Multibranch Block-Based Grain Size Classification Of Hybrid Disk Using Ultrasonic Scattering: A Deep Learning Method

Abstract

To assess the grain size of hybrid disks, we propose a simple network architecture—the wide-paralleled convolutional neural network (WP-CNN)—based solely on multibranch blocks and create a grain size classification model based on it. Multibranch blocks are used to enhance the capability of feature extraction, and the global average pooling layer was implemented to reduce the number of model parameters. To train and test the model, a dataset of ultrasonic scattering signals from a hybrid disk was constructed. The WP-CNN structure and hyperparameter selection were examined using the training set. The experiment demonstrated that, compared to traditional 1D convolutional neural network, 1D ResNet, and InceptionTime, the classification accuracy of this method can reach 92.3%. A comparison is made with the empirical mode decomposition scattering model and frequency spectra tree model. The proposed network provides accurate classification of grain size without physical parameters and specific physical models. The results show the deep learning method has the feasibility to evaluate hybrid disk grain size distribution.