Fault Diagnosis of Gearbox in Multiple Conditions Based on Fine-Grained Classification CNN Algorithm
Algorithm 1
The training algorithm of the proposed method.
Input: fault signal training set , hyperparameter ,, and , the maximum number of iterations , and counter ,
Output: parameter of the CNN model
(1)
Select a mini-batch sample from the fault signal training set
(2)
Execute the forward propagation of the CNN model, for each input signal; calculate the activation values of each layer
(3)
Compute the softmax loss error flows of the fc7; then calculate the backpropagated error flows from fc7 to fc6 and the error flows from fc7 to fc5
(4)
Compute the softmax loss of the fc6
(5)
Compute the overall error flows of fc6, which consist of the softmax loss of fc7 and softmax loss of itself. Then use the BP algorithm to calculate the error flows backpropagated from fc6 to fc5
(6)
Compute the GLM loss error flows of the backpropagation to fc5, and then multiply hyperparameter
(7)
Compute the overall error flows of fc5 which consist of fc6, fc7, and the GLM loss
(8)
Execute the backpropagation from fc7 to conv1 layer, and use the BP algorithm to compute the error flows of these layers
(9)
Based on the activation values and error flow values, calculate through BP algorithm
