Intelligent recognition of rough handling of express parcels based on CNN-GRU with the channel attention mechanism

The rough handling of express parcels increases the risk of damage to goods, brings customer complaints, and causes over-packing problems. The prerequisite for solving the rough handling of express parcels is to identify various typical rough handling intelligently. Therefore, an intelligent recognition method based on the CNN-GRU (Convolutional Neural Networks-Gated Recurrent Units) fusion model with the channel attention mechanism is proposed in this paper. First, the collected triaxial acceleration data of the parcel are intercepted and windowed. Then seven traditional features (mean, variance, kurtosis, skewness, dynamic range, short-term energy, and zero-crossing rate) are extracted in the window. The traditional feature data is arranged in a matrix of 3 axes x 50 time windows x 7 features and normalized. Finally, the three-dimensional traditional feature matrix is input into the model to obtain the recognition results (normal, dropping, throwing, or kicking). A novel channel attention mechanism called CDCE (Channel Dense-Concatenation-Excitation) block is introduced into the CNN-GRU fusion model. Based on the Squeeze-Excitation Net, the CDCE block replaces the global pooling operation with the dense connection operation of sub-channels, and appropriately adjusts the subsequent layers, to achieve more precise parameter learning. Besides, a new data set has been collected and shared. Experiments show that the recognition accuracy of the CNN-GRU model with the CDCE blocks can reach 96.04%, which is about 1.37% higher than that of the CNN model in the previous study. Moreover, the size of the CNN-GRU model with the CDCE blocks is reduced to 7% of the size of the CNN model.