Author Affiliations
College of Physics and Information Engineering, Fuzhou University, Fuzhou, Fujian 350108, Chinashow less
Fig. 1. Foam images in different dosing states. (a) s1; (b) s2; (c) s3; (d) s4
Fig. 2. Flow chart of the NSST-CNN model
Fig. 3. Structure of the RAE-KELM. (a) Self-encoder; (b) RAE-KELM
Fig. 4. Flow chart of the parameter optimization
Fig. 5. Overall framework of our method
Fig. 6. Performances of different optimization algorithms
Fig. 7. Visualization results of CNN features. (a) MNIST handwritten data set; (b) flotation foam image in state s1; (c) flotation foam image in state s2
Fig. 8. Testing accuracies of different feature extraction methods. (a) MNIST handwritten data set; (b) flotation foam image
Fig. 9. Recognition result of flotation dosing states
Fig. 10. Part of the recognized wrong image. (a) s1 is misjudged as s2; (b) s3 is misjudged as s2; (c) s4 is misjudged as s3
Layer | Scale 1 | Scale 2 | Scale N(N=3,4,5,…) |
---|
Step size | Number of cores | Core size | Step size | Number of cores | Core size | Step size | Number of cores | Core size |
---|
C1 | 1 | 64 | 5×5 | 1 | 64 | 5×5 | 1 | 64 | 5×5 | P2 | 2 | 64 | 2×2 | 2 | 64 | 2×2 | 2 | 64 | 2×2 | C3 | 1 | 32 | 3×3 | 1 | 64 | 3×3 | 1 | 64 | 3×3 | P4 | 2 | 32 | 2×2 | 2 | 64 | 2×2 | 2 | 64 | 2×2 | C5 | 1 | 32 | 2×2 | 1 | 32 | 2×2 | 1 | 64 | 2×2 | P6 | 2 | 32 | 2×2 | 2 | 32 | 2×2 | 2 | 64 | 2×2 | Fully connected layer | 1024 neurons |
|
Table 1. Parameters of the CNN model
Δθ/π | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 |
---|
PO /% | MO | PO /% | MO | PO /% | MO | PO /% | MO | PO /% | MO |
---|
1.1 | 97.47 | 213.4 | 97.91 | 205.1 | 98.56 | 211.1 | 99.11 | 196.3 | 98.79 | 215.6 | 1.2 | 98.21 | 196.5 | 97.88 | 208.8 | 99.24 | 204.4 | 99.20 | 204.8 | 99.06 | 204.6 | 1.3 | 97.91 | 201.6 | 98.24 | 199.4 | 98.78 | 199.7 | 99.15 | 201.6 | 99.35 | 198.4 | 1.4 | 98.33 | 208.1 | 98.56 | 197.6 | 98.95 | 196.4 | 99.64 | 198.7 | 99.64 | 202.7 | 1.5 | 99.11 | 204.3 | 99.55 | 204.5 | 99.35 | 196.8 | 99.70 | 196.9 | 99.81 | 205.3 | 1.6 | 98.76 | 197.4 | 99.36 | 210.4 | 99.12 | 201.3 | 99.63 | 206.2 | 99.65 | 203.8 | 1.7 | 99.23 | 196.6 | 99.87 | 201.8 | 99.76 | 195.2 | 99.95 | 190.8 | 99.75 | 206.2 | 1.8 | 99.05 | 192.1 | 99.78 | 195.3 | 99.23 | 198.8 | 99.63 | 194.3 | 99.54 | 197.5 | 1.9 | 98.86 | 195.8 | 99.36 | 199.2 | 99.34 | 205.6 | 99.24 | 201.8 | 99.21 | 198.0 | 2.0 | 98.25 | 202.0 | 98.95 | 206.4 | 98.97 | 207.1 | 99.26 | 202.7 | 98.99 | 201.3 |
|
Table 2. QBFA's average optimal value probability and number of iterations under different parameters
Combination pair | QGA+KELM | QGA+RAE-KELM | QBFA+KELM | QBFA+RAE-KELM |
---|
Training accuracy /% | 97.47 | 99.55 | 98.98 | 99.88 | Training time /s | 5659.55 | 5802.57 | 6050.36 | 6132.32 | Testing accuracy /% | 94.58 | 96.77 | 96.64 | 98.89 | Testing time /s | 5.04 | 5.08 | 5.18 | 5.11 |
|
Table 3. Recognition results of flotation dosing states by different combination pairs
Method | Ref. [2] | Ref. [3] | Ref. [5] | Ref. [7] | Ours |
---|
Training accuracy /% | 89.67 | 86.21 | 98.38 | 98.89 | 99.88 | Training time /s | 1598.41 | 1444.22 | 14365.21 | 6079.52 | 6132.32 | Testing accuracy /% | 82.22 | 76.35 | 93.25 | 95.78 | 98.89 | Testing time /s | 4.47 | 4.25 | 6.43 | 5.32 | 5.11 |
|
Table 4. Recognition results of different methods for flotation dosing states