheme based on blockchain. Firstly， SM2 elliptic curve public key cryptography algorithm and smart contract technology are used to encrypt and store online shopping data to ensure data integrity and security. Secondly， smart contract binding technology and SimHash technology are used to collect and process online shopping data to ensure data authenticity and improve data quality. Finally， the improved ZK-PoD （Zero Knowledge Proof of Delivery） protocol is used to realize the sharing of online shopping data， so as to achieve the fair exchange of data and value without the participation of trusted third parties. Theoretical analysis and experiments show that this scheme is more secure than traditional centralized online shopping data management scheme. Compared with the data sharing scheme implemented by proxy re-encryption technology， this scheme greatly reduces the time required for data sharing.

A Self-optimizing Method for Antenna Coverage of Railway Communication Base Station Based on Reinforcement Learning

ZHANG Zhi-guo

2023, 0(07): 69-72. doi:10.3969/j.issn.1006-2475.2023.07.012

Asbtract ( 71 )

PDF (1285KB) ( 20 )

Related Articles | Metrics

In LTE-R（Long Term Evolution for Railway） communication network， one major factor influencing the coverage is related to the antenna tilt angle. By adjusting antenna tilt， signal reception of user terminal within a cell can be effectively improved. To solve the coverage optimization problem of Self-Organizing Network （SON） proposed in the 3rd Generation on Partnership Project （3GPP）， a method using reinforcement learning algorithm for optimizing the title angle of base station antenna is proposed. Firstly， with the help of the electromagnetic radiation model of the base station antenna， we build a signal coverage optimization model that takes the antenna transmission gain as the objective function. Secondly， we transform the coverage optimization problem into the maximum benefit problem by employing reinforcement learning algorithm， and then accomplish the coverage self-optimization of mobile terminal by finding the best angle value in the range space of the antenna tilt angle. Finally， we have carried out simulation and field verification. The results show that the solution increased 4.11 percentage points compared to no antenna tilt optimization， and increased 3.59 percentage points compared to the method based on the particle swarm algorithm. The proposed method is suitable for the self-optimization of base station antenna tilt angle covered by LTE-R communication network.

Microservices-based Architecture Design for Civil Aircraft Industrial Software

LU Wei-qiang

2023, 0(07): 73-78. doi:10.3969/j.issn.1006-2475.2023.07.013

Asbtract ( 112 )

PDF (5983KB) ( 25 )

Related Articles | Metrics

As the core of large-scale complex equipment manufacturing， industrial software plays an important role in realizing Industry 4.0， but China’s industrial software faces the problem of stuck neck. This paper analyzes the development situation and future strategic needs of civil aircraft industrial software， introduces the Internet micro-services architecture technology， designs the civil aircraft industrial software architecture for the entire civil aircraft industry chain， and applies it to a civil aircraft manufacturing enterprise. The enterprise innovates the concept of “business + technology”， realizes business process-oriented micro-services and technology mid-stage micro-services， and adopts the lean management model to finely manage huge and complex business modules and public technology modules. The system has been in trial operation for half a year， and the success rate of continuous release and delivery version has reached 98%， and the quality factor of development is getting higher and higher. Although affected by the epidemic， the success rate of a release that passes software testing is still significantly increased by 75%.

A Remote Sensing Image Change Detection Model Based on CNN-Transformer Hybrid Structure

XU Ye-tong, GENG Xin-zhe, ZHAO Wei-qiang, ZHANG Yue, NING Hai-long, LEI Tao

2023, 0(07): 79-85. doi:10.3969/j.issn.1006-2475.2023.07.014

Asbtract ( 366 )

PDF (2633KB) ( 75 )

Related Articles | Metrics

The emergence of convolutional neural network and Transformer model has made continuous progress in remote sensing image change detection technology， but at present， these two methods still have shortcomings. On the one hand， the convolutional neural network cannot model the global information of remote sensing images due to its local perception of convolution kernel. On the other hand， although Transformer can capture the global information of remote sensing images， it cannot model the details of image changes well， and its computational complexity increases quadrally with the resolution of images. In order to solve the above problems and obtain more robust change detection results， this paper proposes a CNN-Transformer Change Detection Network （CTCD-Net） based on convolutional neural network and Transformer hybrid structure. Firstly， CTCD-Net uses convolutional neural network and Transformer based on encoding and decoding structure in series to effectively encode local and global features of remote sensing images， so as to improve the feature learning ability of the network. Secondly， the cross-channel Transformer self-attention module （CSA） and attention feedforward network （A-FFN） are proposed to effectively reduce the computational complexity of Transformer. Full experiments on LEVIR-CD and CDD datasets show that the detection accuracy of CTCD-Net is significantly better than that of other mainstream methods.

