1. Selection of residual unit module
In RESNET, there are two kinds of residual unit modules: original residual unit module and pre activated residual unit module. Residual unit module is mainly composed of weight layer, batch normalization (BN) layer and relu. The weight layer is used to extract features, the BN layer is used to speed up the network training, and the relu is used to reduce the interdependence between parameters, improve the nonlinear modeling ability of the network, and prevent the over fitting phenomenon. Addition stands for summation.
In the original residual unit module, the first layer convolutes the vibration data directly, which increases the computational complexity of network training and slows down the speed of training network. The pre activated residual element module adopts the pre activated connection mode, which reduces the complexity of the model and makes the model converge faster. Therefore, the pre activated residual element module is selected in this paper.
2. RESNET structure
RESNET mainly includes convolution layer, pooling layer and classification layer.
2.1 Convolution layer
In the convolution layer, multiple convolution kernels are used to extract the main features of the original vibration signal, and multiple feature vectors are obtained to achieve feature dimensionality reduction. The calculation formula of convolution layer is as follows:
Where xmj is the output of the m-th convolution layer and j is the dimension; N is the number of input eigenvectors; Xm-1i is the input of the m-th convolution layer; Kmij is the convolution kernel; BMJ is bias.
The convolution layer contains three parameters: convolution kernel size, moving step size and convolution kernel number. RESNET has nine layers. In order to extract short-term features, the first convolution layer uses the wide convolution kernel, which can adaptively learn the features for fault diagnosis and improve the classification accuracy. The convolution kernel size of this layer is 64 × The number of convolution kernels is 32. The rest of the convolution layers are included in the residual unit module.
2.2 Pooling layer
Pooling layer can reduce the output sample size of the middle layer, reduce the number of training parameters of the following layers, and speed up the training speed of the model, so that the model can more deeply mine the useful information of the data in the training. Therefore, a maximum pooling layer is added after the first convolution layer of RESNET, and an average pooling layer is added after the last residual unit module. The size of pooling layer is 5 × The moving steps are 5.
2.3 Classification layer
The classification layer is mainly composed of full join layer and softmax regression classifier. The full join layer flattens the output features of the pooling layer, that is, all the feature vectors are connected to form a one-dimensional vector. Softmax regression classifier mainly realizes target fault classification.
3. Fault diagnosis algorithm based on RESNET
The fault diagnosis algorithm based on RESNET consists of data reorganization, model training, model verification and model testing. Firstly, the data of the original sample is recombined, then the recombined sub sample is divided into training sub sample and test sub sample according to the ratio of 8:2, and finally the training sub sample is divided into actual training sub sample and verification sub sample according to the ratio of 4:1.
Data reorganization: the main purpose is to expand the original one-dimensional vibration signal samples. The original sample is divided into several different sub samples in the way of 1 / 20 overlap as the input of RESNET model.
Model training: 80% of the training sub samples are used as the actual training sub samples for model training. In the iteration process, the error between the actual output and the expected output is compared. When the error is greater than the given threshold, the weight of the training model is updated reversely; When the error is less than a given threshold, the training model is output.
Model validation: 20% of the training sub samples are used as validation sub samples for model validation. In each iteration, the algorithm will update the weight value of each layer in the network connection. After each iteration, the algorithm will use the validation sub samples to test the improvement effect of the model and better evaluate the advantages and disadvantages of the model.
Model testing: 20% of the sub samples are used as test sub samples for model testing. The test sub samples are directly used as the input of the training model, and the fault classification is realized by softmax regression classifier.