Tool Wear State Recognition Based on DDBF Network for Industrial IoT

Tool wear monitoring has been increasingly important in intelligent manufacturing to increase machining efficiency in Industrial Internet of Things (IIoT). Existing tool wear monitoring methods fail to leverage the complementarity of features across different domains (time, frequency, and timefrequency), resulting in insufficient representation of the multiscale characteristics of non-stationary signals. To solve this problem, a Dual-Branch Dilated Convolution (DC)-Bidirectional Temporal Convolutional Network (BiTCN) Fusion (DDBF-Net) is introduced. The proposed dual-branch architecture consists of a traditional feature branch and a time-frequency feature branch. The first branch computes traditional time-domain and frequency-domain features using statistical metrics and Fast Fourier Transform, which are then processed by convolutional neural network to capture underlying patterns. The other branch applies wavelet packet decomposition to construct energy-based time-frequency matrices, which DC and BiTCN further process to capture multi-scale time-frequency patterns and model both short-term and long-term temporal dependencies. Finally, the features extracted from both branches are fused using a enhanced cross-attention feature fusion mechanism. Experimental results on the PHM 2010 dataset show that DDBF-Net improves classification accuracy by 10% compared to the baselines.