Research on Classification and Recognition of Micro Milling Tool Wear Based on Improved DenseNet

Due to the small size of micro milling tools and wear sizes even in the micrometer range, monitoring tool wear is extremely challenging. The wear state of cutting tools directly affects machining quality and productivity, becoming a key factor restricting the development of micro milling tools. This article proposes and establishes a method for identifying and monitoring the wear status of micro milling tools based on machine vision and deep learning models. The deep learning model improves the DenseNet-121 network by using deformable convolution, depthwise separable convolution, and partial convolution instead of standard convolution, reducing the number of model parameters and computational complexity. We have developed a new DLDenseBlock and feature transfer method. In order to better utilize shallow features and reduce the risk of feature loss, the output of DLDenseBlock is processed by EMA attention mechanism and input into the feature pyramid for feature fusion. Four types of micro milling tool wear image datasets were created based on micro milling tool milling and wear evaluation criteria. Using this dataset, existing network models and the improved model proposed in this paper were trained and tested. The experimental results show that the proposed wear state evaluation criteria are reasonable and accurate. Compared with existing network models, the improved model in this paper can effectively identify the wear state of micro milling tools with an accuracy rate of up to 100%.