Channel-attention-based TCN-Transformer for recognition of rough handling in parcels

Abstract Manual sorting of express packages often leads to instances of rough handling of express parcels, resulting in package damage, customer complaints, and excessive packaging by businesses. To address this issue, we propose TCN-CAMTNet (temporal convolutional networks-channel attention mechanism-multi-head Transformer), a novel detection model for recognizing rough handling behaviours based on multimodal sensor data fusion. The model introduces a channel attention mechanism (CAM) and an improved temporal convolutional network (TCN) to enhance feature extraction. The core temporal convolutional block dynamically assigns weights to multi-sensor data channels using the proposed CAM, while residual connections effectively mitigate the vanishing gradient problem. These weights are integrated with the original input features along the channel dimension, significantly improving the extraction of complex multimodal information. Additionally, a multi-head attention Transformer network with positional encoding is employed to capture temporal dependencies and enhance feature representation for long-sequence data. During training, the model employs Focal Loss and the Adam optimizer, with gradient clipping to prevent gradient explosion. Experimental evaluations using laboratory-collected data demonstrate that TCN-CAMTNet achieves superior accuracy, precision, and recall. Compared to a CNN-based model, it improves accuracy by 4.86% on average, with a standard deviation decrease of 0.07. Furthermore, our dataset evaluations confirm the model’s effectiveness in detecting rough handling behaviours in logistics operations. The TCN-CAMTNet model outperforms state-of-the-art methods across multiple performance metrics, providing a novel, efficient, and reliable solution for detecting abnormal parcel handling in the courier sorting process.