Grouping sparse filtering: a novel down-sampling approach toward rotating machine intelligent diagnosis in 1D-convolutional neural networks

Abstract

Convolutional neural networks (CNNs) have weight-sharing and feature-learning abilities, and can efficiently and effectively be used for the health monitoring of industrial equipment. However, the pooling operation in a typical CNN can cause the loss of valuable impulse features during data down-sampling. We propose grouping sparse filtering (GSF) to overcome this problem. Instead of using a pooling operation, the GSF splits the channels of features obtained after convolution into equal-length groups. A feature selector with a feature aggregation function based on the channel importance factors and a lasso constraint is used to filter the groups to perform down-sampling. The GSF method preserves the impulse features due to the block sparsity of the vibration signal. Theoretical analysis demonstrates that the GSF has a similar computational complexity to using a pooling layer in a CNN for the same number of layers. Two experimental studies were conducted using data from a laboratory test and industrial environments. The experimental results show that the 1D-CNN with GSF provides better performance for retaining the impulse features of the rotating machinery signals and higher fault identification accuracy than a CNN with a pooling layer.