In order to more clearly reflect the output classification status of features in each layer of RESNET model, t-sne technology is used to reduce and visualize the high-dimensional features of output of maximum pooling layer, pre activated residual unit module and full connection layer of RESNET model. The visualization effect is shown in Figure 1, in which different colors represent different gear states, The number corresponds to the type label.
From Figure 1, it can be seen that after the maximum pool layer and the first residual unit module, the distribution of state signals of gears is disorderly and irregularly. At this time, the feature extraction ability of the model is weak. After the second residual unit module, the state signals of gear gradually began to gather to the same position, but it is still not clear that the different states of gears can be divided. After the third residual unit module, gear wear, gear crack and gear broken signal gradually began to separate from other two gear states. After the fourth residual module, each state signal can be gathered together. After the whole connection layer, the pitting signal of gear can be separated from other 4 gear state signals.
Through analyzing the visualization process of reducing dimension of rocker gear, it is shown that with the increasing of model layers, the nonlinear expression ability of RESNET model is gradually enhanced, and the indivisible features can be mapped to the non-linear separable space. It is shown that RESNET model has strong feature extraction ability in the application of fault classification of shearer rocker gear.