ResnetCPS for Power Equipment and Defect Detection
Routine visual inspection is fundamental to the preventive maintenance of power equipment. Convolutional neural networks (CNNs) substantially reduce the number of parameters and efficiently extract image features for classification tasks. In the actual production and operation process of substations...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2024-11-01
|
| Series: | Applied Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3417/14/22/10578 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1846154508653035520 |
|---|---|
| author | Xingyu Yan Lixin Jia Xiao Liao Wei Cui Shuangsi Xue Dapeng Yan Hui Cao |
| author_facet | Xingyu Yan Lixin Jia Xiao Liao Wei Cui Shuangsi Xue Dapeng Yan Hui Cao |
| author_sort | Xingyu Yan |
| collection | DOAJ |
| description | Routine visual inspection is fundamental to the preventive maintenance of power equipment. Convolutional neural networks (CNNs) substantially reduce the number of parameters and efficiently extract image features for classification tasks. In the actual production and operation process of substations, due to the limitation of safety distance, camera monitoring, inspection robots, etc., cannot be very close to the target. The operational environment of power equipment leads to scale variations in the main target and thus compromises the performance of conventional models. To address the challenges posed by scale fluctuations in power equipment image datasets, while adhering to the requirements for model efficiency and enhanced inter-channel communication, this paper proposed the ResNet Cross-Layer Parameter Sharing (ResNetCPS) framework. The core idea is that the network output should remain consistent for the same object at different scales. The proposed framework facilitates weight sharing across different layers within the convolutional network, establishing connections between pertinent channels across layers and leveraging the scale invariance inherent in image datasets. Additionally, for substation image processing mainly based on edge devices, smaller models must be used to reduce the expenditure of computing power. The Cross-Layer Parameter Sharing framework not only reduces the overall number of model parameters but also decreases training time. To further enhance the representation of critical features while suppressing less important or redundant ones, an Inserting and Adjacency Attention (IAA) module is designed. This mechanism improves the model’s overall performance by dynamically adjusting the importance of different channels. Experimental results demonstrate that the proposed method significantly enhances network efficiency, reduces the total parameter storage space, and improves training efficiency without sacrificing accuracy. Specifically, models incorporating the Cross-Layer Parameter Sharing module achieved a reduction in the number of parameters and model size by 10% to 30% compared to the baseline models. |
| format | Article |
| id | doaj-art-ef3a720635be47e987bbe81e9704b090 |
| institution | Kabale University |
| issn | 2076-3417 |
| language | English |
| publishDate | 2024-11-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Applied Sciences |
| spelling | doaj-art-ef3a720635be47e987bbe81e9704b0902024-11-26T17:49:19ZengMDPI AGApplied Sciences2076-34172024-11-0114221057810.3390/app142210578ResnetCPS for Power Equipment and Defect DetectionXingyu Yan0Lixin Jia1Xiao Liao2Wei Cui3Shuangsi Xue4Dapeng Yan5Hui Cao6The Shaanxi Key Laboratory of Smart Grid, Xi’an Jiaotong University, Xi’an 710049, ChinaThe Shaanxi Key Laboratory of Smart Grid, Xi’an Jiaotong University, Xi’an 710049, ChinaState Grid Information & Telecommunication Group Co., Ltd., Beijing 100031, ChinaState Grid Information & Telecommunication Group Co., Ltd., Beijing 100031, ChinaThe Shaanxi Key Laboratory of Smart Grid, Xi’an Jiaotong University, Xi’an 710049, ChinaThe Shaanxi Key Laboratory of Smart Grid, Xi’an Jiaotong University, Xi’an 710049, ChinaThe Shaanxi Key Laboratory of Smart Grid, Xi’an Jiaotong University, Xi’an 710049, ChinaRoutine visual inspection is fundamental to the preventive maintenance of power equipment. Convolutional neural networks (CNNs) substantially reduce the number of parameters and efficiently extract image features for classification tasks. In the actual production and operation process of substations, due to the limitation of safety distance, camera monitoring, inspection robots, etc., cannot be very close to the target. The operational environment of power equipment leads to scale variations in the main target and thus compromises the performance of conventional models. To address the challenges posed by scale fluctuations in power equipment image datasets, while adhering to the requirements for model efficiency and enhanced inter-channel communication, this paper proposed the ResNet Cross-Layer Parameter Sharing (ResNetCPS) framework. The core idea is that the network output should remain consistent for the same object at different scales. The proposed framework facilitates weight sharing across different layers within the convolutional network, establishing connections between pertinent channels across layers and leveraging the scale invariance inherent in image datasets. Additionally, for substation image processing mainly based on edge devices, smaller models must be used to reduce the expenditure of computing power. The Cross-Layer Parameter Sharing framework not only reduces the overall number of model parameters but also decreases training time. To further enhance the representation of critical features while suppressing less important or redundant ones, an Inserting and Adjacency Attention (IAA) module is designed. This mechanism improves the model’s overall performance by dynamically adjusting the importance of different channels. Experimental results demonstrate that the proposed method significantly enhances network efficiency, reduces the total parameter storage space, and improves training efficiency without sacrificing accuracy. Specifically, models incorporating the Cross-Layer Parameter Sharing module achieved a reduction in the number of parameters and model size by 10% to 30% compared to the baseline models.https://www.mdpi.com/2076-3417/14/22/10578power systemsubstationResNetcross-layer parameter sharing |
| spellingShingle | Xingyu Yan Lixin Jia Xiao Liao Wei Cui Shuangsi Xue Dapeng Yan Hui Cao ResnetCPS for Power Equipment and Defect Detection Applied Sciences power system substation ResNet cross-layer parameter sharing |
| title | ResnetCPS for Power Equipment and Defect Detection |
| title_full | ResnetCPS for Power Equipment and Defect Detection |
| title_fullStr | ResnetCPS for Power Equipment and Defect Detection |
| title_full_unstemmed | ResnetCPS for Power Equipment and Defect Detection |
| title_short | ResnetCPS for Power Equipment and Defect Detection |
| title_sort | resnetcps for power equipment and defect detection |
| topic | power system substation ResNet cross-layer parameter sharing |
| url | https://www.mdpi.com/2076-3417/14/22/10578 |
| work_keys_str_mv | AT xingyuyan resnetcpsforpowerequipmentanddefectdetection AT lixinjia resnetcpsforpowerequipmentanddefectdetection AT xiaoliao resnetcpsforpowerequipmentanddefectdetection AT weicui resnetcpsforpowerequipmentanddefectdetection AT shuangsixue resnetcpsforpowerequipmentanddefectdetection AT dapengyan resnetcpsforpowerequipmentanddefectdetection AT huicao resnetcpsforpowerequipmentanddefectdetection |