Bearing Fault Identification Method under Small Samples and Multiple Working Conditions
Algorithm 3
Our bearing fault identification algorithm.
Inputs: Number of classifiers, training and test samples composed of rolling bearing vibration signals, and the sample labels.
Outputs: Bearing fault identification classifier and the prediction values of the test samples.
(1)
Set β with 0.001 for Algorithm 2;
(2)
For all classifiers do:
(3)
Decompose the bearing vibration signals and construct the new training set, verification set, and test set according to the data preprocessing step described in Section 3.1;
(4)
Divide the training set into support and query sets with the ratio of 1 : 1;
(5)
Initialize the sample weight of the support and query set with ,;,, where and are the numbers of the sample of the support set and query set;
(6)
Set the target loss function with the cross-entropy (, where p is the predictive value, q is the true value, and is the sample weight);
(7)
Update the parameters of the first classifier using the support set and query set according to the modified Meta-SGD;
(8)
Calculate the identification error rate of the training set according to equation (2);
(9)
Calculate the weight coefficient of the classifier according to equation (3);
(10)
Update the sample weight of the training sample, and normalize these weights according to (4);
(11)
Use the network parameters of the previous classifier to initialize the network parameters of the next classifier;
(12)
Train the next classifier using the training set updated with the new weights according to the modified Meta-SGD;
(13)
end
(14)
Calculate the prediction values of the test sample according to equation (5), output the prediction values and the integrated classifier.
