An Intelligent Tool Wear Prediction Model Based on an Improved Regressive Siamese Neural Network with Various Sample Sizes

Tool condition has a great impact on the productivity and operating costs of enterprises. Many data-driven models based on the deep learning or transfer learning framework have already achieved significant results in tool condition monitoring. Nevertheless, the performance of these models typically varies with sample size, leading to poor application extensiveness across different-sized enterprises. This paper proposes a novel intelligent model for tool wear prediction, called Regressive Siamese Neural Network with Multilayer Perceptron-Support Vector Regression (RSNN-MLP-SVR). First, the Stockwell transform is utilized to extract more comprehensive time-frequency information from raw signals, and the absolute difference between tool wear values of paired samples is calculated as the corresponding label. Then, the backbone sub-network in the regressive Siamese neural network (RSNN) sub-model stacks multiple residual network modules to capture tool-wear-sensitive features from input paired samples. Subsequently, a new similarity measurement method based on multilayer perceptron (MLP) is proposed to evaluate the similarity between the input samples and incorporated into the optimization process of the RSNN sub-model. After sub-model training, a support vector regression (SVR) module is constructed to analyze features extracted by the backbone sub-network and predict tool wear values. Actual milling experiments were designed and conducted to gather available samples, and the following performance comparison experimental results prove the superiority of the proposed model under various sample sizes.