Abstract: With the development of intelligent manufacturing technology and the growing demand for personalization, multi-variety and small-batch production has gradually become the mainstream in the manufacturing industry. In this condition, traditional quality management methods which focus on large-batch production and statistical process control are not suitable for small-batch production. In complex production processes, it also exists challenges such as numerous parameters, non-linearity and interactions. To this end, a deep transfer learning method is adopted to predict the quality of target products using small sample data transferred from massive historical production data. First, by using the synthetic minority over-sampling technique (SMOTE) algorithm and an improved kernel principal component analysis (KPCA) algorithm, transferable features from both the source and target domains are selected, balancing feature importance and transferability. It also mitigates negative transfer issues and enhances the generalization capability of the model. Then, a quality prediction model based on deep transfer learning is built using a convolutional neural network, i.e., SeNet, which incorporates a channel attention mechanism. Simulation results demonstrate that as the number of target domain samples increases, the proposed method is significantly superior to prediction accuracy compared with the widely adopted support vector machine modeling method. Additionally, the proposed selection method of transferable features significantly enhances the quality prediction performance of deep transfer learning, providing a novel approach to ensuring the quality of complex small-batch production processes.

Key words: small-batch production quality prediction, deep transfer learning, synthetic minority over-sampling technique (SMOTE), kernel principal component analysis (KPCA), squeeze-and-excitation networks (SeNet)