Author Affiliations
1School of Electrical Engineering, Guizhou University, Guizhou, Guiyang 550025, China2Department of Brewing Engineering Automation, Moutai College, Zunyi, Guizhou 564507, Chinashow less
Fig. 1. Structure of the B-CNN
Fig. 2. Process of the gradient calculation
Fig. 3. Basic structure of the SENet
Fig. 4. Flow chart of the SENet
Fig. 5. Improved network structure
Fig. 6. Structure of the residual unit
Fig. 7. Images of press-plate in different states. (a) guan; (b) kai; (c) NS1; (d) NS2
Fig. 8. Flow chart of network training
Fig. 9. Confusion matrix of different methods. (a) Our method; (b) B-CNN
Fig. 10. Grad-CAM diagrams with different opening and closing angles of the press-plate
Fig. 11. Recognition results of different methods. (a) NS1; (b) NS2; (c) guan; (d) kai
Fig. 12. Accuracies of different methods
Fig. 13. Loss rates of different methods
Fig. 14. Accuracies of different methods in the test set
Name | Configuration |
---|
CPU | Intel i5-5200U 2.20 GHz | RAM | 8 GB | GPU | NVIDIA GeForce 920 M 4.0 G | GPU acceleration library | CUDA 9.0 cuDNN v7.1 | Deep learning framework | Pytorch1.1.0 |
|
Table 1. Parameter of the experimental platform
Experimental parameter | Scene setting |
---|
Activation function | ReLU | Number of samples | 2672 | Learning initial speed | 10-4 | Rate decay coefficient | 0.01 | Input data dimension | 448×448×3 | Loss function | cross entropy | Number of iterations | 3500 | Optimization algorithm | Adam | Activation function | Softmax | Output data dimension | 3 |
|
Table 2. Initial parameters of the experiment
Parameter | Actual value |
---|
Predictive output | P' | XTP | XFP | N' | XFN | XTN | Total | P | N |
|
Table 3. Confusion matrix
Method | PPV | TPR |
---|
B-CNN | 0.92 | 0.91 | B-Se-ResNet | 0.98 | 0.94 | Hog+SVM | 0.88 | 0.87 | B-ResNet | 0.94 | 0.92 | DCL | 0.95 | 0.94 | CIN | 0.96 | 0.91 |
|
Table 4. Evaluation indicators of different methods