Introduction
-
This paper proposes a dual-channel CNN (DC-CNN) model to extract the characteristics of power equipment through two independent CNN models.
-
This paper proposes a random forest(RF) classification method incorporating deep learning for defect recognition of power equipment
-
The proposed DC-CNN and RF classification method is used to classify images of various types of power equipment, and the validity of the proposed method is verified.
Feature extraction algorithm
AlexNet model
DC-CNN model
The extraction of substation features
FP algorithm
BP algorithm
Image recognition of power equipment
Analysis of experimental results of feature extraction
Analysis of experimental results of substation image recognition
Model | Data | Insulator | Transformer | Breaker | Bus | Tower |
---|---|---|---|---|---|---|
Single CNN | Training | 94% | 90% | 91% | 92% | 74% |
Validation | 94% | 91% | 90% | 93% | 73% | |
Testing | 94% | 89% | 89% | 92% | 72% | |
Dual CNN | Training | 97% | 93% | 94% | 96% | 82% |
Validation | 97% | 93% | 95% | 96% | 81% | |
Testing | 97% | 92% | 93% | 96% | 80% |
Model | Average accuracy (%) | GPU running average time (s) | CPU running average time (s) |
---|---|---|---|
Single CNN | 87.2 | 0.9 | 330 |
Dual CNN | 91.6 | 1.2 | 452 |
Results analysis of different recognition methods
Method | Types | Insulator | Transformer | Breaker | Bus | Tower |
---|---|---|---|---|---|---|
1 | A | 90% | 84% | 85% | 83% | 74% |
B | 91% | 86% | 85% | 85% | 76% | |
2 | A | 96% | 91% | 92% | 95% | 79% |
B | 97% | 92% | 93% | 96% | 80% | |
3 | A | 78% | 72% | 71% | 70% | 80% |
B | 80% | 73% | 72% | 70% | 81% |