ShuffleNet v2.3-StackedBiLSTM-Based Tool Wear Recognition Model for Turbine Disc Fir-Tree Slot Broaching

Abstract

At present, deep learning technology shows great market potential in broaching tool wear state recognition based on vibration signals. However, traditional single neural network structure is difficult to extract a variety of different features simultaneously and has low robustness, so the accuracy of wear status recognition is not high. In view of the above problems, a broaching tool wear recognition model based on ShuffleNet v2.3-StackedBiLSTM is proposed in this paper. The model integrates ShuffleNet v2.3, which has been channel shuffling, and StackedBiLSTM, a long and short-term memory network, to effectively extract spatial and temporal features for tool wear state recognition. Based on the innovative recognition model, the turbine disc fir-tree slot broaching experiment is designed, and the performance index system based on confusion matrix is adopted. The experimental research and results show that the model has outstanding accuracy, precision, recall, and F1 value, and the accuracy rate reaches 99.37%, which is significantly better than ShuffleNet v2.3 and StackedBiLSTM models. The recognition speed of a single sample was improved to 8.67 ms, which is 90.32% less than that of the StackedBiLSTM model.

Keywords

• deep learning
• vibration signal characteristics
• broaching
• channel shuffling
• tool state recognition