Environment Understanding Algorithm for Substation Inspection Robot Based on Improved DeepLab V3.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Substation inspection traditionally relies on manual methods, posing challenges for all-weather, real-time, and accurate data collection.
• Manual inspections increase operational risks and the likelihood of safety incidents for maintenance personnel.
• There is a critical need to enhance the intelligence of inspection robots for improved substation monitoring.

Purpose of the Study:

• To develop an advanced environment understanding algorithm for substation inspection robots.
• To improve the segmentation accuracy of object edges within substation environments.
• To create a compressed, efficient model suitable for embedded platforms with limited computational power.

Main Methods:

• An improved DeepLab V3+ neural network architecture was developed.
• The Atrous Spatial Pyramid Pooling (ASPP) module was modified with a new dilate rate combination.
• A Convolutional Block Attention Module (CBAM) was integrated into the up-sampling layers.
• The improved neural network underwent compression for deployment on embedded systems.

Main Results:

• The enhanced DeepLab V3+ achieved a mean intersection-over-union (mIoU) of 57.65% on a substation dataset, an improvement of 6.39% over the original DeepLab V3+.
• The model size was reduced to 13.9 M, a decrease of 147.1 M compared to the original model.
• Experiments were conducted on both the PASCAL VOC 2012 and a dedicated substation dataset.

Conclusions:

• The proposed improved DeepLab V3+ algorithm significantly enhances environmental understanding for substation inspection robots.
• The modifications lead to superior segmentation accuracy, particularly for object edges.
• The model's compression makes it viable for deployment on resource-constrained embedded platforms, advancing robot intelligence in substations.