Analysis of experimental results of shearer rocker gear fault

The vibration data of each state of gear under 50% load is selected for experiment, and the sample length is set as 40000. Before training, the one-dimensional original vibration signals in five states are divided into 2455 total samples in the way of 1 / 20 overlap, and then 2455 samples are divided into 1964 training sub samples and 491 test sub samples in the proportion of 8:2, Finally, the 1964 training sub samples are divided into 1571 actual training sub samples and 393 verification sub samples according to the ratio of 8:2.

The keras framework of Python software is used to build RESNET model. The activation function of hidden layer is relu, the loss function is cross entropy loss function, and the optimizer and classifier are Adam and softmax respectively. The specific training parameters are shown in Table 3. Dropout is a regularization method for neural network model, which randomly ignores some neurons in the training process to weaken the joint adaptability between nodes and enhance the generalization ability of the model. It is verified that the best effect is when the dropout probability is 0.5. Epochs represents the process that a complete data set passes through the neural network once and returns once.

After 100 rounds of training, the accuracy of model training set and verification set is shown in Figure 1. It can be seen from Figure 1 that in the first 50 rounds of epichs, the accuracy of the verification set converges faster than that of the training set. In the last 50 rounds of epichs, the accuracy of verification set and training set is close to 1.0. This shows that the autonomous learning ability of the RESNET model in the training set is gradually enhanced with the increase of epochs rounds.

In order to more clearly reflect the classification status of rocker gear in different states, the confusion matrix of the test set is visually analyzed, as shown in Figure 2. Among them, the diagonal value is the number of test samples whose gear states are correctly predicted. It can be seen from Figure 2 that in the gear wear (type label 1) test samples, one sample was mistakenly measured as gear crack (type label 3), and one sample was mistakenly measured as gear broken (type label 4); In the test sample of gear pitting (type label is 2), one sample was mistakenly measured as normal gear (type label is 0), and one sample was mistakenly measured as gear crack. The recognition rate of the other three gear states is 100%. This shows that RESNET model can well realize fault classification of shearer rocker gear.