Rice Nitrogen Nutrition Diagnosis Based on HSV Color and LBP Texture Features

YANG Sun-zhe, SUN Ai-zhen

2023, 0(07): 86-92. doi:10.3969/j.issn.1006-2475.2023.07.015

Asbtract ( 69 )

PDF (2440KB) ( 23 )

Related Articles | Metrics

In order to realize diagnosis and identification of nitrogen nutrition in rice quickly and conveniently， we propose a diagnostic identification method of nitrogen nutrition in rice based on leaf HSV combined with LBP texture histogram features. The field experiment of early rice is conducted on “Zhong Jiazao” rice species， and four levels of nitrogen application are set. The images of the top three leaf tips of rice at the tillering stage are captured by the camera， and the HSV color features of each leaf tip are obtained separately by using the image processing technique， and with improving the LBP algorithm， the mean_LBP texture features are extracted by using the average gray value of the center point 5×5 range pixels instead of the gray value of the center point pixels as the threshold value. The H， S， V color histogram and mean_LBP texture histogram features are quantized， normalized， and concatenated and merged into a 1D feature vector with 1024 components， and after dimensionality reduction by PCA， GS_SVC， BP， KNN and RF methods are applied to construct a rice nitrogen nutrition diagnosis identification model， respectively. The experimental results show that the improved tip images of mean_LBP texture features combined with HSV color features have the identification accuracy rate reached 95.23% in GS_SVC that better than other models. Since the leaf tip is sensitive in the diagnosis of nitrogen nutrition in rice and image HSV， LBP features are not affected by subjective factors. It shows that the proposed method has good universality and reliability， which provides a feasible and new method for accurate nutritional diagnosis in rice and other crops.

Image Segmentation Method of Residual Film on Cotton Field Surface Based on Improved SegFormer Model#br#

NIU Yu-heng, LI Yong-ke, CHEN Yan-hong, JANG Ping-an

2023, 0(07): 93-98. doi:10.3969/j.issn.1006-2475.2023.07.016

Asbtract ( 97 )

PDF (3556KB) ( 44 )

Related Articles | Metrics

In order to solve the problem of serious pollution caused by residual plastic film during cotton planting， a fast recognition and segmentation method based on improved SegFormer model is proposed. Taking the collected surface residual film of cotton field in Changji City， Xinjiang Uygur Autonomous Region （coordinates 44 ° 23 ′ 1 ″ N， 87 ° 30 ′ 23 ″ E） as the research object， 1047 images are collected at noon on a sunny day after snow and made into a data set. Based on the SegFormer model， a deeper feature layer level is added to obtain more subtle features to solve the problem of the morphologic variation of the residual film and the smaller target. The average crossing and merging ratio of the original SegFormer model has reached 83.00%， the average crossing and merging ratio of the improved SegFormer model has increased by 0.42 percentage points compared with the original model， the die coefficient has increased by 0.3 percentage points and the single detection time is 51.13 ms. The experimental results show that the improved SegFormer model can basically meet the requirements of fast segmentation tasks， and provide a theoretical basis for rapid assessment of residual film pollution in cotton fields.

Lightweight License Plate Detection Algorithm Based on Improved YOLOv4

SHAN Yu, ZHANG Hao-peng, CHI Jing

2023, 0(07): 99-104. doi:10.3969/j.issn.1006-2475.2023.07.017

Asbtract ( 122 )

PDF (1832KB) ( 44 )

Related Articles | Metrics

To address the problem that existing license plate detection algorithms are ineffective in complex environments， a deep learning-based GEG-YOLOv4 lightweight license plate detection model is proposed. The model uses YOLOv4 as the basic framework and GhostNet as the backbone network to significantly reduce the number of model parameters， and incorporates an ECA attention module that can avoid dimensionality reduction and effectively capture cross-channel interaction information to increase the channel weight of license plate information and reduce the interference of license plate information from the complex environment background. Finally， the Ghost module is used to replace some of the normal convolution in the deep network， which further reduces the number of model parameters while better preserving the redundant information in the feature map. The experimental results on a large license plate dataset CCPD show that the GEG-YOLOv4 model reduces the number of parameters by about 88%， increases the AP value by 0.09% and improves the speed by about 55% compared with YOLOv4， which has better detection performance for license plate data in complex environments than other methods and can meet the needs of practical application scenarios.

