Page 92 - 摩擦学学报2025年第8期

1190 摩擦学学报（中英文） 第 45 卷

表 13 结果输出部分网络层参数
Table 13 Results output some network layer parameters
Network layer Parameter setting Activate
Deep series 5 4 inputs 14(S)×14(S)×512(C)
Maximum pooling layer Stride [2, 2], fill [0, 0, 0] 5(S)×5(S)×512(C)
Relu − 5(S)×5(S)×512(C)
Batch normalized layers − 5(S)×5(S)×512(C)
Two-dimensional global mean pooling layer − 1(S)×1(S)×512(C)
Dropout layer 40% discard 1(S)×1(S)×512(C)
Fully connected layer − 1(S)×1(S)×1(C)
Regression output layer Mean-squared-error 1(S)×1(S)×1(C)

表 14 IncepRegCNN神经网络训练参数
Table 14 Training parameters of IncepRegCNN
neural network
Parameters Related Settings
1.3% 1.8% 2.2% 2.9%
Solver Adam
Maximum iterations 50
Small lot size 50
Initial learning rate 0.001
Learning rate reduction factor 0.010
2.9% 3.2% 3.6% 3.8%
Learning rate reduction stage 5
Solution environment GPU
Model verification mode 50% cross verification

表 15 模型评价指标对比 3.8% 4.0% 4.1% 4.3%
Table 15 Comparison of model evaluation indexes

Neural network model RMSE MAE R 2
BP regression neural network 8.6E-3 7.3E-3 0.81
PSO-BP regression neural network 8.0E-3 6.1E-3 0.83
4.4% 4.6% 4.8% 5.1%
IncepRegCNN regression neural network 3.9E-3 3.0E-3 0.95

b. 经主成分分析方法处理后，损伤信号15种特征
值降低至7种主成分且累计贡献率为99.19%，主成分分 5.2% 5.8% 7.1% 7.4%
析方法具有良好的信号降维效果.
Fig. 15 Sample test set
c. 基于CNN网络Inception架构构建的回归神经网 图 15 测试集样本

8

7 6
Percentage loss of cross- sectional area/% 5 4 3

2
Actual value
Predicted value
1
0 2 4 6 8 10 12 14 16 18 20
Test set sample number
Fig. 16 Results of IncepRegCNN test set identification
图 16 IncepRegCNN测试集识别结果