Model for recognizing the wear condition of fir-tree slot broaching tools based on MobileNet v2.5-StackedBiGRU

Broaching is a critical process in the production of turbine disk fir-tree slots for aeroengines. Any abnormal state of the broaching tool, if not addressed in time, can result in serious damage. A wear state recognition model (MobileNet v2.5-StackedBiGRU) of broaching tool is developed to enhance the accuracy and generalization performance of tool wear monitoring and prediction. In this model, the MobileNet v2.5 neural network is designed to effectively extract the spatial feature information from vibration signals and realize feature dimensionality reduction. Then, a bidirectional stacked gate recurrent unit (StackedBiGRU) neural network is designed to achieve temporal feature extraction. Finally, the extracted features are fed into multiple fully connected layers and softmax layers to achieve the recognition of broaching tool wear state. By experimentally verifying the real-time signals collected during the fir-tree slot broaching process, the MobileNet v2.5-StackedBiGRU model, which is based on the merging of spatial feature information and temporal feature information, achieves an accuracy rate of 98.23% in classification accuracy recognition. And the single recognition speed is 34.13% and 90.53% faster than the MobileNet v2.5 and StackedBiGRU models, respectively.