Fine-grained Identification of Maidong Based on Multi-scale ResNet Combining Attention Mechanism

QIN Zhu-yuan, WU Hao-zhong, TAN Dai-qing, HAN Ai-qing, ZANG Hao, WANG Xuan, TANG Yan

2023, 0(07): 105-111. doi:10.3969/j.issn.1006-2475.2023.07.018

Asbtract ( 125 )

PDF (4687KB) ( 48 )

Related Articles | Metrics

The identification of traditional Chinese medicinal materials depends on the experience of Chinese pharmacists， with low efficiency and no unified quantitative criteria. Aiming at the fine granularity classification problem of Sichuan Ophiopogon japonicus， Liriope spicata and Zhejiang Ophiopogon japonicus， an improved MARNet-152（Multiscale-Attention Residual Network-152） model based on ResNet-152 neural network is proposed， which assists artificial identification of three easily-confused maidong decoction pieces automatically. An improved model， MARNet-152 is constructed based on ResNet-152 residual neural network， with group convolution of 3×3 convolutional kernels in the Bottleneck of the ResNet-152 network structure to extract and represent multi-scale features. The convolution attention mechanism module（CBAM） combining space and channel is introduced to make the model pay more attention to the recognition of target object details and have better interpretation. The classification accuracy of the improved network model reached 91.42% in the fine grained recognition of maidong image， which is 6.62 percentage points higher than that of the basic model， and could provide reference for the recognition of maidong image. The improved MARNet-152 model has higher generalization ability， and the recognition effect is significantly improved compared with the original ResNet-152 model.

Remote Sensing Image Road Segmentation Based on CA-TransUNet

GONG Xuan, GUO Zhong-hua, CHEN Wang

2023, 0(07): 112-118. doi:10.3969/j.issn.1006-2475.2023.07.019

Asbtract ( 139 )

PDF (2025KB) ( 57 )

Related Articles | Metrics

Aiming at the problems of missed and misjudgment of the road segment in optical remote sensing image with complex background and rich feature information， this paper puts forward a method of remote sensing image road segmentation based on CA-TransUNet. The semantic segmentation network TransUNet with multi-head self-attention is taken as the benchmark， and the void space pyramid pooling is integrated into the feature extraction module to obtain the feature maps of different horizons. Through the integration of the information of each channel， the extraction of multi-scale features is enhanced. A hybrid attention mechanism is added to the cascaded upsampling module to reduce the loss of process details， suppress the attention to irrelevant boundary information， and enhance the road features. The Dice loss function and binary cross-entropy loss are selected to optimize the road segment of optical remote sensing images more accurately. Experimental results show that the proposed method achieves 56.33% IoU value and 71.32% F1 index on DeepGlobe dataset， and the accuracy is up to 97.32%， which is higher than other classical road segmentation algorithms in remote sensing images. The improved algorithm can effectively segment the remote sensing images with complex surrounding background， obstructed by obstacles and narrow roads.

Improved Solar Cell Defect Detection Algorithm Based on YOLOv5s

LUO Wei, LIU Si-yuan, XU Jian-xiang, DONG Tian-pei

2023, 0(07): 119-126. doi:10.3969/j.issn.1006-2475.2023.07.020

Asbtract ( 158 )

PDF (2142KB) ( 58 )

Related Articles | Metrics

In the process of manufacturing solar cells， due to the imperfect manufacturing processes and operational failure of humans， the defects， such as broken cell， crack， finger failure and silicon material missing might be found in the solar cells. A solar cell defect detection model based on YOLOv5s， namely YOLOv5s_CG， is proposed to improve the precision of the solar cell defects detection. The algorithm introduces convolutional attention mechanism （CBAM） blocks in different positions of the backbone network and feature fusion layer. The attention mechanism of the backbone network focuses on the global information， and the attention mechanism of the feature fusion layer focuses on the local information. At the same time， it enhances the features in both spatial and channel dimensions and uses the GIOU loss function to evaluate the detection effect of the algorithm. The proposed method is tested on the open source solar cell dataset which is re-labeled by the authors. The experimental results show that the overall mean average precision （mAP） of the YOLOv5s-CG algorithm reaches 75.1%. Compared with the algorithm of YOLOv5s， various types of defect detection accuracy have been improved， among which the accuracy of crack and silicon material missing has increased by 0.036 and 0.033 respectively， and the average accuracy （mAP） of all classes has increased by 0.026. Compared with the mainstream target detection algorithm of SSD， the overall mean average precision （mAP） has improved by 0.123. The algorithm can accurately detect the defects of solar cells， which could provide a better defects detection algorithm for real solar cell production.

Table of Content